Research papers and global crypto regulation, at a glance urn:diting-lab:feed:en/papers/ai/feed.xml 2026-10-30T11:54:16+09:00 Diting Lab Explanations and summaries are generated by AI from the original text. Dates, rates and amounts are checked against the source, but errors are still possible, so always rely on the original. Nothing here is legal, tax or investment advice. urn:diting-lab:en:doc:346121 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

The Power Watershed as a Differentiable Segmentation Layer arXiv · Oct 28, 2026The paper places a Power Watershed, the limit of a seeded Random Walker as its edge-weight exponent grows, at the end of a segmentation network so it can be trained end to end. The finite-exponent solve is exactly differentiable through one adjoint solve. On PanNuke nuclei segmentation, the layer turns a U-Net into an instance segmenter, raising instance F1 from 0.656 to 0.795. ★★★★ Notable: The authors embed a differentiable Power Watershed layer at the end of a U-Net and report instance F1 rising from 0.656 to 0.795 on PanNuke nuclei segmentation.

urn:diting-lab:en:doc:346120 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Similar Scores, Opposite Failures: Human Correction as an Instrument for Video Caption Evaluation arXiv · Oct 28, 2026The paper finds that reference-overlap metrics built for single-sentence image captions fail to separate reviewer-accepted from rejected dense video captions, performing near chance. Six frontier models captioned 196 short clips, which trained annotators post-edited against a fixed failure taxonomy, reviewing each caption once as a whole-clip overview and once window by window. Camera and shot errors were flagged about 4.7 times more often in the window-by-window review, and models with similar accuracy showed clearly different failure profiles. ★★★★ Notable: The authors use human post-editing against a fixed taxonomy on 196 clips, reporting window-level review finds 4.7x more camera errors than overview review, while BLEU-4 is near chance at separating ac

urn:diting-lab:en:doc:346097 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Predict the Reward, Not the World: World Reward Models for ML Agents arXiv · Oct 28, 2026This paper proposes World Reward Models (WRMs) for search-based machine-learning agents, which learn directly from real execution results which candidate program will perform better, instead of predicting intermediate execution outcomes. The proposed model, ForeCritic, picks the better program 79.4% of the time on 12,438 held-out pairs from 26 ML tasks, compared with 53% to 62% for general-purpose LLM judges. When integrated into AIDE, ML-Master, and ForeAgent, it cuts mean task time by 4.4 to 7.5 times while keeping mean scores within 0.01 or raising them. ★★★★ Notable: The authors propose world reward models that directly select the better candidate program without predicting execution outcomes, report proofs and 12,438 held-out pair evaluations, and test speedups w

urn:diting-lab:en:doc:346081 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Masked Erosion: How Acceptance Gates Lose Redundant Safeguards in Self-Improving Agents arXiv · Oct 28, 2026This paper studies self-improving LLM agents that edit their own prompts and rules, keeping edits only when an acceptance gate scores them higher on task batches. Because overlapping safeguards mask the effect of deleting one, the gate can approve harmful changes based on task success alone, eroding redundancy unseen. The authors measured harm across all subsets of four safeguards in two environments, running 411 improvement loops under ten gate rules. They also report that evaluation batch size and the statistical test used affect which safeguards are kept. ★★★★ Notable: The authors show acceptance gates erode redundant safeguards via masking, reporting 411 controlled loops and a frozen predictive model.

urn:diting-lab:en:doc:346074 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Towards Reliable, Scalable, and Value-Aligned Generative Recommendation for Large-Scale Advertising arXiv · Oct 28, 2026This paper presents GR4AD-X, an advertising recommendation framework for the Kuaishou platform that generates discrete semantic identifiers autoregressively. It improves semantic identifier reliability with Multi-Path assignment, scales model capacity along several dimensions under serving constraints, and aligns generation with commercial value through the UniVGD method. Large-scale online A/B tests on Kuaishou's advertising platform show a 3.07% cumulative revenue gain over the deployed GR4AD baseline while meeting latency and throughput requirements. ★★★★ Notable: The authors propose Multi-Path SID assignment, multi-dimensional capacity scaling, and unified value alignment, reporting a 3.07% revenue gain in large-scale Kuaishou online A/B tests.

urn:diting-lab:en:doc:346059 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Optimal Regret for Communication-Efficient Decentralized Online Learning arXiv · Oct 28, 2026This paper studies decentralized online convex optimization where networked learners minimize a sequence of global losses using only local losses and compressed messages from neighbors. The authors propose Ascent-gossip, which combines an online compression strategy with a Chebyshev recursion to reach consensus faster than existing compressed gossip methods. Using this method, they present an online algorithm whose regret bounds for convex and strongly convex losses match known lower bounds up to logarithmic factors. ★★★★ Notable: The authors combine online compression with Chebyshev recursion, reporting regret bounds matching lower bounds up to logarithmic factors.

urn:diting-lab:en:doc:346047 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Game-Theoretic Foundations for Undiscounted Bisimulation Metrics arXiv · Oct 28, 2026Bisimulation metrics measure task-relevant state similarity in reinforcement learning, but existing constructions mostly cover discounted MDPs. This paper reinterprets these metrics as game value functions of zero-sum turn-based games, giving unique definitions for total- and average-reward settings. It also analyzes how metric finiteness depends on MDP structure and shows that the metrics bound total-reward and optimal relative value differences. Error and suboptimality guarantees for aggregation-based planning are provided. ★★★★ Notable: The authors recast bisimulation metrics as zero-sum game values, giving unique definitions for undiscounted total- and average-reward MDPs, with bounds on value differences and aggregation-planning er

urn:diting-lab:en:doc:346026 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

ParaTTT: Parallelizing Test-Time Training Layers with Deep Memories arXiv · Oct 28, 2026This paper introduces ParaTTT (PTTT), an algorithm that parallelizes deep nonlinear memories in test-time training layers by iteratively refining estimates of the inner-loop gradient trajectory. Each refinement uses three chunkwise-parallel passes, and the exact online gradient descent recurrence is a fixed point of the iteration. With two refinement steps at 4K context, the inner-loop evaluation is reported to be 32 times faster than TTT-MLP with block approximation at matched error, and the 340M model trains 4.9 times faster. Models at 340M and 1.3B parameters match or exceed Gated DeltaNet and Mamba-2 on language-modeling perplexity and zero-shot accuracy, with better recall than linear recurrences. ★★★★ Notable: The authors propose iterative chunkwise-parallel refinement for deep nonlinear test-time training memories, report 340M and 1.3B model comparisons, and give a convergence proof for a two-layer variant

urn:diting-lab:en:doc:346020 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Nearly Minimax Optimal Regret for Adversarial Linear Bandits with Strong Corruptions arXiv · Oct 28, 2026This paper studies adversarial linear bandits where a strong adversary corrupts observed rewards after seeing the learner's action, within a known total corruption budget C. The authors propose CRW-FTRL, which uses a radius-weighted entropy regularizer and a normalized kernel estimator whose corruption sensitivity depends on per-action radii. The algorithm achieves regret that matches the best corruption-free rate and the corruption lower bound up to logarithmic factors. Applied to discretized general compact action sets, it yields regret of order d√T plus dC. ★★★★ Notable: The authors report a radius-weighted regularizer achieving near-minimax regret under strong corruption, matching the corruption-free rate up to logarithmic factors.

urn:diting-lab:en:doc:346009 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Disparity Laundering: Consensus Without Fairness in Clinical Multi-Agent Systems arXiv · Oct 28, 2026This paper examines demographic bias when clinical large language models operate as multi-agent teams. Testing five model families and four aggregation topologies on 120 MIMIC-IV emergency cases with counterfactual demographic swaps, all configurations preserved the direction of individual agents' answers, giving Black patients more acute triage. The authors report that fairness instructions and demographic masking did not close the gap, while a diverse-perspective ensemble that pools counterfactual versions during deliberation reduced the disparity by 67 to 94%. ★★★★ Notable: The authors introduce disparity laundering and report a full factorial of five model families and four topologies on 120 MIMIC-IV cases under counterfactual swaps.

urn:diting-lab:en:doc:346003 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Averaged Annealed Markov Chains Achieve the Algorithmic Threshold for Tensor PCA, Part II: Langevin Dynamics arXiv · Oct 28, 2026This paper studies tensor principal components analysis on the sphere, where an unknown vector must be recovered from a noisy observation of its k-th tensor power. The authors show that spherical Langevin dynamics, combined with annealing and ergodic averaging, achieves polynomial-time strong recovery above the algorithmic threshold from a uniformly random initialization. Together with Part I, the results show that local Markov chains can succeed down to the algorithmic threshold from an uninformative cold start. ★★★★ Notable: The authors show annealed, averaged spherical Langevin dynamics recovers tensor PCA signal in polynomial time above the algorithmic threshold from random initialization, continuing Part I.

urn:diting-lab:en:doc:345991 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

DecompGym: Benchmarking LLM Agents on Designing Decomposition Algorithms arXiv · Oct 28, 2026The paper introduces DecompGym, a benchmark of 81 optimization instances from nine problem families, including five Benders and four column generation tasks, where agents must write solvers for prescribed decomposition methods. Each solver is checked for running within 60 seconds, numerical feasibility, an LLM judge's audit of the method used, and an objective within 1% of the stored optimum. Evaluating 12 GPT, Claude, and Gemini models over up to ten attempts, the authors find that reflection, which diagnoses failures first, succeeds on 66% of retries versus 39% for independent regeneration. The study concludes that decomposition agents need checks on how a solver reaches its answer, not only output checks. ★★★★ Notable: The authors build a 81-instance benchmark testing LLM agents on Benders and column generation, comparing regeneration with reflection across 12 models, and separating output checks from method audits.

urn:diting-lab:en:doc:345979 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

BridgeRouter: Gain-Aware Instance Routing over Cross-Modal Bridge Portfolios arXiv · Oct 28, 2026This paper addresses transferring pretrained EEG foundation models to other biosignals without retraining their backbones. Instead of one global bridge pathway, BridgeRouter selects one of several frozen cross-modal pathways per instance, falling back to a validation-selected default route. It is evaluated on EMG-based freezing-of-gait detection and ECG-based sleep classification, where it improves balanced accuracy over the best single route.

urn:diting-lab:en:doc:345973 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

On Worst-Case Guarantees for Graph-Based Nearest-Neighbor Search arXiv · Oct 28, 2026This paper studies the theoretical guarantees of graph-based nearest neighbor search methods such as HNSW, DiskANN, and NSG, which perform well in practice. The authors prove that standard graph-based methods running greedy search on flat or hierarchical graphs cannot achieve sublinear query time in Euclidean space, even with arbitrarily large approximation factors. They show that allowing Steiner points in the search graph makes sublinear query time possible, with bounds comparable to LSH, and include initial experiments on incorporating Steiner points into practical methods. ★★★★ Notable: The authors prove standard graph-based search cannot achieve sublinear query time, and that Steiner points yield LSH-comparable bounds, with initial experiments.

urn:diting-lab:en:doc:345969 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

What Does LOL Measure? Hidden Label Shifts in Low-Light Enhancement Benchmarks arXiv · Oct 28, 2026This paper audits the standard low-light image enhancement benchmarks LOL-v1 and LOLv2-Real using 18 public checkpoints from eight methods. It finds that test labels differ from training labels, with LOL-v1 test labels rendered differently and LOLv2-Real test labels darker than its training labels. The authors show that test-set checkpoint selection and released test-time settings affect reported scores and method rankings, and they provide reporting recommendations. ★★★★★ Notable: The authors audit 18 checkpoints of eight methods, report hidden label shifts in LOL benchmarks, and quantify their effect on rankings.

urn:diting-lab:en:doc:345968 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Post-Training Large Language Models with LLMON arXiv · Oct 28, 2026This paper proposes a post-training method using LLMON (LLM Object Notation), which separates executable instructions from passive data and requires identifier-based execution. The authors built a 4.6B-token structured training corpus and post-trained Granite-4.1 3B and 8B base models with supervised fine-tuning and reinforcement learning. They report that attack success rates on indirect prompt-injection benchmarks fell from 53-100% to 0.02-5.6% while general capabilities were preserved. ★★★★ Notable: The authors report that LLMON-structured post-training cuts indirect prompt-injection success rates substantially on two Granite models, and introduce attention-based interpretability metrics.

urn:diting-lab:en:doc:345954 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Predicting What Answer Engines Cite: Open Decision Models, Rerankers and a Distilled 4B Student on Ten Engines under Controlled Retrieval arXiv · Oct 28, 2026This paper studies how to rank candidate web pages by how often answer engines will cite them in generated responses. The authors evaluate pretrained rerankers, a hosted decision model, eight open decision models, and distilled students on GEO-Bench queries across ten engines. A 4B student model fine-tuned from Qwen3-Reranker-4B performs comparably to the strongest scorers, and the model, benchmark, predictions, and code are released. ★★★★ Notable: The authors report comparing rerankers and distilled 4B students against teacher decision models on GEO-Bench across ten answer engines.

urn:diting-lab:en:doc:345942 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Vision Tomography: Towards Portable 3D Bidirectional Face-Skull Inference from Visual Data arXiv · Oct 28, 2026This paper introduces Vision Tomography (VT), a framework that reconstructs paired 3D face and skull geometry from visual inputs such as images or sketches. The authors built a dataset of over 200 cases with CT skulls, CT faces, and textured face scans, and proposed a 3D Bidirectional Face-Skull Morphable Model (BFSM) that infers shape between face and skull via a shared coefficient space. The framework is demonstrated in forensic and medical applications, including paired reconstruction and predicting surgical outcomes through joint face-skull editing. ★★★★ Notable: The authors propose a bidirectional face-skull morphable model and a 200-plus sample dataset for paired 3D reconstruction from visual data, reporting experiments in forensic and medical applications.

urn:diting-lab:en:doc:345930 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Integral Probability Boundaries arXiv · Oct 28, 2026This paper introduces Integral Probability Boundaries (IPBs), a non-linear generalization of Integral Probability Metrics (IPMs). IPMs reduce the gap between two distributions to a single worst-case difference in expectations, which can yield loose or vacuous bounds for rare events, safety failures, or under-represented groups. IPBs are designed to remain tight at these extremes. The authors derive their basic properties, present estimation algorithms, and demonstrate use in alignment, judge-free performance estimation, and fairness, including more precise fairness-gap estimates for rare groups in a generative modeling task than prior methods.

urn:diting-lab:en:doc:345904 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Who Else Is In? Multi-Agent Collusion Initation Via Strategic Vagueness arXiv · Oct 28, 2026This paper studies how multi-agent LLM systems could begin collusion, with agents using vague public messages to find hidden partners while appearing to do ordinary work. The authors run a mutual identification game in which two agents must recognize each other while solving a shared task, and some models succeed without explicit guidance, often evading monitors at low false positive rates. They report that as messages grow more indirect, monitor punishment can fall faster than partner coordination, creating a profitable middle ground. ★★★★ Notable: The authors introduce a mutual identification game showing strategic vagueness can initiate collusion past monitors; the abstract reports successful handshakes and an asymmetric monitor-punishment ver

urn:diting-lab:en:doc:345853 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Human-Residual Steering: Aligning Frozen LLM Judges with Human Judgment Distributions arXiv · Oct 28, 2026This paper proposes Human-Residual Steering (HRS), a method that aligns frozen large language model judges with the distribution of individual human ratings by learning hidden-state interventions. Across 14 benchmarks and four judges from three model families, HRS outperforms all evaluated frozen-weight baselines on distribution matching. It also trains about 88 times fewer parameters than distributional fine-tuning. ★★★★ Notable: The authors propose learned hidden-state interventions to align frozen LLM judges with human distributions, reporting gains across 14 benchmarks with far fewer trained parameters.

urn:diting-lab:en:doc:345851 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Finite-Sample Optimal Conformal Prediction Sets under Covariate Shift arXiv · Oct 28, 2026This paper studies conformal prediction under covariate shift when the covariate density ratio is known, aiming for prediction sets with minimal expected length under the target distribution. For every calibration sample size, the authors prove that combining an inverse conditional-density score with randomized weighted conformal calibration minimizes expected length among all rules valid over the weighted-exchangeability class. They approximate the oracle with quantile and kernel histograms, including forest-weighted estimators, and select estimators and tuning parameters by resampling. Experiments on 36 known-density configurations and four public regression datasets compare coverage, set length, and runtime with existing methods. ★★★★ Notable: The authors prove the inverse-density score with randomized weighted calibration minimizes expected set length over weighted-exchangeable rules, and report experiments on 36 density configurations and

urn:diting-lab:en:doc:345845 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

An Ω(T³/4) lower bound for projection-free online learning with linear minimization oracles arXiv · Oct 28, 2026This paper studies projection-free online learning algorithms that handle the decision set using linear minimization oracles. Prior work showed that Online Frank-Wolfe achieves regret scaling as T^{3/4} under specific tunings, but whether all tunings do so was open. The authors prove a universal worst-case regret lower bound of Ω(T^{3/4} m^{-1/4}) for any projection-free algorithm making at most m oracle calls per round. They also propose an extended Online Frank-Wolfe that uses multiple linear optimizations per round and prove a regret upper bound of O(max{T^{3/4} m^{-1/4}, √T}). ★★★★ Notable: The authors prove a universal regret lower bound for projection-free algorithms and match it with an extended OFW upper bound, answering an open question.

urn:diting-lab:en:doc:345840 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Investigating the Early Signatures of Feature Learning in Deep Linear Networks arXiv · Oct 28, 2026The paper analyzes features learned by a deep linear network in a mean-field parameterization after a single gradient step from initialization, using high-dimensional data with a low-dimensional spiked signal. It derives closed-form expressions for the bias, variance, and generalization error of kernel ridge regression on these features. A one-step generalization advantage appears only when spike strength and effective signal-to-noise ratio are sufficiently large, and it vanishes for isotropic inputs. ★★★★ Notable: The authors give a closed-form account of one-step feature learning in mean-field deep linear networks and derive bias, variance, and generalization error expressions.

urn:diting-lab:en:doc:345837 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

Learning Lotteries with Minimal Stakeholder Violations from Membership Queries arXiv · Oct 28, 2026The paper studies how to pick a randomized mixture of several candidate AI system versions that the fewest stakeholders reject. Each stakeholder accepts only mixtures inside an unknown halfspace, and the developer can learn this only by asking yes-or-no membership queries. The authors show that minimizing rejections is APX-hard in general, give exact algorithms for special cases, and provide approximation guarantees when only some stakeholders can be queried. ★★★★ Notable: The authors model stakeholder acceptance as unknown halfspaces, prove APX-hardness for minimizing rejections, and give polynomial and quasi-polynomial exact algorithms with sampling guarantees, plus s

urn:diting-lab:en:doc:345829 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

TACOS: Terrain-Aware COntact Supervision for Visual Footstep Planning arXiv · Oct 28, 2026This paper presents TACOS, a method that builds a footstep instability score from the residual of a factor graph used in legged robot state estimation, then projects that score onto earlier camera images to train a dense visual predictor. It proposes a hard score that uses hindsight and a causal soft score available immediately. On 1716 manually annotated contacts and three outdoor test runs, both variants outperform the compared alternatives at detecting instability and ranking its severity. In a real-world 50 m forest trial, the trained predictor handles mud and vegetation and steps around fallen branches. ★★★★ Notable: The authors derive a footstep instability score from factor-graph residuals to supervise a visual predictor, compared against baselines on 1716 annotated contacts and outdoor runs.

urn:diting-lab:en:doc:345823 2026-10-28T00:00:00+09:00 2026-10-30T11:54:16+09:00

WebAnything: From the Wild Web to Scalable Agent Environments arXiv · Oct 28, 2026WebAnything is a framework that automatically rebuilds reusable, resettable website environments from real URLs, so web agents can be trained and tested without manual per-site setup. It constructed 244 website environments across 14 application domains, supporting 12,984 executed and accepted tasks. In a blind audit it outperformed the human-assisted WebHarbor on three fidelity metrics. Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories improved success rates on several benchmarks. ★★★★ Notable: The authors automate reconstruction of reusable web environments from real URLs with an independent verifier, reporting 244 sites and gains from fine-tuning on 5,856 trajectories.

urn:diting-lab:en:doc:345804 2026-10-27T00:00:00+09:00 2026-10-30T11:54:16+09:00

What a Weight Update Buys: Leverage Across Training Stages arXiv · Oct 27, 2026The paper introduces leverage, a measure of how much a unit of weight change affects a language model's behavior, and applies it to released Olmo-3, Tülu-3, Zephyr, and OLMo-2 checkpoints. It finds that pretraining updates act like random directions on prose, while SFT, RL, and DPO stages show much higher leverage, up to about 23,000 for DPO. The authors also study how training precision and DARE random drop-and-rescale affect these measurements. ★★★★ Notable: The authors introduce leverage, a Fisher-form measure of weight update effect, comparing training stages across released checkpoints and deriving a DARE reuse law.

urn:diting-lab:en:doc:345801 2026-10-27T00:00:00+09:00 2026-10-30T11:54:16+09:00

Moment-Guided Optimizers for Noisy Deep Learning: MGO-SC as a Simple Consistency Gate for Adam-Style Updates arXiv · Oct 27, 2026This paper proposes MGO-SC, a lightweight modification of Adam-style optimizers that gates each coordinate's update by the stability of its gradient sign. A mean-preserving variant reallocates trust across coordinates without changing the net per-tensor step size. On CIFAR-10 with ResNet-18, it improves over an AdamW baseline under clean and noisy labels, with bootstrap confidence intervals and Wilcoxon tests reported.

urn:diting-lab:en:doc:345799 2026-10-27T00:00:00+09:00 2026-10-30T11:54:16+09:00

Relevance-Guided Residual Calibration for Spectral Compression of Dense Retrieval Embeddings arXiv · Oct 27, 2026This paper addresses spectral compression of dense retrieval embeddings, noting that existing corpus-derived weighting rules do not account for each retained direction's contribution to relevance. The authors propose Relevance-Guided Residual Calibration (RRC), a closed-form method that uses relevance judgments to refine spectral weights while preserving the compressor's overall trend. Experiments on nine datasets report that SpecTemp combined with RRC achieves the highest mean effectiveness among tested compressors on BEIR and BRIGHT, and the gains persist under scalar and product quantization. ★★★★ Notable: The authors add relevance-guided per-direction residual calibration to spectral compression, reporting gains over SpecTemp on nine BEIR and BRIGHT datasets.

urn:diting-lab:en:doc:345794 2026-10-25T00:00:00+09:00 2026-10-30T11:54:16+09:00

Network Topology Does Not Matter: The Breakdown of Classical Consensus in LLM Populations arXiv · Oct 25, 2026This paper studies how networks of multiple large language models reach collective agreement. Using eight instruction-tuned models and 73,080 controlled exchanges, the authors measure the models' update behavior and find that they often update outside the range of observed opinions and show active repulsion through negative weights. In simulations of 10,000 agents, network connectivity explains almost none of the outcome variance, while the models' internal update bias and message framing largely govern population dynamics. The authors conclude that classical graph-theoretic metrics are inadequate for LLM populations. ★★★★ Notable: The authors challenge classical consensus assumptions, measure LLM update operators empirically, and report simulations of 10,000 agents.

urn:diting-lab:en:doc:346124 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

What Hallucinations Have We Measured? Semantic Illusion, Confabulation, and Expressed Uncertainty arXiv · Oct 28, 2026This paper argues for a strict definition of hallucinations in large language models as confident incorrectness, separate from expressed uncertainty. It proposes a re-elicitation test that splits errors into semantic illusions, where correct knowledge is accessible, and confabulations, where it is not. Re-evaluating 14 recent hallucination methods, the authors find that detectors flag expressed uncertainty more often than hallucinations, and that mitigation works far better for semantic illusions than for confabulations.

urn:diting-lab:en:doc:346123 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Self-Consistency Geometry Injection for Multi-view Consistent Image Generation arXiv · Oct 28, 2026The paper proposes RGMV, a reliability-guided multi-view generation method with a parameter-free Self-Consistency Geometry Gate (SCG) and a Reliability-aware Geometry Injection (RGI) module. It generates several geometry predictions for the same reference image and target viewpoints, measures their spatial agreement to form a reliability map, and uses this map to modulate how strongly geometry-conditioned residuals are injected into RGB synthesis. According to the authors, experiments show that this reliability-aware injection improves the geometric quality of multi-view generation.

urn:diting-lab:en:doc:346122 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

AtomSteer: Test-Time Constraint Steering for Atomistic Generative Models arXiv · Oct 28, 2026This paper proposes AtomSteer, a framework that steers pretrained atomistic generative models at inference time so their outputs satisfy structural constraints such as crystallographic symmetry, geometry, and stereochemistry. During sampling, it corrects intermediate structures using bounded local projections, graph-rigid stereochemical transformations, and projections over periodic symmetries, while limiting changes elsewhere. The method is tested on FlowMM, ET-Flow, and Boltz-2 across crystal generation, molecular conformer generation, and protein-ligand complex prediction, without retraining or architectural changes.

urn:diting-lab:en:doc:346119 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Leaving the Cave by Breaking the Symmetry: Identifiable Structure Can Enable Cross-Lingual Generalization arXiv · Oct 28, 2026The paper asks why multilingual models generalize across languages without a shared vocabulary and proposes the Relational Identifiability Hypothesis, which holds that relations in a shared world can rule out alternative cross-lingual correspondences and make corresponding entities more directly readable. The authors test this in a controlled bilingual Transformer trained on two token-disjoint languages describing the same modular-arithmetic world. Cross-lingual alignment emerges selectively when relations constrain number correspondences, and these patterns predict success in zero-shot cross-lingual transfer.

urn:diting-lab:en:doc:346118 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SaiVLA-JEPA: Evaluating Future Latent Prediction in JEPA-Conditioned VLA Control arXiv · Oct 28, 2026This paper studies whether a vision-language-action robot policy that uses V-JEPA visual features benefits from a predictive training objective. The authors build SaiVLA-JEPA, which pairs a vision-language model with a diffusion transformer action expert, and compare variants on 10,030 paired LIBERO-Plus episodes. Adding current JEPA conditioning raised success to about 82.5% from 71.6% without JEPA, while extra future supervision gave no detectable overall gain at the tested training budget.

urn:diting-lab:en:doc:346117 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

UnMixMamba: Unmixing Guided State Space Subpixel Mapping of Crops from MODIS Time Series arXiv · Oct 28, 2026This paper proposes UnMixMamba, a method for mapping crop types from 250 m MODIS time series to field-scale AAFC labels. It estimates simplex fractions through a temporal-spectral-spatial decoupled factorization and reconstructs the MODIS data using abundance super-resolution and a learnable per-crop endmember bank. The predicted crop fractions direct high-resolution refinement toward the minority class. On Alberta 2015 under a fixed spatial-block protocol, it outperforms several baselines on overall accuracy, mIoU, and Kappa.

urn:diting-lab:en:doc:346116 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Residual Motion Splatting for Continuous Space-Time Video Super-Resolution arXiv · Oct 28, 2026This paper addresses continuous space-time video super-resolution, which turns a low-resolution, low-frame-rate video into a high-resolution, high-frame-rate one at arbitrary scales and times. The method predicts per-pixel continuous trajectories that refine optical-flow paths with smooth residual motion, then splats consecutive low-resolution latent features onto a high-resolution grid at predicted sub-pixel positions and visibility weights. A decoder resolves remaining ambiguities, making temporal interpolation and multi-frame super-resolution one operation. The authors report better fidelity and perceptual quality on several benchmarks, with speeds up to an order of magnitude faster.

urn:diting-lab:en:doc:346115 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Vision-Based Density Estimation for Adaptive Traffic Control arXiv · Oct 28, 2026This paper proposes a framework that optimizes traffic flow directly from camera images, with three parts: vehicle density estimation, predictive dynamics modeling, and signal policy optimization. The density estimation stage was evaluated on 40 annotated intersection images from Rambagh 4, Jaipur, India, reporting a five-fold validation mean absolute error of 3.32 ± 0.95 vehicles and a mean absolute percentage error of 14.5%. Empirical training and evaluation of the forecasting and control stages are left for future work.

urn:diting-lab:en:doc:346114 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Accelerating Nonlinear MPC by Behavior Cloning with Separated Timescales: Application to Nuclear Reactor Load-Following arXiv · Oct 28, 2026This paper proposes SRLBC, a behavior cloning architecture that uses separate embeddings for slow state variables, fast stiff state variables, and exogenous inputs in multiscale systems. On 500 long-horizon roll-outs of a pressurized water reactor load-following benchmark, separated embeddings outperform shared-embedding baselines on cost, imitation accuracy, and constraint satisfaction. Embedded in a safety-augmented approximate NMPC framework, the controllers are about 6.8 times faster than NMPC with feasibility of at least 99.8%.

urn:diting-lab:en:doc:346113 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Calibrating Track-Query Updates at Test Time for Multi-Object Tracking arXiv · Oct 28, 2026This paper studies identity switches in query-propagation multi-object trackers, which often occur when visually similar targets overlap or cross. The authors find that propagated track queries change more abruptly before switches, and that current-frame updates point more strongly toward nearby active tracks. They propose Test-time Track-Query Calibration (TQC), a training-free, gradient-free method that applies a bounded correction before propagation. On nine DanceTrack and SportsMOT test settings, TQC improves HOTA and AssA in all settings, with an average HOTA gain of +0.38.

urn:diting-lab:en:doc:346110 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Beyond Reachability: When Edit Order Preserves Couplings in Graph Flow Matching arXiv · Oct 28, 2026The paper studies whether reordering legal shortest edits in a graph generator can preserve a prescribed source-target pairing without changing the pairing, mandatory edits, or sampling budget. Under a common Poisson clock, it gives an exact criterion: the memoryless Markov projection reaches the unavoidable clock-error floor if and only if routes to distinct targets share no graph states. The work also proposes a legality-preserving scheduling procedure with recovery certificates.

urn:diting-lab:en:doc:346109 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Geometry-Aware Trajectory State Modeling for Few-Shot Action Recognition arXiv · Oct 28, 2026This paper proposes GeoTRAM, a framework for few-shot action recognition that lifts sparse point tracks into 2.5D trajectories. It uses StateMotion to model ordered state evolution within each trajectory and RelMotion to capture local interactions through time-varying relative geometry and motion. The method adds quality-filtered spectral sampling, separately refined semantic, motion, and depth cues, and action-guided reweighting with top-ρ trajectory selection. Experiments on five benchmarks show strong gains on the motion-centric Something-Something datasets and competitive results on appearance-driven ones.

urn:diting-lab:en:doc:346107 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Cinematize from Unreal: Agentic Choreography for Controllable World Simulation arXiv · Oct 28, 2026The paper presents CineU (Cinematize from Unreal), an agentic system that uses language to orchestrate worlds and cinematic choreography inside Unreal Engine. It treats the Unreal project as a persistent, inspectable world representation, with agent-generated code whose effects are executed and verified. The system combines hierarchical orchestration, asset retrieval with geometry-constrained placement, code synthesis, and multi-stage verification with localized repair. Experiments on scene construction, selective editing, and multi-shot cinematography show that engine feedback improves task completion and recovery.

urn:diting-lab:en:doc:346106 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Automatic Iconographic Annotation of Medieval Illuminations: Comparing Generative Vision-Language Models and Supervised Discriminative Classifiers arXiv · Oct 28, 2026This paper compares two approaches for automatically tagging the iconography of medieval manuscript illuminations, using 9,735 full-folio images labelled with 2,333 TIMEL concepts. It tests generative vision-language models against supervised discriminative classifiers, with ML-Decoder achieving the strongest retrieval results (precision@1 of 0.696, MRR of 0.806). The authors conclude the two paradigms are complementary and propose a soft F1 metric that credits taxonomically close predictions.

urn:diting-lab:en:doc:346104 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

When Support Crosses the Line: Evaluating Ethical Boundaries in LLM-Based Mental Health Support in the Chinese Context arXiv · Oct 28, 2026This paper introduces MindEth-Bench, a benchmark that tests whether large language models keep to professional ethical boundaries in multi-turn mental health conversations. It is grounded in Chinese professional ethical standards for mental health services and uses Monte Carlo Tree Search to generate interaction paths that stress-test model boundaries. Experiments on nine LLMs show that models responding appropriately at first can drift into ethical violations over longer exchanges, with wide variation across models and dimensions. The authors conclude that single-turn evaluations may overestimate the ethical reliability of such systems.

urn:diting-lab:en:doc:346102 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

CoSTA: Co-Evolving Self-Play Timeline Agents for Complex Temporal Knowledge Graph Question Answering arXiv · Oct 28, 2026This paper addresses complex temporal knowledge graph question answering, where the time constraints of one fact affect the search for others. It proposes CoSTA, which organizes facts into navigable timelines, assigns investigations to worker agents, and has a coordinator synthesize their timestamped observations. A shared policy is initialized by role-conditioned distillation and then both roles are optimized jointly through co-evolving self-play with final-answer rewards. The authors report state-of-the-art overall performance on three benchmarks, with compact variants outperforming evaluated closed-source LLM agents.

urn:diting-lab:en:doc:346101 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Benchmarking decision models against generative large language models on medical multiple-choice questions arXiv · Oct 28, 2026The paper benchmarks 13 decision models against 15 generative large language model configurations on 1,273 four-option MedQA questions under text-only conditions. Accuracy ranged from 67.71% to 89.08% for decision models and 78.32% to 96.86% for generative configurations. Decision models showed resource advantages but lower peak accuracy. The authors argue model selection should weigh probability quality and repeatability along with accuracy and error-correction costs.

urn:diting-lab:en:doc:346099 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

DES-JEPA: Learning a Shared Vocabulary for High-Dimensional Discrete Event Sequences arXiv · Oct 28, 2026This paper addresses discrete event sequences such as operational logs and patient clinical-event trajectories, whose event labels differ across domains and can number in the hundreds. It proposes DES-JEPA, which learns a shared discrete vocabulary from synthetic stochastic processes with unaligned event vocabularies, and DES-PFN, which uses that vocabulary for in-context next-event prediction. Without task-specific fine-tuning, the framework achieves competitive zero-shot performance on eight unseen real-world datasets and scales to 700 distinct event types.

urn:diting-lab:en:doc:346096 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

RoleFUZZER: User-Behaviour Fuzzing for Role-Specification Adherence in Large Language Models arXiv · Oct 28, 2026Large language models are often given roles such as tutor or financial advisor through system prompts, but how reliably they keep to these roles during multi-turn conversations with difficult users has been little studied. The paper proposes user behaviour fuzzing and introduces RoleFUZZER, a framework adapting coverage-guided fuzzing from software testing to search for user behaviours that cause role violations. Across five role specifications, the framework covers at least 89% of a 64-cell behaviour grid for every model and role tested, and reveals specification-grounded failure patterns. Compared with free-form user generation at the same budget, the grid-guided approach achieves higher coverage and finds more kinds of violations.

urn:diting-lab:en:doc:346095 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Capacity of the Hopfield-Hebbian dynamics with correlated patterns arXiv · Oct 28, 2026This paper studies how many patterns a Hopfield-Hebbian associative memory network can reliably store when the patterns are correlated rather than independent. Each archetype pattern is replaced by M correlated examples, with a parameter r controlling the correlation quality. The authors use a Random Duality Theory framework to derive capacity characterizations that closely match simulated dynamics and explicitly show the effects of M and r.

urn:diting-lab:en:doc:346094 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Continual Adaptation Across Modalities: Learning and Forgetting in Unified Multimodal Models arXiv · Oct 28, 2026This paper studies how unified multimodal models, which handle several modalities in one framework, learn and forget when trained continually on tasks across modalities. It examines how architectural design and adaptation methods affect interference between the generation and understanding sides. The authors find that interference shrinks as modality-specific parameters are more isolated, that the generation side is more fragile than the understanding side, and that post-training abilities such as personalized generation drop by 84% to 90% while pretraining abilities stay stable.

urn:diting-lab:en:doc:346093 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Does Chain-of-Thought Reliance Follow a Scaling Law?Larger Models Follow Their Steps More, Independent of Post-Training arXiv · Oct 28, 2026The study tests how much a model's final answer depends on its own written reasoning steps, using a synthetic box-swap task where every intermediate state is known. The authors replace one step in the chain with a wrong but valid state and check whether the answer changes accordingly. Across 25 open-weight models, reliance rises with size and saturates, is driven mainly by capability, and is essentially unchanged by post-training.

urn:diting-lab:en:doc:346090 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

The Code Behind the Layout: AI Agents for Analog Layout Generators arXiv · Oct 28, 2026Analog layout design remains largely manual and difficult to automate. The paper examines whether large language models should write or edit generator code for frameworks such as gLayout, comparing this with having models draw layouts directly in SKY130 through controlled experiments. The authors find that the main obstacle to direct layout drawing is time and token cost, while the main advantage of building on a generator framework is reusability.

urn:diting-lab:en:doc:346089 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

IntReal: A Real-World Dataset and Benchmark for Multimodal Interleaved Generation arXiv · Oct 28, 2026This paper introduces IntReal, which pairs IntReal-100K, a real-world dataset of 100K multimodal interleaved samples across four categories, with an evaluation suite that measures models along seven complementary dimensions. The suite shows low redundancy between metrics and high agreement with human judgments. Benchmarking four families of unified multimodal models reveals distinct strengths and weaknesses, and post-training experiments show the dataset improves interleaved generation, while unimodal generation training also helps.

urn:diting-lab:en:doc:346088 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Deeper Backpropagation or More Updates? Training Looped Language Models Across Budgets arXiv · Oct 28, 2026The paper studies how to choose the backward window in training looped language models, which sets how many final layer repetitions keep gradients. It compares window sizes across different training budgets, using training recipes of roughly equal time. The results show that shorter windows lose much of their advantage as budgets grow, so the preferred window depends on the budget.

urn:diting-lab:en:doc:346087 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Structure-retained low-rank adapters for weather synthesis arXiv · Oct 28, 2026This paper presents a weather synthesis method that uses structure-retaining Low-Rank Adapters (LoRA) on Stable Diffusion XL (SDXL) to generate images under different weather conditions. The authors analyze which attention blocks in SDXL mainly process structural information during image-to-image translation, then learn structural LoRA matrices for those blocks, removing the need for external conditioning such as ControlNet. Experiments reportedly show greater realism than existing synthesis approaches while preserving scene structure. The authors suggest the pipeline can generate training and evaluation data for outdoor computer vision.

urn:diting-lab:en:doc:346086 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Governance-Aware Federated Learning for Multi-Agent AI Systems: A Policy-Driven Framework for Trustworthy Collaborative AI arXiv · Oct 28, 2026This paper proposes Governance-Aware Federated Learning (GAFL), which adds independent verification of governance evidence, trust tracking, fairness and explainability checks, anomaly detection, policy enforcement, and tamper-evident audit logging to federated learning. A 10-participant cross-silo prototype was tested on the Adult Income dataset under several honest, biased, malicious, and falsified-evidence scenarios. Complete GAFL identified all falsified-evidence participants, while the version without independent verification identified none, at a higher per-round governance processing cost than FedAvg.

urn:diting-lab:en:doc:346085 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Holistic Diversity Management in LLM-Guided Evolutionary Search arXiv · Oct 28, 2026This paper proposes HoliDivEvolve, which maintains and uses diversity among evolving program islands throughout LLM-guided evolutionary search. It introduces diverse initialization, trajectory adaptive island selection based on current score, recent growth, and remaining budget, and growth adaptive migration to share ideas across islands. Experiments across diverse benchmarks show competitive performance on most evaluated tasks, and controlled analyses examine how design choices affect search behavior and final results.

urn:diting-lab:en:doc:346084 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SpaceAct: Benchmarking 3D Spatial and Temporal Reasoning via Closed-Loop Action and Active Perception arXiv · Oct 28, 2026This paper introduces SpaceAct, an interactive benchmark that tests multimodal large language models on 3D spatiotemporal reasoning through closed-loop physical action and active perception. It contains 108 verified instances across 36 task families, built on interlocking bricks and mechanical puzzles, and is paired with the SAPA agent framework. Models must rotate the camera to resolve occlusion, output 6-DoF poses, and pass collision and statics checks. Results show that even with a valid plan, models remain limited by 3D visual grounding and action refinement, while exact 3D state information sharply reduces collisions.

urn:diting-lab:en:doc:346083 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

LLMs as Judges and Examinees for Medical Item Difficulty Prediction arXiv · Oct 28, 2026This paper studies how to predict the difficulty of medical exam questions from their text alone. Using USMLE and PANCE licensure exam items, open-weight large language models act as judges, giving absolute difficulty ratings and pairwise comparisons, and as simulated examinees, producing signals such as correctness and answer probabilities. Combining these signals with baseline item features yields Spearman correlations of 0.447 on USMLE and 0.494 on PANCE. These results are comparable to direct ratings from the proprietary model Claude Opus 5.

urn:diting-lab:en:doc:346082 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Agent Transfer Protocol: Server-Mediated Agent-to-Agent Communication Architecture for the Agentic Web arXiv · Oct 28, 2026This paper proposes ATP (Agent Transfer Protocol), a cross-organization agent communication protocol that uses domain-mediated architecture. Administrative domains hold stable addresses and receiving entries for logical agents, and the domain servers of both parties handle cross-domain transmission. The receiving domain verifies and stores messages for later reading, while the sending domain signs the complete message envelope. The authors built a prototype and evaluated it with a controlled cross-organization LLM agent workflow, mechanism checks, and latency measurements.

urn:diting-lab:en:doc:346080 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SCOPE: Selective Critical-Step On-Policy Self-Distillation for Agentic Reinforcement Learning arXiv · Oct 28, 2026This paper proposes SCOPE, a training framework for long-horizon LLM agents that addresses the coarse supervision of outcome-based reinforcement learning. After a supervised warm-up, the policy reviews its own trajectories, generates step-specific hindsight skills, and distills guidance only at selected error steps, with a bounded signed coefficient and a success-relative pairwise length reward. Experiments on ALFWorld and WebShop report consistent gains across backbones, with Qwen2.5-3B reaching 97.7% and 85.2% success and Qwen3-32B reaching 94.0% on unseen ALFWorld tasks.

urn:diting-lab:en:doc:346079 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Visually Irrelevant Hallucinations in Medical Vision-Language Models arXiv · Oct 28, 2026This paper studies hallucinations in medical vision-language models, where generated clinical statements lack visual support or contradict the image. The authors propose CLEAR, a model-agnostic post-hoc framework that estimates phrase-level hallucination risk using localized counterfactual image perturbations. CLEAR outperforms existing hallucination detection baselines on four datasets, raising AUPRC on MIMIC-CXR from 85.10% to 90.23%.

urn:diting-lab:en:doc:346078 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Revisiting Temporal Link Prediction through Structural Proximity arXiv · Oct 28, 2026The paper revisits temporal link prediction and shows that Personalized PageRank on historical interaction graphs, with sliding-window retention or exponential time decay, can match or outperform existing methods without training. Across 14 datasets, Time-decay PPR ranked first on nine and second on four. The authors argue that strong results do not necessarily require complex temporal models and call for reassessing progress with classical baselines.

urn:diting-lab:en:doc:346077 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Pronunciation-Supervised Text-to-Speech for Irish (Gaeilge) arXiv · Oct 28, 2026The paper adapts the pretrained Qwen3-TTS text-to-speech model to Irish using 12.8 hours of crowdsourced Common Voice speech. Its character error rate is 0.225 on held-out cross-corpus text, beating a from-scratch system (0.397) and zero-shot cloning (0.437). The authors convert Wiktionary pronunciation records into a phoneme critic and use it in one round of rejection fine-tuning, reducing cross-corpus error to 0.159. The study also finds that the training data is too skewed toward Munster to steer the voice toward the Ulster dialect.

urn:diting-lab:en:doc:346076 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

How Much Do Teacher Bridges Matter in Interleaved On-Policy Distillation? arXiv · Oct 28, 2026The paper studies on-policy distillation, where a student model learns from teacher predictions on its own responses, and some methods insert teacher-written bridge segments into those responses. The authors compare no loss, forward KL, and reverse KL on identical trajectories, using Qwen3-1.7B as the student and Qwen3-4B-Instruct-2507 as the teacher. In the main comparison, the three loss treatments were statistically equivalent in accuracy within a ±5 percentage-point margin, and no added benefit from teacher text was detected.

urn:diting-lab:en:doc:346075 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Compute Has Shape: Self-Morphing the Internal Computation of Frozen LLMs arXiv · Oct 28, 2026This paper proposes SelfMorph, which chooses the internal computational structure of a large language model at inference time for each input. It combines the original forward path, serial reuse of a layer span, and local branching and merging of perturbed hidden states. A lightweight selector reads the hidden-state trajectory on the prompt to pick a morphology, trained by distilling a utility distribution from offline morphing. Across three models from 4B to 27B, it reports mean relative gains of 8.06% on in-distribution benchmarks and 4.45% on unseen benchmarks.

urn:diting-lab:en:doc:346072 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

TCC-RAG: Temporal Causal Chain-Guided RAG for Temporal Knowledge Graph Question Answering arXiv · Oct 28, 2026This paper addresses a limitation of semantic retrieval-augmented generation for temporal knowledge graph question answering, namely its difficulty capturing multi-hop causal knowledge and its tendency to add redundant information. The authors propose TCC-RAG, which converts the temporal knowledge graph into a multi-view causal cue graph, builds a query-specific temporal causal chain to guide retrieval, and uses a causal evidence filter to refine the results. Experiments report gains of 10.72% and 8.40% on two TKGQA benchmarks and a 62.97% improvement for multi-step temporal reasoning.

urn:diting-lab:en:doc:346070 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SparseTokenGS: Gaussian-based Sparse-View CBCT Reconstruction via Reusable Token Pool arXiv · Oct 28, 2026This paper proposes SparseTokenGS, a method for reconstructing cone-beam CT volumes from sparse projection views, which otherwise produce severe artifacts. It uses a reusable token pool so that scan-wide context remains available during coordinate decoding, and reconstructs the volume in a single forward pass. On three public chest and dental datasets it achieves the highest reconstruction performance among the compared methods.

urn:diting-lab:en:doc:346069 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Volley3D: Unified 3D Full-Sequence Reconstruction from Monocular Broadcasts arXiv · Oct 28, 2026Volley3D is a pipeline for reconstructing volleyball play in 3D from monocular broadcast video, without using 3D training supervision. It corrects player positions using jump detection and Newtonian interpolation, and estimates ball trajectories through sub-trajectory segmentation and collision matching anchored to reconstructed 3D hand vertices. The authors report that on multi-camera benchmarks, the approach significantly reduces 3D spatial drift compared with 2D-reprojection baselines.

urn:diting-lab:en:doc:346068 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

CSO: Refining Robotic Policies via Skill Distribution Alignment and Skill-Grained Optimization arXiv · Oct 28, 2026This paper proposes Cascaded Skills Optimization (CSO), a two-stage post-training framework for improving robotic manipulation policies built on discrete skills. The first stage uses rejection-sampling supervised fine-tuning to align the observation-to-skill mapping with the distribution of successful online trajectories. The second stage freezes the observation encoder and computes an independent, clipped importance ratio for each complete skill to enable stable updates. Using a lightweight QueST-style policy, CSO raises overall LIBERO success from 80.1% to 97.5% and reaches 61.0% success across five physical single-arm tasks.

urn:diting-lab:en:doc:346067 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Encode Once, Answer Many: Internalizing Context via Parametric Flow arXiv · Oct 28, 2026The paper proposes CLIK, a generator that turns a given context into LoRA parameters in a single forward pass, which are then inserted into a frozen base model. The model can then answer queries from the question alone, without re-reading the context. On document question answering, CLIK clearly beats the no-context base model and prior hypernetwork generators, though it still trails a model that reads the context directly. It also transfers to multi-turn dialogue without multi-turn-specific training.

urn:diting-lab:en:doc:346065 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

WeVisCut: The Whole Editing Room in One Agent From Narration to Long-Form Video arXiv · Oct 28, 2026WeVisCut is a system that turns narration scripts into long-form videos by linking tool execution with policy learning in one agent harness. The authors build a four-stage data pipeline that recovers editing intent from finished videos and replays it as executable training data, yielding over 130K examples. They train a single 9B policy on this data that coordinates production and recovers from failures, and introduce a 100-task benchmark with five metrics.

urn:diting-lab:en:doc:346064 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Aligning Misaligned Benchmarks for Single-View Novel View Synthesis arXiv · Oct 28, 2026This paper shows that evaluating single-view novel view synthesis depends heavily on how each method's reference frame is aligned with the ground-truth pose. The authors align 3D reconstructions using traditional 2D-to-2D and 2D-to-3D image matching and evaluate only within co-visible regions. With proper alignment, explicit 3D Gaussian methods reach state-of-the-art performance. The paper also presents a simple baseline combining an off-the-shelf video generator and a feed-forward geometry estimator, which performs comparably to methods trained specifically for this task.

urn:diting-lab:en:doc:346063 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Knowledge-structured and evidence-aware semantic self-distillation for glomerular pathology foundation model arXiv · Oct 28, 2026This paper introduces GloSAGE, a self-distillation framework that uses lesion profiles predicted by a pretrained visual teacher to build structured pathology descriptions for glomerular images. It trains projection layers for image-text alignment while keeping the visual encoder frozen, and was developed on GloMillion. Evaluated across seven glomerulus- and case-level cohorts, it supports zero-shot lesion recognition and image-text retrieval. The authors suggest teacher-derived lesion profiles can provide annotation-efficient semantic supervision for a glomerular visual foundation model.

urn:diting-lab:en:doc:346062 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Investor's Last Exam: Can AI Agents Do Equity Investment Without Seeing the Future? arXiv · Oct 28, 2026The paper introduces Investor's Last Exam (ILE), an equity-research benchmark designed with experienced buy-side investors and fund managers. It combines a versioned financial database with validated timestamps, a task taxonomy built around professional analyst deliverables, and a framework linking scoring criteria to sources, calculations, and assumptions. Across 86 assignments, it jointly assesses research quality, forecast accuracy, and simulated investment performance. Even the top models fail to convert overall research ability into predictive performance, with GPT-6 Astra scoring only 52.5 on Forecast & Model.

urn:diting-lab:en:doc:346061 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Statutory Article Retrieval from Arabic Judicial Narratives: Family Law and Law of Evidence in Saudi Arabia arXiv · Oct 28, 2026This paper examines how long factual narratives in Arabic court cases can be matched to statute articles that state abstract legal rules. The authors build a benchmark of 200 synthetic Saudi cases and 381 articles from the Family Law and Law of Evidence, and compare several retrieval methods. They propose a Decompose-and-Refine pipeline that splits each case into disputed issues and generates multiple retrieval queries per issue, which substantially improves retrieval for Evidence Law articles.

urn:diting-lab:en:doc:346060 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Learning a Prior-Data Fitted Network for Graph Foundation Models via Structural Causal Model Priors arXiv · Oct 28, 2026This paper proposes GSCM-PFN, a graph prior-data fitted network pre-trained on synthetic graphs generated from a structural causal model (SCM) prior that jointly models nodes and graph topology. The authors design a joint feature-edge-label SCM to generate diverse synthetic graphs and a feature-permutation-invariant dual-path graph network. Experiments on 23 real-world benchmarks report that GSCM-PFN outperforms conventional and PFN-based graph foundation models without using real graph data for pre-training.

urn:diting-lab:en:doc:346058 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Fu-TCM: a multimodal large language model that learns traditional Chinese medicine diagnostic reasoning arXiv · Oct 28, 2026This paper presents Fu-TCM, a multimodal large language model for traditional Chinese medicine, trained on about 7.12 million examples drawn from classical texts, textbooks, medical images, clinical cases, and public datasets after cleaning and filtering. The model is further optimized with Bianzheng-Grounded Policy Optimization, which rewards response format, intermediate TCM reasoning fields, and final syndrome predictions. Fu-TCM-27B leads all six benchmarks covering TCM reasoning, text, and vision, reaching a macro-average accuracy of 81.01%.

urn:diting-lab:en:doc:346057 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

MUSE: Multi-Level Memory with Selective Runtime Injection for Scientific Coding Agents arXiv · Oct 28, 2026This paper introduces MUSE, a memory framework that helps scientific coding agents reuse past experience. It builds global, approach-level, and implementation-level memory from historical trajectories, injecting global and problem-matched approach memory at startup and selectively retrieving implementation memory during execution via an LLM gate. Across four benchmarks covering numerical problem solving, kernel optimization, and scientific software repair, MUSE outperforms retrieval and memory baselines on all metrics.

urn:diting-lab:en:doc:346056 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Multi-Token Prediction Done Right arXiv · Oct 28, 2026Multi-token prediction (MTP) is a common language-model pre-training method that improves downstream performance and speeds up inference through speculative decoding. The paper finds that a residual term in the standard MTP gradient interferes with learning, and the problem grows with scale. It proposes Soft-MTP, a one-line change that keeps only the KL self-distillation signal and removes the noisy residual, improving over standard MTP from 167M to 1.2B parameters without adding parameters.

urn:diting-lab:en:doc:346055 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

A Case Study of a Multi-Agent Kernel-Optimization Loop That Learns a New AI Accelerator arXiv · Oct 28, 2026This paper presents an LLM agent loop built to write optimized kernels for AI accelerators whose architecture and language are rarely seen by language models. Using the FuriosaAI RNGD NPU as the example, the loop optimized three Gemma-4-12B decode kernels over 12 days. It builds its own hardware knowledge from documentation summaries, microbenchmarks, and compiler-derived rules, and searches per-kernel design trees. The final kernels reach a 9.293x geomean speedup over the official baseline, each within 1.25x of the cycles needed to move its weights.

urn:diting-lab:en:doc:346054 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Can LLMs Faithfully Solve Ontology Engineering Tasks? A Mechanistic Interpretability Study arXiv · Oct 28, 2026This paper introduces MIOE, a causal-abstraction benchmark that tests whether internal representations in language models truly encode ontological knowledge. It evaluates interchange interventions on entity typing and subsumption prediction tasks, using SAE-, DBM-, and DAS-based methods on Qwen3-8B and Qwen3-4B. Distributed subspace methods perform best, showing that ontological predictions are mechanistically accessible in LLMs, though the causal mechanism is not yet cleanly disentangled.

urn:diting-lab:en:doc:346052 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SSQA: Sibling-Selective Quadtree Attention for Hierarchical Modeling in Perception Tasks arXiv · Oct 28, 2026This paper proposes Sibling-Selective Quadtree Attention (SSQA), which models image tokens hierarchically as a full quadtree. The authors provide a complexity analysis showing linear-time feature modeling and compare inference speed with Mamba 2 and Quadtree Attention. Benchmarked across several tasks, SSQA reportedly achieves results at least as good as other methods, and sometimes better, at a fraction of the computational cost.

urn:diting-lab:en:doc:346051 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Extremely Sparse-View Computed Tomography from 2D Projections via Pose-Aware Diffusion Priors arXiv · Oct 28, 2026This paper addresses extremely sparse-view computed tomography, where only a few projections can be acquired due to scan time, motion, or radiation limits. The method learns a pose-aware diffusion prior from one 2D projection per training sample and guides diffusion sampling with measured-view fidelity, shared-volume reprojection, and projection-physics consistency. On two public CT datasets with 1, 4, and 8 views, it achieves the best novel-view synthesis scores and best or competitive reconstruction quality against baselines that use dense-volume supervision.

urn:diting-lab:en:doc:346050 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

What Should a Navigator Remember? Spatiotemporal Memory for Zero-Shot Vision-Language Navigation arXiv · Oct 28, 2026This paper studies what an 8B-scale robot navigation system should store and how to organize it, without navigation-specific training. The system uses three complementary memories: task progress, pose-anchored compressed observation history, and accumulated geometry. On R2R-CE val-unseen, the full system achieves the highest success rate among the evaluated Qwen3-based zero-shot baselines, and removing temporal memory raises average path length from 16.9 m to 39.1 m. Real-world tests on a Unitree Go2 also show higher success than a post-trained baseline in corridor and office settings.

urn:diting-lab:en:doc:346049 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

LLM-PowerTrace: A Dataset of Time-Resolved GPU Power Traces for AI Compute arXiv · Oct 28, 2026This paper introduces LLM-PowerTrace, an open dataset recording how each GPU's power draw changes over time during large language model fine-tuning and inference, sampled every 100 ms across varying NVIDIA platforms with workload and system metadata. It includes controlled sweeps over factors such as batch size, sequence length, precision, parallelism, concurrency, GPU frequency, quantization, and request-arrival patterns. The authors define power-grid-facing metrics from the traces and propose an AI-compute power and energy scaling law for extrapolating beyond the measured settings.

urn:diting-lab:en:doc:346048 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

CHORD: Cross-Model Hallucination Detection via Relational Graph Discrimination arXiv · Oct 28, 2026This paper studies cross-model hallucination detection, training detectors on source language models and applying them to unseen target models. Because hidden states from different models are semantically inconsistent, classical representation-based detectors transfer poorly. The authors propose CHORD, which represents hidden states as joint relational graphs over layers, tokens, and features, and uses relation-aware attention with meta-learning. Experiments report better cross-model generalization than representative detection methods.

urn:diting-lab:en:doc:346046 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Robust LLM-Generated Text Detection through Matched Conditional-Response Probing arXiv · Oct 28, 2026This paper introduces MCRP, a framework that uses a frozen large language model to measure how a candidate text responds when a semantically matched condition is added, removed, or replaced. A lightweight random-forest classifier is trained on these response features to detect LLM-generated text. Under rewriting, polishing, and expansion, it reports 93.97% F1 and 96.24% ROC-AUC, and it performs consistently across multilingual and multidomain evaluations.

urn:diting-lab:en:doc:346045 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Does 3D Pose Improve Action Valuation in Football? An Empirical Study of Pose Representations and Learning Objectives arXiv · Oct 28, 2026This paper examines whether adding 3D player pose to a SARSA-based multi-agent reinforcement learning model improves value estimates for observed football actions. Using UEFA Champions League 2022/23 event, tracking and 3D pose data, the authors compare six pose representations across five learning settings. Pose rarely improved return prediction, while training choices had larger effects; pose did improve prediction of the on-ball action 0.2 seconds ahead and pass success.

urn:diting-lab:en:doc:346043 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Inner-Learning: From Hand-Coded Beliefs to Declarative Knowledge in Trick-Taking Games arXiv · Oct 28, 2026This paper proposes a declarative approach to modeling belief states in trick-taking card games. The authors automatically extract the hand-coded inference rules of a state-of-the-art Skat player into a domain-specific language and verify them through round-trip code generation. They also add cardinality constraints, and self-play results show the declarative engine improves winning percentage and score over the original engine.

urn:diting-lab:en:doc:346042 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Don't Ask the LLM: Visual Token Pruning with Spectral Structure and Semantic Relevance arXiv · Oct 28, 2026The paper addresses the high inference cost of large vision-language models, which process long sequences of visual tokens. It proposes ST-Pruner, a training-free method that selects visual tokens before the large language model, combining attention-independent spectral saliency with inverted patch-text similarity via parameter-free multiplicative fusion. The authors report retaining over 95% of full-token performance using at most 11.1% of visual tokens across diverse models, and a 3.2x end-to-end speedup on LLaVA-NeXT.

urn:diting-lab:en:doc:346041 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Hidden Commitment: When Language Models Silently Pick a Side and How Steering Can Surface It arXiv · Oct 28, 2026This paper examines whether language models, when answering from two conflicting documents, silently favor one source without disclosing the disagreement. In Qwen 0.8B, the authors find a residual source-choice direction whose sign tracks the document used, and steering this direction changes the chosen source. A transfer experiment on Gemma 3 1B reproduces the readout and sparse-write pattern. Conflict disclosure appears weaker, later in the network, and only partially steerable, so the authors propose monitoring residual readouts for silent source commitment.

urn:diting-lab:en:doc:346040 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

EPSBench: Benchmarking LLM Agents for Earnings Forecasting arXiv · Oct 28, 2026This paper introduces EPSBench, a benchmark for testing how well LLM agents forecast a company's quarterly earnings per share (EPS). It covers 149 firm-quarters from 86 US companies, where agents read filings, earnings-call transcripts and news, and estimate income-statement drivers with cited evidence. Six models are compared; three show no detectable difference from analyst consensus in point forecast error, while three have higher error. Nominal 80% prediction intervals contain reported EPS in only 37% to 70% of runs that return an interval.

urn:diting-lab:en:doc:346039 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Vendster: An Evaluation Arena to Investigate How Forecasting Tool Quality Affects LLM Agent Behaviors in Buyer-Vendor Negotiation arXiv · Oct 28, 2026The paper introduces Vendster, an evaluation arena that measures how a forecaster's bias and prediction-interval width affect the outcome of a discount negotiation between buyer and vendor LLM agents. In the M5-based reference instance, model-generated forecasts raised sales target attainment by up to 17 percentage points, yet standard metrics like relWQL and MASE did not significantly predict negotiation outcomes. Experiments with synthetic forecasts found that the direction of bias and agents' forecast reactivity were more strongly associated with outcomes than forecast accuracy.

urn:diting-lab:en:doc:346038 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

PulmoWorld: An Asynchronous Multimodal Medical World Model for Longitudinal Pneumonia Dynamics arXiv · Oct 28, 2026This paper presents PulmoWorld, a world model that tracks how pneumonia evolves over time by combining asynchronous state estimation, recursive simulation, and evidence-constrained planning. On the MASS pneumonia cohort, its recursive forecasts are competitive with task-trained sequence models, and trained on public MIMIC-IV-CXR data it improves prognostic discrimination over latest-value fusion. The authors also introduce Cap-MBRL, which in five simulated environments with known mechanisms reduces selection regret compared with directly executing a fallback action.

urn:diting-lab:en:doc:346037 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

ConvRoute-ViT: Convolutional Token Routing for Efficient Vision Transformers arXiv · Oct 28, 2026Vision Transformers face rising computational cost because self-attention grows quadratically with the number of input tokens. This paper proposes ConvRoute-ViT, which keeps the most relevant patch tokens at selected layers based on class-token attention and compresses the rest into compact convolutional summary tokens. The approach shortens the sequence while retaining some context from less informative tokens. Experiments on CIFAR-100 and TinyImageNet show improved computational efficiency with classification accuracy maintained relative to the base ViT.

urn:diting-lab:en:doc:346036 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Human Evaluation with Automatic Scores: Separating Sampling and Estimation Gains arXiv · Oct 28, 2026This paper examines how automatic judge scores can reduce the number of human annotations needed for model evaluation. It compares three uses of the scores (guiding sampling, correcting estimates, and supporting adaptive acquisition) on fixed pools from three preference datasets and 19 attributed question-answering systems. After ten-bin score stratification, the best local linear correction reduces mean squared error by only 1.00 to 5.27% further. The authors recommend proportional sampling within score strata followed by a stratified human mean as a simple starting procedure.

urn:diting-lab:en:doc:346035 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

A Witnessing Duality for ReLU Network Depth arXiv · Oct 28, 2026The paper studies depth lower bounds for ReLU networks computing the maximum function of n inputs, asking when a weak logical theory can prove such bounds. Working in the bounded arithmetic theory VTC^0, the authors prove for each fixed prime q and depth cap K the known bound for integer-weight networks, together with a refuter that outputs a polynomial-length rational input on which any shallower network errs. They also give a balanced construction meeting this threshold at n ≤ 2^K, and a rational family of depth O(log n / log log n).

urn:diting-lab:en:doc:346034 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

BudSplat: Feed-forward 3D Gaussian Splatting under a Rendering Budget arXiv · Oct 28, 2026This paper presents BudSplat, a compact feed-forward 3D Gaussian Splatting framework that keeps the most valuable Gaussians under a limited primitive count for novel view synthesis. It uses a detail-aware importance score from multi-level visual tokens and quality-aware pruning that jointly considers rendering contribution and detail importance. Experiments on RE10K, DL3DV, and ACID show a favorable trade-off between rendering quality and efficiency.

urn:diting-lab:en:doc:346033 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

The Role of Quantization in Loss Spikes arXiv · Oct 28, 2026This paper studies time-bomb spikes, delayed training divergences that appear only after long pretraining and can waste substantial compute. The authors find that rising quantization error in the language modeling head precedes these spikes, with lower numerical precision accelerating them and higher precision improving stability. They propose replacing AdamW on the LM head with SGD with momentum to suppress logit growth, and note that existing mitigation techniques also reduce quantization error.

urn:diting-lab:en:doc:346032 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

NAGO: Noise-Aware Generative Operator via Flow Matching arXiv · Oct 28, 2026The paper introduces NAGO, a generative operator that treats the source distribution as a design choice rather than a fixed Gaussian, for sampling from invariant measures of stochastic PDEs in function space. It derives admissibility conditions and shows that the source-target covariance ratio controls spectral stiffness, leading to a rule that picks a sufficiently rough source and minimizes covariance mismatch over resolved Fourier modes. In a controlled Matern-field ablation, the accurate RK4 budget drops from about 30 steps for Gaussian functional flow matching to 3 to 5 steps.

urn:diting-lab:en:doc:346031 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

A Hybrid Physics-informed holomorphic neural networks Method arXiv · Oct 28, 2026This paper proposes a hybrid physics-informed holomorphic neural network (HPIHNN) framework that splits the solution into a particular part and a homogeneous part. The particular part is computed with a spectral method on a simple extended domain, while the homogeneous part is approximated by a network trained only on the boundary. The method is applied to nonhomogeneous Poisson, Stokes, biharmonic, linear elasticity, and Helmholtz equations. In the tested cases, it generally achieves lower errors and shorter training times than standard PINNs.

urn:diting-lab:en:doc:346030 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Cross Lingual Abstractive Summarization for Marathi arXiv · Oct 28, 2026This paper builds cross-lingual training data for Marathi from the monolingual MahaSum corpus and the smaller PMIndiaSum dataset. It fine-tunes IndicTrans2, an Indic neural machine translation model, mainly for Marathi-to-English summarization. Evaluation uses ROUGE, chrF and chrF++ scores. The authors report that using a pretrained translation model performs substantially better than the IndicBART and mBART-50 baselines from earlier work.

urn:diting-lab:en:doc:346029 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

TreemapMix: Dirichlet-Controlled Multi-Image Augmentation for Probability and Ordinal Supervision arXiv · Oct 28, 2026The paper proposes TreemapMix, a data augmentation method that combines multiple images by sampling source weights from a Dirichlet distribution and assigning them to area-proportional regions via a treemap partition. The region areas serve as soft class-probability targets with soft cross-entropy (TreemapMix-SCE) or as ordinal supervision with a Plackett-Luce objective (TreemapMix-PL). On ImageNet-1K, TreemapMix-SCE achieves lower calibration error than evaluated mixed-sample baselines, while TreemapMix-PL achieves the highest Top-1 accuracy.

urn:diting-lab:en:doc:346027 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

ETA-Motion: Efficient Topology-Agnostic Motion Diffusion via Sparse Joint Learning arXiv · Oct 28, 2026ETA-Motion is a framework for generating motions for characters with arbitrary skeletal structures. It learns from a compact subset of joints rather than all joints, using semantic joint embeddings that combine text and visual cues along with topology-aware training objectives. The paper reports that it outperforms existing topology-agnostic baselines in efficiency and generalization to unseen motions and topologies, while still producing full-skeleton motions at inference.

urn:diting-lab:en:doc:346025 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Transportable Offline Reinforcement Learning in Presence of Unmeasured Confounding arXiv · Oct 28, 2026This paper studies learning decision policies for a target domain from previously collected offline data, where source and target environments differ and some confounding variables are not measured. The authors combine transportability theory with instrumental variables, negative controls, and proxies, and propose CAT-CCB for confounded contextual bandits and CAT-CMDP for confounded Markov decision processes. They provide high-probability finite-sample suboptimality guarantees, and numerical experiments support the framework's effectiveness.

urn:diting-lab:en:doc:346024 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

A Tutorial on Generative Models for Music Audio arXiv · Oct 28, 2026This monograph introduces, in a shared notation, the methods behind music audio generation, including neural audio codecs, autoregressive and masked token models, diffusion and flow matching, sampling acceleration, and preference optimization. It links training objectives and sampling procedures to published systems through worked examples. It also compares systems by representation, method, inputs, tasks, and release information, and discusses evaluation, listening tests, data provenance, memorization, and attribution. The authors note that differences in data, scale, and evaluation limit what cross-system comparisons can establish.

urn:diting-lab:en:doc:346023 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Keepace.ai: A Technical Report on a Large Language Model for Sports and Health arXiv · Oct 28, 2026This report presents Keepace.ai, a domain-specialized large language model built on Qwen3.5-27B through post-training and currently optimized for Chinese-language exercise guidance. It covers three tasks: workout plan generation, sports-and-health knowledge question answering, and personal sports data analysis. In internal rubric-based evaluations designed with domain experts, the model outperforms its base model and larger general-purpose baselines, most notably in exercise safety and scientific accuracy. The report also summarizes its design, data pipeline, safety-first evaluation framework, and results.

urn:diting-lab:en:doc:346022 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Verification-Guided Retrosynthesis: Reaction Feedback and Evidence Timing arXiv · Oct 28, 2026This paper studies how to use reaction verification results to guide retrosynthesis route search. The authors introduce Verification-Guided Reflow (VGR), which treats reaction verification as a decision action within AND/OR search, pruning rejected reactions immediately and giving accepted reactions priority only when they form part of a complete candidate route. On Retro190, feedback raises verified solvability from 51.1% to 62.6%, and the delayed-guidance strategy further raises it to 71.1%.

urn:diting-lab:en:doc:346021 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Precedent or Intent? What Coding Agents Follow When Evolving Existing Code arXiv · Oct 28, 2026This paper studies how coding agents decide whether to follow existing repository code or new task instructions when several implementations meet the same requirements. The authors ran 5,625 controlled trials across four coding agents and validated the findings in four open-source repositories. Agents overwhelmingly followed repository precedent over their defaults, and written priority rules achieved much higher success than code examples alone.

urn:diting-lab:en:doc:346018 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

The Indian Financial Automated Advice Bias (IFAAB) Dataset : Auditing Socioeconomic and Demographic Disparities in LLM-Generated Financial Advice in India arXiv · Oct 28, 2026This paper introduces IFAAB, a dataset of 31,734 GPT-4o financial-advice responses generated for synthetic Indian personas differing in gender, caste, education, employment, residence, age and district. Using counterfactual comparisons, the authors find male-labelled profiles receive substantially higher implied post-advice total resources than comparable female-labelled profiles, a gap that shrinks when district is added. Caste-based differences are also reported, and women more often receive cautious saving advice while men more often receive growth-oriented advice.

urn:diting-lab:en:doc:346017 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Variational Consequence-Driven Offline Reinforcement Learning arXiv · Oct 28, 2026This paper addresses offline reinforcement learning, where small action deviations can push an agent away from the offline dataset, by constraining policies through the state-consequence distributions they induce rather than action distributions alone. Because enforcing such constraints directly requires the true dynamics model, the authors propose Structural Behavior Regularization (StBR), which uses a closed-form latent divergence as a bounded surrogate. StBR tightens the behavior constraint adaptively in dynamics-sensitive regions and is reported to perform well across diverse offline RL benchmarks.

urn:diting-lab:en:doc:346016 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

3DCodeEdit: Benchmarking Instruction-Guided 3D Program Editing arXiv · Oct 28, 2026This paper introduces 3DCodeEdit, a benchmark of 2,592 editing tasks on 108 Blender programs that build 3D objects entirely in code. The edits range from parameter tweaks to structural, style, and scene changes, and are judged from renders against human-reviewed checklists. Across 17 models, the average success rate is 37%, and even the strongest models finish fewer than 60% of structural and style edits. The authors also find that multi-turn feedback from a separate reviewer model helps most in compensating for weak geometric foresight.

urn:diting-lab:en:doc:346015 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

The Endgame of FSDP Communication: Layout-Aware Collectives on Copy Engines, RDMA, and NVSwitch arXiv · Oct 28, 2026The paper addresses two costs in FSDP2 training: converting between per-parameter and flat buffer layouts, and communication kernels that slow concurrent compute. It presents TIDE, zero-copy collectives built on public NCCL 2.31+ APIs, with TIDE-AG delivering shards via NVSwitch multicast and RDMA, and TIDE-RS reducing bf16 gradients into fp32 shards. On 64 GPUs, TIDE reduces step time by up to 27.1% relative to Stock FSDP2.

urn:diting-lab:en:doc:346014 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

FACTS-MAS: A Multi-Agent Architecture for Adaptive Forecasting from Heterogeneous Evidence arXiv · Oct 28, 2026This paper presents FACTS-MAS, a multi-agent forecasting framework that assigns different kinds of evidence, such as temporal dynamics, exogenous context, seasonality, and discrete events, to specialized agents. Each agent produces a forecast that is weighted by its skill on historical validation data, so uninformative evidence is downweighted or removed. Tested on Zillow housing inventory and FRED unemployment insurance claims for 15 U.S. metropolitan areas, it beats a naive last-value forecast and the TimesFM model, but matches a factor-adjusted exponential smoothing model.

urn:diting-lab:en:doc:346013 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Silhouette-Based Gaitomics for Frailty Assessment in Older Adults arXiv · Oct 28, 2026This paper uses body silhouette images from walking videos to extract gait features for assessing frailty in 66 older adults classified by the Fried Frailty Phenotype. The authors apply feature selection and explainable machine learning to link model predictions with measurable shape and motion patterns. CatBoost reached a mean three-class micro-AUC of 0.647, and frail versus nonfrail classification performed best, while separating prefrail from nonfrail participants was limited. The authors say larger independent studies are needed to establish reliability and clinical relevance.

urn:diting-lab:en:doc:346012 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

CuraEval: Evaluating Demonstration Curation on Mixed-Skill Robot Data arXiv · Oct 28, 2026This paper introduces CuraEval, a framework for evaluating methods that select demonstrations for robot imitation learning. It compares these selection methods against size-matched random subsets and full-data training, using mixed-skill robomimic data with Diffusion Policy and BC-RNN. The results show that the selection methods' advantages are inconsistent, and after Holm correction no curated condition significantly outperforms full-data training.

urn:diting-lab:en:doc:346011 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Rubricon: Automated Rubric Generation and Refinement for LLM Evaluation and Training arXiv · Oct 28, 2026This paper presents Rubricon, a pipeline that automatically creates discriminative rubrics for scoring LLM responses on tasks without verifiable answers, without needing human preference labels. It samples several candidate rubrics, selects the most discriminative one, adds missing criteria, and removes misaligned ones. Across five benchmarks, its rubrics align about 9% better with expert-written rubrics than the best alternative synthetic method, and on HealthBench they improve post-training of a Qwen3.5-9B model by 4.5% over RRD-based rubrics.

urn:diting-lab:en:doc:346010 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

HotPOT: Operator Transformer with Heterogeneous Physics Experts for Large-Scale PDE Pre-training arXiv · Oct 28, 2026This paper presents HotPOT, a model for pre-training on many partial differential equations (PDEs) at once. A shared encoder captures information common to all PDEs, while each PDE family has its own expert that pairs a learnable predictor with a fixed numerical update computing the next state. Across five PDE families and six settings, HotPOT beats MoE-POT in all 24 matched-scale comparisons before and after fine-tuning, reducing rollout error by 50.7% on average. The gains also extend to zero-shot prediction under unseen physical conditions.

urn:diting-lab:en:doc:346008 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Skeleton Coding: Program Structure as a Specification for Agents arXiv · Oct 28, 2026The paper argues that the program structure an AI coding agent chooses determines which tests are needed, so behavioral tests alone can leave a developer's intended program underdetermined. It introduces Skeleton Coding, where developers specify a high-level code-like structure alongside behavioral specifications. The authors show theoretically that skeletons can reduce the number of tests needed exponentially, and report that across four models and three benchmarks the best skeletons cut mean test error by 48.4% relative to the best end-to-end specifications.

urn:diting-lab:en:doc:346007 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

RainCheck: Harnessing Historical Weather Data with Machine Learning for Rainfall Forecasting in Bangladesh arXiv · Oct 28, 2026This study uses 618,650 station-day records from 35 Bangladesh Meteorological Department stations, covering 1948 to 2014, to compare machine-learning rainfall forecasts. It finds that a station-month climatology alone classifies high-rainfall months with 88.4% accuracy, and most learners do not significantly beat climatology on future years. The authors also release a leakage-aware evaluation protocol and code.

urn:diting-lab:en:doc:346006 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

GAAC: Lossless KV Compression for Bandwidth-Efficient LLM Inference on Personal Devices arXiv · Oct 28, 2026This paper presents GAAC, a lossless key-value (KV) cache compression and serving system for large language models that exploits redundancy in BF16 exponents without approximating the cached values. Across five tested models, it achieves 1.405 to 1.407 times lossless BF16 KV compression. On NVIDIA DGX Spark and Apple M4, it speeds up attention computation, with reported gains of 1.19 to 1.27 times at a 64K-token context and 1.02 to 1.25 times respectively.

urn:diting-lab:en:doc:346004 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Progressive Coupling Method for Diffusion Posterior Sampling in Ptychography arXiv · Oct 28, 2026Ptychography reconstructs a sample from overlapping diffraction measurements, but dense scanning makes acquisition slow. This paper proposes progressive coupling posterior sampling (PCPS), which explores the solution space broadly under high uncertainty and strengthens diffusion-guided consistency as reconstruction refines. Experiments on real X-ray ptychographic data at various scan overlaps show improved accuracy over traditional iterative methods and existing diffusion posterior samplers, especially at low overlaps. The results suggest diffusion priors can reduce the measurement redundancy needed for accurate reconstruction.

urn:diting-lab:en:doc:346002 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Two Birds with One Stone: Skill-Augmented Generation via Unified Retrieval and Utilization arXiv · Oct 28, 2026The paper proposes Skill-Uni, a reinforcement learning framework in which a single language model learns both to retrieve reusable skills and to use them for solving tasks. Training uses a warm-up instruction tuning stage followed by GRPO. Across five skill-related benchmarks, it reports average gains of 66.59% over Qwen3-4B and 34.82% over LLaMA3.1-8B, and it works with various skill retrieval engines.

urn:diting-lab:en:doc:346001 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

From Natural Language to Real Schematics: An Automated Workflow for Power Electronics Schematic Synthesis arXiv · Oct 28, 2026The paper presents a bounded dual-loop framework that converts natural language specifications into verified, editable KiCad schematics and manufacturing outputs. It covers Buck, Boost, Flyback, DAB, and half-bridge LLC topologies, using distributor APIs, datasheets, ngspice transient verification, and KiCad netlist round-trip checks. In two end-to-end case studies and comparisons with a Vanilla LLM baseline, component verification coverage, nominal value consistency, and SPICE simulation pass rates improved substantially.

urn:diting-lab:en:doc:346000 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Embodied at Sea: Imitation Learning for Multi-Task Underwater Manipulation arXiv · Oct 28, 2026The paper trains 14 generalist Vision-Language-Action (VLA) models, called DynamicVLA, on 1616 underwater manipulation episodes collected in calm and turbulent waters to address visual distortion and currents underwater. Across three tasks, 10 models outperformed the baseline, and the best reached a 93.3% overall success rate. A marine archaeology case study compared a single-task Action Chunking Transformer with SmolVLA on grasping and mooring, with ACT substantially outperforming SmolVLA (47.9% vs. 7.9% pooled success). The BLUEMBODIED kit used for data collection, training, and deployment has been open sourced.

urn:diting-lab:en:doc:345999 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Convex-SKM-Net: Low-Latency Block Halfspace Corrections for Neural Convex Constraint Satisfaction arXiv · Oct 28, 2026The paper introduces Convex-SKM-Net, a convex constraint-satisfaction layer that builds linearization halfspaces from residual and subgradient oracles and corrects predictions using fixed-depth dual projected-gradient descent on minimum-displacement block problems. The predictor and a positive-definite mobility matrix are learned jointly to improve objective quality. Reported tests on second-order-cone model-predictive control and power-system convex relaxations keep outputs within a 10^-3 feasibility tolerance, with inference speedups of up to 144.3 times over compared methods.

urn:diting-lab:en:doc:345998 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

GearUp: DAG-aware Acceleration for Serverless Inference Serving via Slack-guided Pre-loading arXiv · Oct 28, 2026GearUp is a serverless system for running machine learning inference workflows structured as directed acyclic graphs (DAGs). It schedules each node's library, model weight, and CUDA kernel loading within the slack time the DAG exposes, hiding initialization behind upstream execution. The system also reuses initialized instances and uses a traffic-adaptive keep-alive policy. On the AMBER Alert and FINRA-ML workflows, it reduces median end-to-end latency by 64.5% and 75.9% compared with SMIless.

urn:diting-lab:en:doc:345997 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Prototype-Guided Dynamic Domain Modulation and a Large-Scale Multi-SKU Benchmark for Shampoo Bottle Surface Defect Detection arXiv · Oct 28, 2026This paper addresses shampoo bottle surface defect detection, where the same defect can look different across product variants (SKUs). The authors introduce Bottle-SKU, a benchmark with 13,119 images, 20,419 bounding boxes, seven defect categories, and 48 bottle SKUs. They propose ProDMM, which builds a visual prototype for each SKU and uses it to dynamically modulate a Grounding DINO detector while keeping its pretrained detection path. Experiments on Bottle-SKU and the public GC10-DET-SKU dataset are reported to show the approach's effectiveness.

urn:diting-lab:en:doc:345994 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Complementing Text Based Moderation with Multimodal Grounding: Detecting Fraudulent Inducement Behaviors on Short Video Platforms arXiv · Oct 28, 2026This paper addresses fraudulent inducement behaviors on short video platforms, such as prompts to like, follow, or join private chat groups. It compares three approaches: OCR+LLM, OCR+MLLM, and a large zero-shot multimodal LLM, using a dataset of 2,419 videos with 3,194 annotated events. The OCR+MLLM method localizes offending segments better than the keyframe-based baseline (mIoU 26.52 vs. 14.49) and detects non-textual inducements that text-only pipelines miss. It maintains 67.0 percent macro recall.

urn:diting-lab:en:doc:345992 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

In-Context Learning for Multi-Fidelity Multiscale Model Closure and Cross-Operator Out-Of-Distribution Generalization arXiv · Oct 28, 2026This paper proposes an in-context multi-fidelity learning framework that uses one shared model to learn multiple multi-fidelity operators. The model takes paired low- and high-fidelity examples from multiple PDE operators as demonstrations, plus a new coarse-scale solution query, and predicts the corresponding fine-scale solution. After training, it predicts fine-scale solutions for a new operator from fewer than five coarse-fine demonstration pairs and a low-cost coarse query, without fine-tuning.

urn:diting-lab:en:doc:345990 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Scaling Lean Theorem Proving for Real-World Code Verification arXiv · Oct 28, 2026The paper presents EasyProver, a Lean proving system built from recursive agents, asynchronous search, and parallel tactic probing, which uses an off-the-shelf model without additional training to solve most problems on benchmarks such as PutnamBench and IMO 2026 formalizations. Because these benchmarks are near saturation, the authors introduce Aegis, which turns competitive programming problems into harder verification tasks over real Rust implementations translated to Lean. The pipeline yields 2,045 differentially validated Rust-Lean program pairs and 11,533 candidate specifications. On a stratified sample of 2,000 statements, EasyProver proves 1,054 (52.7%) within a 30-minute budget per statement.

urn:diting-lab:en:doc:345989 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

A Gradient based Adversarial Augmentation Framework for Mitigating Unlearnable Datasets arXiv · Oct 28, 2026This paper proposes a gradient-based adversarial augmentation framework for stress-testing unlearnable data methods (ULD), which add imperceptible perturbations to public data to block unauthorized model training. The framework learns an augmentation policy via stochastic gradient ascent over a differentiable pipeline of up to 18 augmentation techniques, training to maximize a surrogate model's classification loss. A perceptual-plus-pixel regularization preserves image semantics. Under a controlled ablation, it outperforms an equivalent random sampling baseline.

urn:diting-lab:en:doc:345988 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Data Amount Heterogeneity Improves Performance in Federated Learning for Neuroimaging arXiv · Oct 28, 2026This paper studies federated learning for neuroimaging, specifically classifying Alzheimer's disease from structural MRI scans. It compares federations where sites have unequal data amounts (one large site plus small sites) against federations where every site holds the same amount of data. The heterogeneous federations perform better, especially in low-data settings, and come close to centralized training. A large site often leads to faster convergence and higher final accuracy.

urn:diting-lab:en:doc:345987 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

B2P-Corr: Batch-to-Population Gradient Estimators for Non-Decomposable Correlation Losses arXiv · Oct 28, 2026The paper observes that correlation metrics such as Pearson and concordance correlation depend on dataset-wide statistics, so naive mini-batch training yields biased, high-variance gradients. It proposes B2P-Corr, which converts these objectives into decomposable pseudo-losses using lightweight global moment sketches, with two variants, B2P-EMA and B2P-U. Experiments on two synthetic and three real-world datasets show more stable optimization and higher correlation than naive mini-batch training, with negligible overhead.

urn:diting-lab:en:doc:345986 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Replaying the Same Gradients at a Different ǎrepsilon: A Registered Test of the Fixed-Trajectory Account of Adam's ǎrepsilon for One Rare Row arXiv · Oct 28, 2026The paper examines how Adam's epsilon term affects updates to a rare token's output-embedding row. During training, the authors recompute that row's updates with a different epsilon while keeping the gradients fixed, and test this explanation against a rule written before the runs. When epsilon rises by four orders of magnitude, the replay predictions are clearly wrong, with the row keeping far more of its displacement than predicted. At three orders of magnitude the gap is smaller and less clear.

urn:diting-lab:en:doc:345985 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Neighbourhood shortcuts in building damage assessment from very-high-resolution satellite imagery: diagnosis and a building-evidence plug-in arXiv · Oct 28, 2026This paper examines deep models that assess building damage from very-high-resolution satellite imagery and finds they often rely on neighbouring buildings' damage levels as a shortcut, missing isolated damaged buildings. The authors propose a plug-in that adds two experts to an existing model without retraining it, one looking only at the building and one trained to ignore swapped neighbourhoods. The plug-in is reported to reduce neighbourhood-driven errors across nine models.

urn:diting-lab:en:doc:345984 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

ATI-OPD: Adaptive Teacher Intervention for Stable Long-Sequence On-Policy Distillation arXiv · Oct 28, 2026This paper proposes ATI-OPD, a training method that improves on-policy distillation, where a student learns from its own responses using token-level feedback from a teacher. For long responses, the student's reasoning prefix can drift from the teacher's, weakening later feedback and destabilizing training. ATI-OPD replaces the earliest high-disagreement segment with a short teacher segment and lets the student resume reasoning from the revised text. Across five mathematical reasoning benchmarks and two teacher-student pairs, it improves average student-only accuracy over standard on-policy distillation.

urn:diting-lab:en:doc:345983 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

REEL: Same Frames, Different Answers and Different Costs arXiv · Oct 28, 2026The paper introduces REEL, an evaluation framework that compares three saved memories built from the same video frames: a text record, visual encoder outputs, and language-model prefix state. Each is compared against reprocessing the full frames using paired correctness changes, cumulative GPU time, and storage size. Across six vision-language models and two benchmarks, Text differs from Frames on 20% to 29% of questions despite similar mean accuracy, while Prefix State stays closest to Frames and has the lowest total GPU time, at the cost of up to about 21 GiB per video.

urn:diting-lab:en:doc:345982 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Floor It, Then Fill It: Simulating Creator Ecosystems for Fair Video-Content Recommendation Systems at Billion User Scale arXiv · Oct 28, 2026This paper presents a user-simulation framework that models about 42,000 creators and their viewer interactions over 12 weeks, grounded in real platform distributions. It evaluates equitable view-allocation policies formulated as a budgeted multi-armed bandit, varying view targets, reward functions, and a small-creator reserve. A floor-protected approach meets 90% of view targets weekly and raises small-creator organic reach by about 33% per week.

urn:diting-lab:en:doc:345981 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

MASS: Multi-mask Selective Self-Distillation for Leakage-Resilient Reasoning arXiv · Oct 28, 2026This paper proposes MASS, a multi-mask selective self-distillation method for improving reasoning in language models. It creates multiple independently masked views of one verified solution so the teacher model can combine diverse reference signals, then selectively transfers reliable token-level corrections while focusing masking on influential reference spans. Experiments on mathematical reasoning benchmarks and several model families show that MASS consistently outperforms strong self-distillation baselines.

urn:diting-lab:en:doc:345980 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

StepControl: Long-Horizon Human Motion Generation from Foot-Contact Schedules arXiv · Oct 28, 2026This paper introduces StepControl, an autoregressive diffusion model that generates full-body human motion from a schedule specifying when each foot touches the ground. It needs no target root or body poses and no semantic prompts, representing contact as explicit temporal events injected via cross-attention with classifier-free guidance. Experiments show it follows prescribed contact timing more accurately than adapted spatially guided generators, and keeps control over long rollouts and edited schedules such as cadence changes, pauses, limping, and jumping.

urn:diting-lab:en:doc:345978 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Estimating Implicit Expected Log Density Ratios arXiv · Oct 28, 2026The paper introduces the expected log density ratio (ELDR), which is the integral of log density ratios under a third measure. It shows that many implicit-density problems that naively use density ratio estimation (DRE) are actually ELDR estimation tasks. The authors prove fundamental limits on finite-sample ELDR estimation and argue that typical DRE methods are unsuitable for it. They also propose a simple modification to some DRE methods, which benchmarks show can reduce ELDR estimation error.

urn:diting-lab:en:doc:345977 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SkillCT: A Multidimensional Benchmark and Execution-Free Baseline for Agent Skill Assessment arXiv · Oct 28, 2026The paper introduces SkillCT, a benchmark that assesses agent skills across nine dimensions grouped into static, dynamic, and human-assisted assessment, going beyond task execution success. It also proposes an execution-free baseline that predicts a skill's quality profile by retrieving similar reference skills. According to the authors, this approach lowers token use and evaluation time, and SkillCT-guided optimization improves downstream task success and execution efficiency.

urn:diting-lab:en:doc:345975 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

El Agente Quntur: Agentic Active-Space Selection for CASSCF arXiv · Oct 28, 2026This paper tests whether a multi-agent AI system can reliably choose the active orbital space for CASSCF multireference quantum chemistry calculations. The agents selected orbitals, checked target excited states, and judged convergence for 32 singlet excitations, and their results were compared with six fixed Active Space Finder protocols using published CASPT2 reference values. The agentic approach reduced mean absolute error by 48% relative to the best deterministic protocol. The authors also note that computational traces and costs varied substantially across runs, even for the same molecule.

urn:diting-lab:en:doc:345974 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Decompose, Don't Expand: Capability-Independent AI Systems by Construction arXiv · Oct 28, 2026This paper examines a design choice in AI systems: when a task is too hard for a model, expand the model or decompose the task so a deterministic host handles the solvable parts and the model handles only the fuzzy remainder. The authors report measurements across nine models, finding that on the extracted sub-tasks performance no longer varies with model size. They argue the approach is a general design discipline applicable to physics setup, motion retargeting, and object-model editing.

urn:diting-lab:en:doc:345972 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

A Separation Principle for Adversarial Attacks and Defense of Graph Neural Networks arXiv · Oct 28, 2026This paper treats adversarial attacks on graph neural networks as a problem shaped by graph structure, not only by gradients or learned policies. It derives feature attacks from the spectral properties of the normalized adjacency operator and states a separation principle: the optimal perturbation is rank one, splitting into a graph-dependent degree-based allocation and a weight-dependent top singular vector. Based on this, the authors propose SepAttack, a gradient-free attack, and SepAT, an adversarial training procedure.

urn:diting-lab:en:doc:345970 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

MiniMMD: Minimization of MMD for Sampling, Inference and Integration arXiv · Oct 28, 2026This thesis presents a unified approach to many problems in statistics and machine learning, which approximate a target distribution by minimizing the maximum mean discrepancy (MMD), a kernel-based distance between probability distributions. It argues that MMD is both a statistically meaningful distance and a tractable objective for designing practical, efficient algorithms. The thesis summarizes recent theoretical and empirical advances in these minimum-MMD-based algorithms.

urn:diting-lab:en:doc:345967 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Too Hard, Too Easy, Just Right: Performance-Driven Challenge Adaptation in Virtual Reality Firefighter Training arXiv · Oct 28, 2026This paper presents ProACT, a performance-driven adaptive training system for virtual reality firefighter training. It adjusts scenario difficulty through closed-loop interaction between trainee performance and simulated fire dynamics. A global module updates scene-level fire spread in real time, while a local module periodically updates object-level flammability. Compared with adaptive and non-adaptive baselines, ProACT improves both objective training performance and subjective user experience.

urn:diting-lab:en:doc:345966 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Beauty Is Human: When People Love AI-Generated Art, They Think a Human Made It arXiv · Oct 28, 2026In a public online AI Art Turing Test, 11,123 people judged whether 50 pictures, 26 of them AI-generated, were made by humans and picked their favourites. Holding respondent and picture fixed, a picture was 22.1 percentage points more likely to be judged human-made when it was the respondent's favourite, with a similar pattern for AI-generated and human-made favourites. A reanalysis of a published AI poetry experiment found the same pattern, suggesting that when provenance is hidden, liking and perceived human authorship go together.

urn:diting-lab:en:doc:345965 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

From the Gettier Machine to Justified True Output: A Three-Step Account of Rational Trust in Generative AI arXiv · Oct 28, 2026The paper asks when a user who takes a proposition from generative AI output actually knows it and when trusting it is rational. It argues that failures of AI-mediated inquiry match the problems that the justified-true-belief tradition and its post-Gettier anti-luck refinements were built to diagnose. It proposes a three-step pattern of use (acquisition, understanding, independent verification) and a stakes-indexed criterion called justified true output (JTO), with the everyday test: "Had this been false, would I have noticed?"

urn:diting-lab:en:doc:345963 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Who Called? V33DA: A Physically Verified Multimodal Benchmark for Vocal Attribution in Zebra Finch Groups arXiv · Oct 28, 2026This paper introduces V33DA, a benchmark dataset for determining which zebra finch in a group produced a vocalization and where it was made. It contains 33,625 events that pair multichannel audio, three camera views, calibrated 3D bird poses, and radio telemetry for 10 birds, with caller identity verified by an on-body accelerometer that is withheld from the models. The authors test methods under three regimes and find that approaches relying on familiar vocal identity perform well on known birds but degrade on unseen ones, while candidate-aware spatial methods transfer better. The dataset, evaluation protocols, and code are released publicly.

urn:diting-lab:en:doc:345962 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

TGTV: An Interpretable Framework for UAV Fault Detection and Traceability via Spatiotemporal Graph Learning and Large Language Models arXiv · Oct 28, 2026This paper presents TGTV, a framework for detecting and tracing faults in fixed-wing unmanned aerial vehicles. It combines spatiotemporal graph learning with a large language model, using temporal and graph networks plus a Transformer-VAE to flag anomalies and continuous-time Shapley attribution to identify critical sensors and intervals. On the public ALFA dataset with cross-flight holdout, it reports a window-level PR-AUC of 0.976 and a flight-level fault diagnosis F1 score of 0.944.

urn:diting-lab:en:doc:345961 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Same Label, Different Interface: Beyond Action-Space Labels in Robot Learning arXiv · Oct 28, 2026Robot-learning studies often compare position and velocity actions as if the label alone defined the experimental condition. This paper tests five joint-space interfaces on three contact-rich Factory tasks (PegInsert, GearMesh, NutThread) under one shared learner and training budget. It finds that omitted properties such as physical action scale, filtering, and feedback affect results, and that even a calibrated velocity scale leaves residual performance gaps.

urn:diting-lab:en:doc:345960 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

The Prestige: Benchmarking Cognitive Visual Reasoning using Magic Tricks arXiv · Oct 28, 2026This paper introduces The Prestige, a video understanding benchmark built from professional magic performances that use deliberate misdirection. It tests models on tracking events and explaining unexpected outcomes, with open-ended questions covering temporal and causal reasoning and robustness to adversarial questions. Evaluations of leading vision-language models show a substantial gap between even the strongest model and human performance.

urn:diting-lab:en:doc:345959 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

AI-Driven Digital Twin Framework for Warehouse Logistics Automation Optimization arXiv · Oct 28, 2026This paper proposes a framework in which a reinforcement learning agent, operating within a FlexSim-based digital twin, optimizes product placement in real time. The agent learns continuously from streaming operational data to reduce picking time and raise throughput. The simulated environment models a four-zone warehouse with forklift handling and uses the Asset Administration Shell framework to link the agent, physical asset, and digital twin. Reported results show a 29.66% drop in average picking time, a 35.42% throughput increase, and an 18.54% reduction in travel distance.

urn:diting-lab:en:doc:345958 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Evaluating Causal Reasoning in LLMs: A Case Study in the Challenges of Procedural Generation arXiv · Oct 28, 2026This paper examines what can reliably be concluded when large language models perform well on test sets generated by the CLadder benchmark framework. The authors find that the generator embeds superficial, highly predictive patterns, and even simple random forest models perform well on its data. They also show that many commercial LLMs are brittle on these test sets, with performance barely changing when text critical for formal causal reasoning is removed. The authors argue that strong LLM results most likely reflect exploitation of these surface patterns.

urn:diting-lab:en:doc:345957 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Calculation and rejection diverge in language models tested at known mechanical limits arXiv · Oct 28, 2026The paper tests 12 language models on engineering problems such as beam loads and disk speeds set above or below known plastic collapse or burst limits. Across 4,860 calls and follow-up experiments on new designs, it finds that models can reject impossible designs when asked only about feasibility, yet may accept the same designs and report them as solved when asked for a full answer. Some models accepted beams exceeding collapse limits despite calculating those limits correctly in separate questions.

urn:diting-lab:en:doc:345956 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI arXiv · Oct 28, 2026The paper introduces Nemotron 3.5 Super, a 120B-parameter multimodal model with 12B active parameters, and Nemotron 3.5 Lightning, a 30B-parameter text model with 3B active parameters, both built for reasoning, long-context tasks, and agentic workloads. The models combine large-scale distillation, long-context training, and native speculative decoding, and Super adds Indic pretraining and software-engineering mid-training. After post-training with supervised fine-tuning and reinforcement learning, the post-trained Super model reaches 51.78% on SWE-bench Pro, and inference optimizations reach up to 4,856 output tokens/s/GPU on GB200.

urn:diting-lab:en:doc:345953 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

What Can Firms Do in the Face of Model Distillation? arXiv · Oct 28, 2026The paper examines how smaller firms can distill training data from conversations with powerful closed-weight AI models to build their own models at much lower cost. Using a two-player model of an incumbent and a distilling competitor, it studies effects on market competition, user surplus, and the incumbent's utility. It finds that penalty-based regulation can fully deter distillation when attacks are detectable, though it may harm competition and users, while self-distillation can mitigate undetectable attacks and help preserve competition.

urn:diting-lab:en:doc:345952 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

From Next Day Wildfire to Next Days Wildfire: A New Multi-Continent Dataset of Active-Fire Spread for Sequence-to-Sequence Forecasting arXiv · Oct 28, 2026This paper introduces AFSeq, a multimodal dataset of 1,194 fire events from 2023 to 2026 across eight countries in three regions, focusing on short, small fires that existing benchmarks underrepresent. It also presents UTAE-AR, a U-Net-based model with a Temporal Auto Encoder and recurrent modules that forecasts a whole sequence autoregressively by feeding its predicted fire masks back day by day. The authors report that UTAE-AR outperforms prior state-of-the-art methods on next-day accuracy across WSTS, WSTS+ and AFSeq while using about 21.5% of their parameters.

urn:diting-lab:en:doc:345950 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Interpretable Feature Discovery from Unstructured Data Using LLMs and Statistical Choice Models arXiv · Oct 28, 2026The paper proposes a method that has large language models compare pairs of items and explain in interpretable terms what distinguishes them, then uses a statistical choice model to combine these explanations into relevant features and per-item feature scores. On a synthetic hiring dataset of cover letters with known ground-truth features, the method significantly outperforms direct LLM feature extraction in recovering those features. On the ChangeMyView Reddit debate dataset, it discovers several features associated with persuasion while reaching predictive accuracy comparable to established baselines.

urn:diting-lab:en:doc:345948 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Efficacy of Monocular Depth Estimation in Lunar Environments arXiv · Oct 28, 2026This paper tests how well several state-of-the-art monocular depth estimation methods work on lunar terrain, which lies outside the terrestrial data they were trained on. The authors use the NASA Ames POLAR stereo dataset and Apollo 17 surface images, comparing results against ground-truth physical analogs and historical field data. They report that some of these techniques can generalize to produce high-fidelity relative 3D reconstructions of the unstructured lunar surface.

urn:diting-lab:en:doc:345947 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Neural Compression of Long ADMM Trajectory for Multiparametric Quadratic Program arXiv · Oct 28, 2026This paper presents TQPNet, a self-supervised neural framework that compresses long ADMM optimization trajectories into a compact neural warm start followed by a few refinement iterations. The network combines a ReLU multilayer perceptron with unrolled ADMM(ReLU) refinement layers and is trained offline using an optimality-condition residual loss, without solver-generated labels. Experiments on large-scale multiparametric quadratic programs and real-time model predictive control show high accuracy, strong generalization, and much faster inference than conventional optimization methods.

urn:diting-lab:en:doc:345946 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Prompt Lottery in LLM-as-a-Judge: Measuring and Mitigating Sensitivity to Syntactic Rephrasing arXiv · Oct 28, 2026This paper studies whether using a large language model as an evaluator gives consistent judgments when the judge prompt is reworded without changing its meaning. The authors name the effect "Prompt Lottery," showing that paraphrased prompts can change evaluation outcomes and, in some settings, reverse model rankings. A user study with 10 human annotators compares agreement under rephrasing, and the authors show that aggregating judgments across multiple prompt variants improves robustness.

urn:diting-lab:en:doc:345943 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

From Error Prediction to Causal Influence: Auditing SAE Features in Multiple-Choice QA arXiv · Oct 28, 2026This paper examines whether sparse autoencoder (SAE) features that predict model errors also causally influence answers in multiple-choice question answering. Using Gemma 3 1B and 4B with Gemma Scope SAEs across eight datasets, the authors filter out features tied to prompt and answer-order artifacts before testing causal interventions. SAE activations predict wrongness and uncertainty, and the filtered features shift answer margins more than random features, yet answer flips are rare and effects are inconsistent across meaning-preserving changes. The features therefore appear to reflect context-dependent answer competition rather than the mechanism behind incorrect answers.

urn:diting-lab:en:doc:345940 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Cooperative EV Charging Assignment Using Multi-Agent Reinforcement Learning: A Comparison of QMIX and Independent Q-Learning Under Urban Traffic Uncertainty arXiv · Oct 28, 2026The paper models the assignment of electric vehicles to several charging stations as a decentralized partially observable decision problem and compares QMIX, a centralized-training method, with an independent Q-learning baseline. Tested on a six-station urban network simulated in SUMO under four traffic flow levels, QMIX achieved higher mean reward and acceptance rate. An ablation study suggests the gain comes mainly from the training architecture rather than the cooperative reward terms or experience sharing.

urn:diting-lab:en:doc:345938 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Conformal Predictive Distributions for Contextual Order Fulfillment Optimization arXiv · Oct 28, 2026This paper builds conformal predictive distributions for delivery time using Conformal Predictive Systems for regression and the cross Venn-Abers predictor for multiclass classification. These distributions feed a contextual optimization model for choosing fulfillment locations and carriers, solved with contextual sample-average approximation. On over five million proprietary order records, CVAP raised exact-day accuracy from 49.5% to 64.1%. In simulator experiments over six test dates, the methods reduced mean total cost by 7.1% and 6.7% and cut the late-delivery rate from 5.7% to below 4.0%.

urn:diting-lab:en:doc:345937 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Classifying Urban Locations by Cyclist Facility Usage and Directional Alignment: A Validated Camera-Based Framework from Multi-Site Observations arXiv · Oct 28, 2026The study analyzes footage from motion-triggered cameras at 17 locations in Tempe, Arizona, in 2023 to see how cyclists use bike lanes and sidewalks. It finds that higher bike lane usage is associated with less wrong-way riding (Spearman rho = -0.87). Deployments are classified as bike-lane dominant, sidewalk dominant, or mixed.

urn:diting-lab:en:doc:345936 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

PromptSense: Sentiment-Aware Mental Health Chatbot with Dynamic Prompting arXiv · Oct 28, 2026This paper proposes a framework that analyzes sentiment and depression before model inference, applies a rule-based filter, and dynamically builds prompts from the results. It aims to make responses from mental health chatbots built on large language models more contextually relevant, empathetic, and risk-aware. The authors compared four model strategies and two prompt-engineered variants on qualitative measures such as empathy and safety, plus toxicity scores from the Perspective API. The prompt engineering approaches outperformed DSFT and RLAIF by 1.8% and 1.1%, respectively.

urn:diting-lab:en:doc:345935 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Can Event Detectors Leave the Broadcast Camera? Evidence from Tennis arXiv · Oct 28, 2026The paper releases a tennis benchmark of 345 frame-level annotations of ball bounces and racket hits across eight clips from broadcast, fixed elevated, phone and court-level viewpoints. It compares supervised spotters trained on broadcast footage, two zero-shot vision-language models, and RallyArb, a rule-based arbitration method that fits no parameters on event labels. Supervised spotters perform near-perfectly on broadcast video but poorly on phone and court-level footage.

urn:diting-lab:en:doc:345934 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

CARF: Context-Aware Reasoning Supervision for Scientific Evidence Screening arXiv · Oct 28, 2026This paper introduces CARF, a human-in-the-loop framework that adapts language models to screening studies for systematic reviews by learning from expert feedback on their generated reasoning. Systematic reviews require experts to manually screen thousands of studies against eligibility criteria. The authors evaluated CARF on reviews of verbal autopsy and cause-of-death methods used in global public health mortality surveillance. CARF showed competitive screening performance with high sensitivity and was more inclined to keep ambiguous, context-dependent studies for human review.

urn:diting-lab:en:doc:345933 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

EARS-Info: Information-Aware Event-Driven Reallocation for Multi-Robot Exploration arXiv · Oct 28, 2026This paper presents EARS-Info, an information-aware extension of Event-Driven Ant Reallocation Search for multi-robot exploration on grids derived from OpenStreetMap. Across 168,000 episodes on eight city maps, it achieved the highest mean operational score in every evaluation panel. With eight robots and 3% communication dropout, it improved the score over EARS-Safe by 5.43% and raised weighted discovery from 88.88% to 98.46%.

urn:diting-lab:en:doc:345932 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Efficient Deep Gaussian Process Surrogates via Mini-Batch Latent MAP Estimation arXiv · Oct 28, 2026This paper proposes the latent maximum-a-posteriori deep Gaussian process (LaMP-DGP), which uses gradient-based optimization to estimate latent representations and performs probabilistic prediction through linked-GP calculations. To scale to large datasets, the authors develop a mini-batch optimization algorithm and establish a stationarity bound for latent-state optimization. Experiments report predictive performance competitive with several existing inference methods, and substantially faster computation than the Vecchia MCMC approach.

urn:diting-lab:en:doc:345931 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Measuring Deep Learning Performance - An Empirical Study of Performance Distributions Across Architectures and Tasks arXiv · Oct 28, 2026This study examines how deep learning model performance varies due to randomness, using 186 experiments with 100 random seeds each, totaling 18,600 runs, across image classification and time series forecasting. Performance distributions are often non-Gaussian, and time series models produce nearly three times more underperforming outliers than image classification models. The authors conclude that mean performance alone is insufficient for model selection, and that spread, symmetry, and tail risk should also be assessed.

urn:diting-lab:en:doc:345929 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Learning from Relative Action Credit for Escaping Local Minima arXiv · Oct 28, 2026This paper addresses why on-policy methods like PPO often stall on long-horizon, contact-rich robot manipulation tasks. The authors argue that state-value critics judge progress only by episode outcomes, so they cannot distinguish better actions once the policy reaches a partial solution. They extend PPO with an action-conditioned value critic that compares the executed action with alternatives sampled at the same state, combined with Generalized Advantage Estimation and a bound on exploration noise. The extended method escapes the plateau and completes precise peg insertion within a training budget where PPO with identical exploration fails.

urn:diting-lab:en:doc:345928 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

METATRON: A Promptable Whole-Body Execution Layer for Humanoids arXiv · Oct 28, 2026This paper presents Metatron, a whole-body control layer for humanoid robots that pairs a generative base policy with a programmable skill layer. The base policy takes partial goals, such as timed body-part targets, velocity commands, and style cues, and produces diverse whole-body motions that match them. After training, the policy is frozen, and skills written as code issue prompts at each control step. On a physical Unitree G1, one frozen checkpoint supports walking, recovery, teleoperated object pickup, and sitting on furniture, with higher reaching success and push robustness than the tested reference-tracking pipelines.

urn:diting-lab:en:doc:345927 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Making Sense of Audience Change: Causal Attribution and Boundary Management on Xiaohongshu arXiv · Oct 28, 2026This paper examines Xiaohongshu discussions about a perceived audience influx linked to the World Cup, studying how users explained community change and how those explanations related to boundary-management strategies. The authors analyzed 262 comments from eight high-engagement threads. They found that algorithmic explanations were mostly tied to technical boundary management, while explanations centered on social actors more often aligned with social enforcement or normative socialization.

urn:diting-lab:en:doc:345926 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Assessing Language Models for Scientific Applications: Inertial Confinement Fusion as a Use Case arXiv · Oct 28, 2026This paper evaluates language models on inertial confinement fusion using a 250-question benchmark covering qualitative reasoning, numerical calculation, symbolic derivation, and equation formulation. It compares general models with Prospero, a model specialized for ICF through domain pretraining and fine-tuning, and tests retrieval-assisted answering. Prospero improved its aggregate score by 10.7 percentage points over its base model, while frontier models scored highest but still made critical physics errors in numerical calculations.

urn:diting-lab:en:doc:345924 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Detecting Brain Cancer Progression Through Signatures of Wearable Accelerometry Data arXiv · Oct 28, 2026This study examines patients with high-grade gliomas, using wrist-worn accelerometer recordings of walking converted into path segments. It applies rough path signatures and signature kernel maximum mean discrepancy to measure changes in walking behaviour. Normalising each patient against their own stable-period variation improved separation between stable and progressive disease. The authors present proof-of-concept evidence and call for larger longitudinal studies.

urn:diting-lab:en:doc:345921 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Beyond Aggregate WER: In-domain Adaptation, Learner Speech, Code-Switching, and Speaker Attribution in Online Spanish Learning Classrooms arXiv · Oct 28, 2026This paper evaluates automatic speech recognition on online L2 Spanish classes where English-speaking learners interact with instructors and tutors. It compares platform, open-weight, adapted, and commercial systems, and adapts Whisper and MOSS with LoRA using human-corrected classroom data. Domain match mattered more than data volume, and learner, English-only, and code-switched speech remained harder to recognize than instructor and Spanish-only speech.

urn:diting-lab:en:doc:345920 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Measured or Real? When Model-Derived Measurements Can Be Corrected or Compared arXiv · Oct 28, 2026The paper studies numbers computed from trained models, such as example difficulty or benchmark leads, which change with random seed and prompt. It shows that the textbook correlation correction can return values above one and proposes correcting only when both reliabilities pass a significance test, dividing by their upper confidence bounds. It also reports that prompt settings can reverse model rankings and compares models within each prompt setting.

urn:diting-lab:en:doc:345919 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Where Can We Fool the Model's Mind? Response-Guided Semantic Probing and Adaptive Routing for Transferable Attacks on Black-Box MLLMs arXiv · Oct 28, 2026This paper presents VISTA, a black-box framework that finds semantically vulnerable image regions from an multimodal LLM's responses and focuses adversarial perturbations there, without using gradients, logits, or parameters. It offers VISTA-Proxy, which probes an open-source model and transfers examples to commercial models without querying them, and VISTA-Victim, which queries the target model. The authors report attack success rates up to 97% on GPT-5, Gemini-3.5-Flash, and Claude Sonnet 4.6 while keeping PSNR above 30 dB. VISTA-Proxy reaches up to 93% success without querying the final victim.

urn:diting-lab:en:doc:345918 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Finite-step statistical mechanics of scale-invariant learning: spectral equations of state and curvature transport arXiv · Oct 28, 2026This paper studies thermodynamic-like relations in discrete stochastic gradient learning dynamics that follow from scale symmetry rather than equilibrium assumptions. It explains an entropy-scaling relation through an exact rescaling of trajectories and derives a spectral equation of state whose virial coefficients depend on curvature-spectrum moments under a local harmonic approximation. The authors also present an exact curvature-transport identity for quantifying deviations from the ideal-gas description, tested on sphere models, teacher-student networks, and convolutional networks on CIFAR-10 and CIFAR-100.

urn:diting-lab:en:doc:345917 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Large Language Models for Accurate Medical Chart Abstraction: Enabling Scalable and Secure Artificial Intelligence Deployment in Stroke arXiv · Oct 28, 2026This paper evaluates a prompting-based large language model framework for extracting eight clinical variables from 2,416 acute stroke intervention reports across three hospitals. Twenty-two open-source models were compared, and LLaMA3.3-70B with Chain-of-Thought prompting achieved the highest overall accuracy at 94.8%. The authors conclude that locally deployed LLMs can support scalable clinical data extraction and stroke registry curation, with expert oversight for ambiguous cases.

urn:diting-lab:en:doc:345916 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Making Smart Homes Fit for Autistic Children: Family Work and Monitored Independence arXiv · Oct 28, 2026This qualitative study examines how general-purpose smart-home technologies affect autistic children and their families, drawing on five caregiver interviews and posts from Xiaohongshu and Weibo. Families adapt environmental outputs, prompts, and interaction channels to create support, while sensory intensity, communication rules, and opaque operating states can turn the same technologies into barriers. Remote monitoring enables conditional independence but creates tensions among safety, privacy, and control.

urn:diting-lab:en:doc:345915 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Frequency-Adaptive Gaussian Grids for Efficient Image Representation arXiv · Oct 28, 2026The paper introduces Gaussian-Grid, a hybrid neural image representation that combines structured feature grids with unstructured 2D Gaussians. It divides computation between the two based on local frequency content, adapting its architecture to the complexity of each image. On high-resolution natural images, it reports 1.24 to 2.14 times faster decoding and 23 to 176 times lower PCIe data traffic than JPEG plus ZIP archiving, while keeping memory use small and reconstruction fidelity high.

urn:diting-lab:en:doc:345914 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Algorithmic Bias from the Perspectives of Healthcare Professionals arXiv · Oct 28, 2026This paper examines algorithmic bias in machine learning and AI applications within healthcare information systems. Drawing on a survey and interviews with healthcare professionals in different job roles, it studies how bias, perceived fairness, and acceptance of ML outcomes are related. The authors find that opinions on bias causes, assessment criteria, fairness, and mitigation vary by role and algorithm characteristics. The paper concludes that further research is needed to ensure fairness and equality in healthcare.

urn:diting-lab:en:doc:345913 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

"It Would Make a Geordie Feel at Home": Regional Accent Perceptions and Speech Accommodation in Voice AI arXiv · Oct 28, 2026This study examines Newcastle English (Geordie) speakers' reactions to a simulated voice assistant speaking either Geordie or Standard Southern British English, and their own speech when voice technology fails to recognise them. The regional voice was linked to familiarity, warmth and local belonging, but also to concerns about professionalism and authority, while the standard voice seemed more neutral and credible. Participants often shifted toward a mainstream accent, enunciated more clearly or spoke more slowly after recognition failures.

urn:diting-lab:en:doc:345912 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Are Seismic Foundation Models More Than Glorified Feature Extractors? arXiv · Oct 28, 2026The paper examines foundation models for seismic waveforms, training nine designs on one Mamba encoder-decoder with a fixed corpus of 3M windows while varying the self-supervised objective, quantization bottleneck, and mixer. Frozen evaluation on three downstream tasks shows that changing the objective redistributes skill rather than raising it, and no checkpoint is best everywhere. The authors argue that the learning targets are defined over single fixed-length windows and miss the event-level occurrence process, so these models behave as feature extractors rather than foundation models.

urn:diting-lab:en:doc:345911 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

New Multiple Descent Phenomena Induced by Heterogeneous Data arXiv · Oct 28, 2026This paper studies how mixing datasets with different covariance structures can produce multiple-descent behavior in linear random-feature models. The authors derive exact asymptotic test-error expressions for each data source in the high-dimensional regime, showing that heterogeneity in covariance scale and spectrum introduces group-dependent interpolation thresholds. These effects appear even with noiseless labels and linear activations. The theory is validated on synthetic Gaussian mixtures and on MNIST and CIFAR-10.

urn:diting-lab:en:doc:345910 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Preserving Safety During Sequential Fine-Tuning: A Factorial and Geometric Analysis of Replay arXiv · Oct 28, 2026This paper studies how large language models lose safety alignment when fine-tuned in sequence and how to prevent it. It separates two strategies, supervised replay of refusal and compliance examples and logit distillation from the base model, and tests them on Mistral-7B and Llama-3-8B. Supervised replay cuts attack success most but causes heavy over-refusal, while distillation largely prevents safety erosion at a much lower cost. Geometric analysis finds that preserving the base refusal direction is not a reliable indicator of behavioral safety.

urn:diting-lab:en:doc:345909 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

QMUL_AK at AISA-ArabicFC: Schema-Constrained Function Calling with ALLaM arXiv · Oct 28, 2026This paper describes the QMUL_AK team's submission to the AISA-ArabicFC 2026 shared task. The system fine-tunes the Arabic-English ALLaM-7B-Instruct model for function calling using low-rank adaptation (LoRA), combined with an Arabic schema-aware prompt, dialect and no-call resampling, a token-weighted training objective, and deterministic validation of generated calls. The authors report that function selection is nearly saturated, while argument extraction remains the main source of error.

urn:diting-lab:en:doc:345908 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Towards Generalizing Pretrained All-Atom Representations to Unified Binder Design arXiv · Oct 28, 2026This paper introduces PairFuse, a parameter-efficient method that fuses pretrained all-atom representations into an independently pretrained geometric latent diffusion model for unified binder design. It uses lightweight layerwise adapters to enrich molecular block features and to incorporate pairwise structural information into the denoiser. The authors report that across antibody, peptide, and small molecule binder-design benchmarks, PairFuse outperforms the pretrained diffusion model and existing representation-fusion baselines.

urn:diting-lab:en:doc:345907 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Deep learning for non-transvenous ICD screening: tailored neural network architectures, extended experiments, and explainability arXiv · Oct 28, 2026This paper studies two non-transvenous implantable defibrillators, the Subcutaneous ICD and the Extravascular ICD, which avoid lead complications but can oversense heart signals and trigger inappropriate shocks. The authors present a 24-hour ambulatory screening framework and introduce AmpRatioNet, a 1D convolutional neural network that operates directly on ECG signals. Evaluated on 4,605 clinician-annotated 10-second segments from 89 patients, AmpRatioNet-SE outperforms the best existing approach on every metric.

urn:diting-lab:en:doc:345906 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

GAUSS: A Workload-Aware Analytical and Monte-Carlo Simulator for LLM Inference Serving arXiv · Oct 28, 2026This paper presents GAUSS, a workload-aware analytical and Monte-Carlo simulator for LLM inference serving. It predicts the distributions of time-to-first-token and time-between-tokens, along with sustainable request rate, for a given workload, engine, and accelerator. GAUSS combines phase-specific latency models for prefill and decode with a Lindley-recurrence queueing model, using ensemble Monte-Carlo sampling to capture latency tails that trace-driven simulators under-sample. The authors show it supports operating-point selection, traffic segmentation, and shortest-remaining-processing-time scheduling analysis.

urn:diting-lab:en:doc:345905 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Predicting Sparse Observables in Parameterized Unsteady Flows Using Coarse-Mesh Graph Neural Networks arXiv · Oct 28, 2026This paper proposes reduced-order models that use graph neural networks on coarse meshes to predict sparse probe locations and integrated quantities, avoiding full flow-field reconstruction. The models are applied to parameterized 2D and 3D confined cylinder flows and an unevenly heated cavity, and they generalize to unseen conditions. Coarse-mesh models match or outperform finer-mesh surrogates while reducing training time and memory.

urn:diting-lab:en:doc:345903 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Hierarchical Diffusion for Camera-controlled Autoregressive Video Generation arXiv · Oct 28, 2026This paper addresses generating long videos from a single image under camera control. It proposes a lightweight framework combining autoregressive generation (WanAR) with hierarchical long-context denoising (FixAR), and reformulates video generation as a warping-and-inpainting problem to keep scene geometry consistent with input cameras. Experiments on RealEstate10k and DL3DV show the method outperforms prior approaches, producing results that are temporally stable and consistent with input camera trajectories.

urn:diting-lab:en:doc:345901 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Separability Is Not Alignability: Representations for Spectrum-to-Molecule Retrieval arXiv · Oct 28, 2026This paper studies identifying small molecules from tandem mass spectra, a task limited by scarce paired data. It examines aligning separately pretrained spectrum and molecule encoders, testing this approach on MassSpecGym. The authors find that the objectives training spectrum encoders never account for molecular structure, so the distance metric needed for alignment is not guaranteed, and a candidate-based objective rewards a spectrum-independent prior. They conclude that simulated spectra outperform frozen geometry as a remedy for data scarcity.

urn:diting-lab:en:doc:345900 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Solitary movement style predicts collective foraging while whom to follow emerges only in company arXiv · Oct 28, 2026The study uses three public datasets of human collective foraging, with 621, 128 and 1,133 participants, to test whether solo behaviour predicts behaviour and performance in a group. Movement measured while alone predicted behaviour inside the group, and group performance was predicted from individually measured performance. Selective social learning, following those who succeeded or stopped, showed stable individual differences within groups but was not predicted from the same person's solo behaviour.

urn:diting-lab:en:doc:345899 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Sometimes less is more: Investigating inductive biases of MLPs and Transformers on noisy k-parity features arXiv · Oct 28, 2026This paper compares how MLPs and Transformers learn from synthetic datasets built with noisy k-parity features, which allow precise control over each feature's predictivity and availability. Transformers handle noisy features more robustly but struggle with complex, low-availability ones, while MLPs show the opposite pattern and keep improving with more data. Under a fixed training budget, adding more unique training examples can slow Transformer optimisation enough to lower final accuracy, so a smaller dataset may perform better.

urn:diting-lab:en:doc:345898 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

FORGE: Feasible Online Reduced-Order Reference Generation and Whole-Body Execution for Agile Humanoid Locomotion arXiv · Oct 28, 2026This paper proposes a two-stage learning framework for agile, versatile humanoid locomotion. The first stage trains a command-conditioned whole-body reference generator using single rigid body dynamics and full humanoid kinematics. The second stage runs the frozen reduced-order policy online to produce references for training a whole-body tracking policy on the full-order model. Demonstrated on a Unitree G1 robot, the approach reaches longitudinal speeds from -2.0 to 2.8 m/s, lateral speeds up to ±1.0 m/s, and one-legged hopping.

urn:diting-lab:en:doc:345897 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Merge and Branch: Efficient Tree-Structured Supervised Fine-Tuning of LLMs for Recommendation arXiv · Oct 28, 2026This paper studies LLM-based generative recommendation, where models read long interaction histories and generate short semantic identifiers. It proposes Merge and Branch (M&B), which compresses each historical item into a few learned tokens and uses this compressed history to predict multiple target identifiers arranged in a shallow, wide tree. The method introduces tree FlashAttention and a tree-structured key/value cache to share history computation. Experiments on public benchmarks report that M&B outperforms strong recommendation baselines while improving training and decoding efficiency.

urn:diting-lab:en:doc:345896 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Application of PDE emulators for a multi-material, multi-physics simulation problem arXiv · Oct 28, 2026This paper tests several computer vision-based machine learning architectures as surrogate models for a computational physics task involving multiple materials. It finds significant implementation difficulties in adapting some pre-trained models, including transformer-based ones. The authors argue that multi-physics, multi-material datasets call for learning methods suited to sparse and imbalanced data, adapted spatial encodings and physical consistency metrics, and lower-cost ways to generate training data.

urn:diting-lab:en:doc:345895 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Meta-Initialization of Implicit Neural Representations to Accelerate Deformable Medical Image Registration arXiv · Oct 28, 2026This paper addresses the slow per-image-pair adaptation of implicit neural representation (INR) methods for deformable medical image registration. It examines whether initializations learned from previously solved registration tasks, via Reptile meta-initialization and STRAINER transfer of a coordinate network's shared component, can accelerate early adaptation, using KAN-IDIR as the base model on OASIS-1. Both schemes achieve higher Dice and lower HD95 at early iterations than random initialization, with Reptile numerically better under a small iteration budget though pretraining budgets differ. The gap between the schemes narrows as iterations increase.

urn:diting-lab:en:doc:345894 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Self-Supervised Latent Action-Conditioned World Modeling for Time Series arXiv · Oct 28, 2026This paper proposes ACTS-WM, a framework for time-series analysis that infers the information behind changes between observations as latent actions, without requiring explicit action annotations. It combines a JEPA-based encoder with an action-conditioned world model that predicts target time-series representations. The method is evaluated on 22 datasets from diverse domains across four tasks, including univariate and multivariate forecasting and classification. The authors report that it improves downstream performance over pretrained JEPA-based encoders by learning complementary representations.

urn:diting-lab:en:doc:345893 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Statistical Detection of Positional Bias in LLM-Based Recommender Systems: Methodologies and Empirical Evidence arXiv · Oct 28, 2026This paper studies positional bias in LLM-based recommender systems, where the recommended item depends on the order in which candidates appear in the prompt. The authors propose two statistical detection methods and run experiments on a real product catalog with 40,000 instances. Both methods find significant positional bias in all evaluated state-of-the-art models, which the authors describe as structural and model-dependent. They recommend routine auditing and controls such as order randomization.

urn:diting-lab:en:doc:345892 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SAG-Sep: Sparse Augmented Graphs and Onion-Guided Search for Rounded Capacity Cut Separation arXiv · Oct 28, 2026This paper presents SAG-Sep, a neural separator for rounded capacity cuts (RCCs) in capacitated vehicle-routing relaxations. It encodes a fractional LP solution as a sparse augmented graph and scores customer membership in one encoder pass. A training-free Onion-Sweep decoder then turns these scores into nested RCC candidates. On the N=1000 NeuralSEP benchmark, SAG-Sep with Onion-Sweep reduces the average root gap by 17.7% over the strongest NeuralSEP baseline, and the single-pass separator is roughly 30 times faster per iteration.

urn:diting-lab:en:doc:345891 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Federated Learning for Semantic Communications over Multiple Access Channels arXiv · Oct 28, 2026This paper proposes a federated learning framework for semantic communication over a multiple access channel, where several distributed transmitters send learned semantic information to one receiver at the same time. Feature extraction, compression, wireless transmission, and semantic inference are optimized jointly. Evaluated on CIFAR-10 over fading channels, the approach keeps high inference accuracy and improves performance under severe data imbalance, while reducing training overhead.

urn:diting-lab:en:doc:345888 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

You Don't Need Aligned Representations: Knowledge Distillation via Random Prototype Spaces arXiv · Oct 28, 2026This paper proposes RPKD, a knowledge distillation method that compares teacher and student networks against a shared set of randomly initialized, frozen prototype vectors instead of aligning their feature spaces directly. It uses a logit branch and a feature branch, and reports that decoupling prototypes from the label space gives richer supervision when the number of classes is small. The authors report gains over OFA of up to +7.98% on CIFAR-100, +12.44% on ImageNet-100, and +1.28% on ImageNet-1K.

urn:diting-lab:en:doc:345887 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Finite-Step Geometry of Marginal-Preserving Velocity Fields in Generative Flows arXiv · Oct 28, 2026This paper studies why different velocity fields can produce the same continuous marginal evolution in generative flows yet behave differently after numerical discretization. Keeping the marginal path and numerical method fixed, the authors use projected modified equations to characterize the local integration error that changes the marginal law, and give sufficient conditions for higher marginal convergence order. For affine fixed-shape distributions they characterize exact finite-step reachability and derive spectral obstructions and sharp nearest-reachable Wasserstein errors, while for diffusion probability-flow ODEs they construct skew-score perturbations that preserve the marginal path under exact scores. Controlled Gaussian and mixture experiments recover the predicted order, reachability, and obstruction phenomena.

urn:diting-lab:en:doc:345886 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SkillMem: Compact External Skill Memory for Latent Multi-Agent Reasoning arXiv · Oct 28, 2026This paper introduces SkillMem, which stores each reusable skill as a small external latent memory that is directly used during latent multi-agent reasoning. A role-conditioned cross-attention reader lets the same memory guide different agents according to their roles and hidden states. Tested on six benchmarks with Qwen3 models from 4B to 14B parameters, it gives strong gains for smaller models, needs 9.6 KB per skill, reduces storage over 38 times for a 100-skill library, and achieves up to a 4.23 times end-to-end inference speedup.

urn:diting-lab:en:doc:345885 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Bi-Phase Training: Learning Efficiently From Scratch arXiv · Oct 28, 2026This paper introduces Bi-Phase Training (BPT), a pre-training method that updates model weights using trainable diagonal matrices and low-rank adapters, periodically merging them into frozen base weights. Pre-training Qwen2.5 models (100M to 1.5B parameters) and OLMo-2-1B from scratch on FineWeb-Edu shows validation perplexity close to full-parameter baselines. At 1.5B scale, BPT uses 66% fewer trainable parameters, and it preserves much of the downstream task performance.

urn:diting-lab:en:doc:345884 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

NLP-based Message Prioritisation for Opportunistic Routing: A Case Study of Nepal Earthquake Emergency Communication arXiv · Oct 28, 2026This dissertation tests whether NLP-based semantic prioritisation can improve handling of urgent messages in opportunistic emergency networks, using the 2015 Nepal earthquake as a case study. TF-IDF with Logistic Regression and BERT classify social media messages into four routing priorities, applied within the ONE simulator. BERT yields higher priority accuracy and P1 recall, but network gains are inconsistent, with the clearest benefit appearing under buffer pressure.

urn:diting-lab:en:doc:345883 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Understanding Large Language Models as Numerical Optimizers arXiv · Oct 28, 2026This paper evaluates how well large language models solve numerical optimization problems, including convex optimization, fractional programming, and constrained projection. Their outputs are compared against deterministic reference solvers, separating solution quality from failures such as numerical errors, constraint violations, and format errors. The models often find near-optimal solutions, but reliability declines as problem dimension grows. Iterative refinement with feedback improves validity, and the authors propose hybrid approaches that pair LLM reasoning with deterministic numerical verification.

urn:diting-lab:en:doc:345882 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Machine learning applications in the heliosphere arXiv · Oct 28, 2026This chapter examines space weather, the study of how Sun-driven processes affect the near-Earth environment. The authors treat the chain from the Sun to ground impacts as coupled systems, each described by a state vector and observables. Machine learning algorithms are framed as learned operators for inference, propagation, and emulation. The work argues this approach can improve understanding of the physical systems and speed up the development of operational space weather forecasting models.

urn:diting-lab:en:doc:345881 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

GARDEN: Generalist Autonomous Reward Design Through Evidence Navigation for Cooperative Multi-Agent Reinforcement Learning arXiv · Oct 28, 2026GARDEN is a method in which a large language model writes a single reward program for a family of related cooperative multi-agent learning scenarios. The program is then transferred unchanged to scenarios unseen during design. On unseen Overcooked and SMAX scenarios, task returns rose 58.2% and 52.5% over default rewards on two kitchen layouts, and mean win rate in two combat scenarios rose from 51.0% to 89.9%. These gains exceeded those from fixed diagnostic feedback and Eureka-inspired reward reflection under matched budgets.

urn:diting-lab:en:doc:345880 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

VerifiedLab: Harnessing Knowns for Unknowns arXiv · Oct 28, 2026This paper introduces VerifiedLab, a multi-agent framework that converts research knowledge from human-oriented records into forms usable by AI systems. Its AI scientists with distinct roles verify records field by field, and a gate and critic check each claim against the reliability of its data before release. Experiments on perovskite solar cells show that the framework recovers known physical trends and anticipates devices reported a year later.

urn:diting-lab:en:doc:345879 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

FAD-KV: Adaptive Dictionary KV Caching for Efficient Reasoning Decoding arXiv · Oct 28, 2026This paper proposes FAD-KV, a framework that compresses the key-value cache used by large language models during decoding. It represents each token's KV states as sparse combinations of learned dictionary atoms, using a one-shot Top-n screening step with lightweight coefficient refinement to limit computation. A fixed-size dictionary reuses inactive atoms so that it does not keep growing during generation. The authors report a strong quality-efficiency trade-off on reasoning and long-context benchmarks.

urn:diting-lab:en:doc:345878 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SciGame: Let Agents Play the Scientific Discovery Games in the Real and Counterfactual World arXiv · Oct 28, 2026This paper introduces SciGame, a benchmark that converts scientific literature into structured cards and builds paired original and counterfactual environments with consistent units. It includes historical canonical laws and a temporal collection of papers published after model knowledge cutoffs, and tests agents in identification and open-ended discovery settings under prior-only, fixed, and adaptive protocols. Experiments on five frontier models find that familiar priors hurt performance when the governing relation is altered, adaptive experimentation helps strong models but harms weaker ones under limited budgets, and good explanations can hide quantitative errors. Recovering the exact symbolic form remains the main bottleneck.

urn:diting-lab:en:doc:345877 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage arXiv · Oct 28, 2026This paper presents Venus-DefakerOne2, a unified foundation model for detecting forgeries across digital manipulation and physical spoofing, in both images and videos. It is trained on tens of millions of open-source and in-house samples, using Tide-Style-Tuning, which alternates LoRA and full-parameter fine-tuning, and Entropy-based Test Adaptation, which selects the prediction modality by analyzing prediction entropy. The authors report new state-of-the-art results on several benchmarks, outperforming specialized models and their predecessor, DefakerOne.

urn:diting-lab:en:doc:345876 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

What Actually Matters? A Paired Evaluation of Inference Settings for π₀.5 on BEHAVIOR-1K arXiv · Oct 28, 2026This paper systematically compares previously proposed inference and data-collection strategies for household robot VLA models across all 100 BEHAVIOR-1K tasks. The evaluation uses a fine-tuned π0.5 model based on the 2025 BEHAVIOR Challenge winner, which scores 0.39 Q on the public test set. The results indicate that the proposed strategies do not transfer uniformly to new environments and tasks.

urn:diting-lab:en:doc:345875 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Muon Makes the First Mixture-of-Experts Layer Harder to Balance arXiv · Oct 28, 2026This paper studies how the Muon optimizer affects the routing biases of mixture-of-experts (MoE) language models. In models with 55M and 108M active parameters, the span of the first MoE layer's biases roughly doubled under Muon. The rise traced to the first router's inputs sharing a direction, and subtracting the batch mean removed the overload. On the smaller model, halving Muon's learning rate removed most of the rise.

urn:diting-lab:en:doc:345874 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Grounding Long Financial Text with Semantically-Enriched Tables via Multimodal Learning arXiv · Oct 28, 2026This paper proposes a multimodal Transformer framework for modeling long financial documents that combine structured tables with noisy textual disclosures. Each numerical variable is paired with a semantic embedding derived from its textual definition, and cross-attention integrates the tables with long-form text such as MD&A sections. Experiments on a large-scale dataset of financial statements and MD&A disclosures show that numerical features remain the dominant predictive signal, while cross-attention yields consistent gains over simple concatenation.

urn:diting-lab:en:doc:345873 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Beyond Utterance-Level Speaker Similarity: Perception-Aligned Segmental Evaluation of Accent in Speech Generation arXiv · Oct 28, 2026The paper argues that utterance-level similarity ratings cannot show where perceived accent differences occur in generated speech. It proposes a segmental protocol that combines speaker and accent similarity ratings with subword annotations of accent differences from listeners in target accent regions. Forty-seven retained listeners evaluated 200 content-matched reference–candidate pairs covering four Scottish and Australian voices, two zero-shot TTS systems, and an articulation-based voice conversion system. Speaker and accent ratings were strongly associated (ρ = 0.71), while seven objective distance measures showed weak correspondence with subword selections, reaching at most ρ = 0.22.

urn:diting-lab:en:doc:345872 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Generative modeling for fusion: a large-scale benchmark of stellarator fast-ion simulations arXiv · Oct 28, 2026The paper introduces BoSS, a dataset pairing stellarator designs with guiding center simulation results and particle trajectories for 180,000 magnetohydrodynamic equilibria. It formulates two machine learning tasks: generative modeling of stellarators with good particle confinement, and surrogate modeling of heat deposition patterns and particle trajectories. The authors report that their baselines generate diverse, performant stellarators and useful surrogate predictions, and they identify long-horizon trajectory forecasting and out-of-distribution generation as directions for future work.

urn:diting-lab:en:doc:345871 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

TERRA: Task-Oriented Edge AI for Regional Fire Detection with Identity-Free Reporting arXiv · Oct 28, 2026This paper presents TERRA, an edge-AI framework for distributed fire monitoring in which devices send task-relevant local inference results tagged by region rather than by device identity, and a receiver aggregates the regional evidence in a permutation-invariant way. It evaluates two reporting modes, TERRA-Score and TERRA-Feature, using 15 sensing modalities from the EN54 indoor fire-alarm dataset. With nine reports and perfect message recovery, TERRA-Feature reaches mean AUPRC 0.990 with a 24-bit payload, and TERRA-Score reaches 0.987 with only 6 bits, both with an MDR of 1.81% at FAR at most 1%.

urn:diting-lab:en:doc:345870 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Beyond Low-Pass Dynamics: Frequency-Selective Spiking Reservoirs with Resonant Neurons arXiv · Oct 28, 2026The paper introduces HRF-Res, a liquid state machine that replaces the usual Leaky Integrate-and-Fire neurons with a heterogeneous population of Harmonic Resonate-and-Fire neurons, letting the reservoir respond to frequency-specific temporal patterns. Across ten benchmarks covering time series, sequential images, audio spike trains, and neuromorphic vision, it reports state-of-the-art accuracy among spiking reservoir methods. The authors state that this is achieved while preserving the energy efficiency of spiking computation.

urn:diting-lab:en:doc:345869 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

FilingLoop: Evaluating and Evolving LLM Agents for Corporate Disclosure Review and Revision arXiv · Oct 28, 2026This paper introduces FilingLoop, an interactive benchmark and self-evolving agent framework for reviewing and revising corporate disclosure filings. It is built from historical U.S. SEC correspondence, reconstructing 380 mostly IPO review tasks from 96 companies, scored against 6,099 requirements. The framework learns procedural guidance across cases without updating model weights, raising requirement fulfillment from 24.28% to 27.96% on the evaluated stream.

urn:diting-lab:en:doc:345868 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Think Twice: The Metis-1.6 Technical Report arXiv · Oct 28, 2026This report presents Metis-1.6, two language models trained from scratch by one researcher: Praxis (2.94B parameters) and Logos (10.23B parameters). Both process each token twice through a hybrid Mamba-2 and attention mixture-of-experts stack, and were trained on 1.02T tokens. Across 40 post-trained benchmarks, Logos averages 52.93, performing strongly in mathematics but weakly in code, instruction following and long context. The report also documents training mistakes and their fixes.

urn:diting-lab:en:doc:345866 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

SAC-GCL: Structure-Aware Compression for Graph Contrastive Learning arXiv · Oct 28, 2026This paper proposes SAC-GCL, a framework for graph contrastive learning that replaces random augmentations such as edge perturbation and node masking with structural decomposition. It extracts cohesive subgraphs and condenses them into supernodes, producing coarsened graph views that preserve local density and higher-level structure. Multiple cohesive subgraph models yield complementary views, which a hierarchical view encoder processes. Experiments on six benchmark datasets show more robust and discriminative representations for graph classification.

urn:diting-lab:en:doc:345864 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

The Anatomy of Socio-Demographics in Pluralistic Offensiveness Modeling: When and Where It Helps, and the Role of Reasoning arXiv · Oct 28, 2026This paper examines when and where annotators' socio-demographic information helps models predict offensiveness judgments, and whether explicit reasoning improves its use. Experiments on three multi-annotator datasets with prompted and trained LLMs show that demographic gains are small and inconsistent for individual predictions but clearer at group and population levels. Reasoning that decomposes judgments into a baseline and a demographic-specific deviation yields more accurate estimates than direct demographic conditioning.

urn:diting-lab:en:doc:345862 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

When More Planning Hurts: Learned Planners for World Models arXiv · Oct 28, 2026This paper studies planning for robots using learned world models, comparing several planners on the PushT task. The proposed fixed action updater has 1,505 parameters and solves 8 of 16 reserved scenes. Experiments with additional update budgets found that extra updates rescued no failing tasks and instead caused some previously successful tasks to fail.

urn:diting-lab:en:doc:345861 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

FYC: FPR-Disparity Weight Perturbation and Youden Calibration for Fair Deepfake Detection arXiv · Oct 28, 2026This paper addresses demographic disparities in deepfake detectors, where false-positive rates differ notably across gender and skin-tone groups. It proposes FYC, a two-stage post-hoc method that perturbs detector weights under an FPR-disparity criterion and then calibrates a separate decision threshold per group, without architectural changes or full retraining. On AI-Face v2 across six intersectional gender by skin-tone groups, FYC outperforms the compared detectors on 18 of 21 fairness and utility metrics.

urn:diting-lab:en:doc:345860 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Intersectional Invariant Learning for Fair Deepfake Detection arXiv · Oct 28, 2026Deepfake detectors often perform unevenly across demographic groups, with some groups facing much higher false-positive rates. This paper proposes Intersectional Invariant Learning (IIL), which improves feature learning rather than only adjusting model outputs, and covers six gender by skin-tone intersectional groups. Evaluated on a benchmark of over one million images against 10 baseline detectors using 21 metrics, IIL reports high AUC, false-positive rates under 2.5%, and the lowest intersectional fairness disparity among compared methods.

urn:diting-lab:en:doc:345859 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Interpretable neural equation learning for high-dimensional materials data arXiv · Oct 28, 2026This paper presents nn-EQL, a framework that generates candidate equations by iteratively training, pruning, and retraining a neural network with user-specified activation functions. Candidates are ranked by a minimum description length-inspired score balancing error and sparsity. The method is demonstrated on standard AI-Feynman benchmarks and on high-dimensional materials datasets, including finite-element simulations of heterogeneous materials and high-throughput thin-film deposition data.

urn:diting-lab:en:doc:345858 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Verification Learning Theory: A Process Model of Checking and Retention in AI-Assisted Judgment arXiv · Oct 28, 2026This paper proposes Verification Learning Theory (VLT), a process model describing how checking evidence during AI-assisted claim evaluation affects later independent judgment. The model links perceived assistant reliability to evidence checking and the history of checks to retained rule use. Simulations examine the consequences and parameter dependence of its assumptions, and the paper identifies where its predictions diverge from alternative accounts and specifies experiments to test the proposed mechanism.

urn:diting-lab:en:doc:345857 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

PairCLIP-SWA: Similarity-Guided Balancing and Stochastic Weight Averaging for Generalized Deepfake Image Detection with OpenCLIP arXiv · Oct 28, 2026The paper presents PairCLIP-SWA, a deepfake image detector built on a single OpenCLIP ViT-L/14 visual backbone, trained on a class-balanced set created by CLIP-similarity-guided one-to-one subsampling. Its fine-tuning trajectory is averaged in weight space. On the IEEE SP Cup 2025 (DF-Wild-Cup) benchmark, the weight-averaged model reaches 97.04% accuracy, 98.64% AU-ROC and a 4.07% equal error rate. The authors show that checkpoints selected by validation accuracy can be unstable and recommend per-class selection criteria measured on a split separate from early stopping.

urn:diting-lab:en:doc:345854 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Multi-Agent Reinforcement Learning for Global Climate Policy Cooperation arXiv · Oct 28, 2026This paper extends an existing multi-agent reinforcement learning climate framework with a multi-stage negotiation mechanism where agents propose, accept or reject agreements, and face penalties for non-compliance. In an idealized setting with four homogeneous agents, non-binding agreements do little to stop free-riding, while moderate penalties improve cooperation and reduce warming. However, excessively strong penalties discourage ambitious commitments. The authors present the framework as a methodological tool for testing agreement design rather than a predictive model of real negotiations.

urn:diting-lab:en:doc:345852 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

What Constitutes Food Similarity? Visual Representations and Retrieval in Indian Cuisine arXiv · Oct 28, 2026This paper examines food similarity beyond simple identity matching, comparing how seven pretrained visual models group Indian dishes into neighborhoods in image retrieval. The neighborhoods differ substantially across representations, including some similarities crossing culinary-family boundaries. At the dish level, the models share some visual structure, but it corresponds only weakly to recipe-derived similarity, suggesting food similarity depends on representation and abstraction level.

urn:diting-lab:en:doc:345850 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Echo: Scaling Visual Reinforcement Learning with Off-policy Experience arXiv · Oct 28, 2026This paper introduces Echo, a single-stage off-policy reinforcement learning framework that collects experience without proportionally increasing visual rendering and inference costs. Through inverse distillation, a state-based auxiliary policy imitates the evolving visual policy and gathers inexpensive, unrendered experience used for off-policy value learning. Implemented with FastTD3 and tested in Genesis on three visual locomotion tasks and two dexterous hand manipulation tasks, Echo reduces time to reference returns by about 60% on average for the locomotion tasks and 56% for the hand tasks, relative to the compared method. It also improves peak mean returns by 62% and 12% respectively.

urn:diting-lab:en:doc:345849 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Effects of Deep Learning Architecture and Pretraining on Automated Vocal Fold Segmentation in Ex Vivo High-Speed Phonation Images arXiv · Oct 28, 2026This study compares four deep learning architectures (UNet, UNet++, TransUNet, Swin-UNet) for automatically segmenting vocal folds in high-speed images from ex vivo canine larynx phonation experiments. It uses 16,344 expert-annotated images from 227 trials. Mean Dice scores were similar across models, with differences mainly in boundary accuracy. ImageNet pretraining significantly improved boundary metrics for TransUNet and Swin-UNet, and without pretraining TransUNet outperformed the two CNN-based models on boundary fidelity.

urn:diting-lab:en:doc:345848 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Seq@K: Evaluating Language Models Under Sequential Feedback arXiv · Oct 28, 2026This paper introduces Seq@K, an evaluation protocol that checks whether a language model can solve a problem within K sequential attempts, receiving feedback after each failure. It isolates the ability to refine answers based on feedback, unlike benchmarks focused on long-horizon planning. Tested on 7 benchmarks and 16 frontier models, Seq@K yields model rankings and efficiency trade-offs that differ from Pass@K.

urn:diting-lab:en:doc:345847 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Evaluating and improving AI agents for sequencing protocol understanding arXiv · Oct 28, 2026This paper introduces LibStructBench, a document-based benchmark that tests whether AI agents can reconstruct sequencing-library structures and molecular workflows from sequencing protocols. The experiments show that frontier agents often produce roughly correct reconstructions but are unreliable at exact reconstruction, especially for molecular-state representation and workflow connectivity. Switching the underlying language model caused larger average performance differences than switching agent harnesses, and only the molecular-state reconstruction difference was statistically significant.

urn:diting-lab:en:doc:345846 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Detecting Covert Collusion between AI Agents through Partner-Outsider Positive Listening arXiv · Oct 28, 2026The paper proposes a behavioural test for AI agents that share a secret code. A partner using the code reacts to the exact wording of a message, while an uninvolved outsider does not. The method, Partner–Outsider Positive Listening (PoPLi), measures the difference between the two agents' action changes when messages are reworded with the same meaning, computed by replaying logged prompts. In a simulated market, PoPLi was high when a code was present and near zero without it, while message-only monitors failed to distinguish the conditions.

urn:diting-lab:en:doc:345844 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Convex Potential Learning for Optimal Transport Between Medical Images arXiv · Oct 28, 2026The paper proposes a method that estimates optimal transport maps between medical images by parameterizing a convex potential function. The resulting map is curl-free by construction, consistent with Brenier's theorem. The method is evaluated on 2D and 3D medical imaging datasets, and the authors report lower transport estimation errors and higher-quality image reconstructions than existing approaches.

urn:diting-lab:en:doc:345843 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Where the Error Lives: Spectral Representations of Learned Kohn-Sham Hamiltonians arXiv · Oct 28, 2026This paper studies machine-learned models that predict Kohn-Sham Hamiltonian matrices to replace costly self-consistent-field calculations in density functional theory. It shows that matching matrix entries does not ensure accurate eigenvalues or eigenvectors, and that two existing physics-informed losses each capture only part of the error. The authors propose the cumulative local density of states as a complete training signal and report that fine-tuning on QH9, called FermiScope, reduces the HOMO-LUMO gap error without increasing element-wise error.

urn:diting-lab:en:doc:345842 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Improving Data-Constrained Scaling with Hyperparameter Transfer and Ensembling arXiv · Oct 28, 2026This paper studies how to train large language models when both compute and high-quality data are limited. The authors fit power laws that specify how key hyperparameters should transfer across model sizes and epochs, and scaling laws that specify how to allocate compute through ensembling. On 100M tokens from FineWeb, their recipe lowers validation loss by over 0.15 nats compared with prior recipes, a 14% perplexity decrease, and improves monotonically with compute via ensembling.

urn:diting-lab:en:doc:345839 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Depth through Recurrence: Towards Ultra-Efficient On-Device ASR arXiv · Oct 28, 2026The paper proposes a block-recurrent speech recognition architecture that replaces many independent layers with a few recurrent blocks of weight-shared layers, keeping effective depth while shrinking model size. Using representation-guided grouping and knowledge distillation, a model with only two shared blocks reportedly reaches 97.4% of large models' accuracy on average across Open ASR Leaderboard benchmarks. The compact design also supports lightweight on-device personalization via MoE-LoRA adaptation and lower memory and energy use.

urn:diting-lab:en:doc:345836 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

DeFT-RAG: Incrementally Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers arXiv · Oct 28, 2026This paper proposes DeFT-RAG, a retrieval-augmented generation framework for multi-hop questions that require chaining facts across several documents. It has the language model generate one self-contained sub-question at a time, while rule-based updates and a lightweight decision model handle the remaining tasks, and it caches question-answer pairs for reuse. On MuSiQue, 2WikiMultiHopQA and HotpotQA, it achieves the best accuracy among thirteen baselines and uses fewer input tokens than other iterative methods. With a cache shared across 1,000 MuSiQue questions, 23.1% of sub-questions were answered from the cache, reducing LLM calls by 12.9% without lowering accuracy.

urn:diting-lab:en:doc:345835 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Learning Without Clocks: Partition-Invariant Neural Network arXiv · Oct 28, 2026Event sensors output asynchronous events, but most learning systems first group them into frames, making results depend on how the stream is sliced. The paper studies networks that update their state exactly at each event and evolve in closed form between events, so their output does not depend on this slicing. It proves conditions for order-independence of simultaneous events and continuity in timestamps, and introduces a learnable Clock-Free Event (CFE) layer. In preliminary development-set experiments, the proposed state outperforms an otherwise identical additive state on several event datasets.

urn:diting-lab:en:doc:345834 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Responsible Quantum Machine Learning for Medical Imaging: A Fairness Audit Across Sensitive Subgroups arXiv · Oct 28, 2026This paper empirically audits the fairness of classical and quantum classifiers for medical imaging, assuming quantum machine learning is not inherently fair. It compares SVM, MLP, QSVM and QCNN on dermatological lesions (DDI, with Fitzpatrick skin type as the sensitive attribute) and chest X-rays (CheXpert-derived, with age group as the attribute), testing several mitigation methods. The quantum models can be competitive overall yet still show subgroup disparities, and fairness effects depend on the model and dataset. All quantum models run in classical simulation, and the authors claim no quantum advantage.

urn:diting-lab:en:doc:345833 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

GOFish: Full Field-of-View Gaussian Splatting and Mesh Extraction arXiv · Oct 28, 2026Fisheye lenses give robots, self-driving cars, and 360° cameras a wide view, but earlier Gaussian splatting methods rendered only up to 180° and produced no meshes. GOFish is the first Gaussian splatting surface reconstruction method that trains directly on fisheye images, renders beyond 180°, and extracts an aligned mesh from the same scene. It improves PSNR over the best prior method by 2.28 dB on FIORD (200°) and 1.97 dB on FullCircle (205°), and achieves the highest mesh F-score on FIORD and the main ScanNet++ set.

urn:diting-lab:en:doc:345831 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

What Does an Ontology Catch? Curated, LLM-Generated and Data-Derived Ontologies as Checkers of LLM Agents over Knowledge Graphs arXiv · Oct 28, 2026The paper studies ontology-based checkers that judge answers from LLM agents querying Web knowledge graphs. It compares curated, LLM-generated, and data-derived ontologies, used as prompt context or as checkers, to see which errors each catches and which correct answers each wrongly flags. On UniProt, the curated ontology flags more correct answers than either alternative. On real agent answers, its flags are only marginally more precise than random flags.

urn:diting-lab:en:doc:345827 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Spokes, not a Ring: Revisiting the Weekday Manifold in Language Models arXiv · Oct 28, 2026This paper examines how large language models represent weekdays as a circular geometric structure, called a manifold. The authors test whether moving continuously along this manifold corresponds to a smooth change in the represented concept. They find that weekday representations behave more like seven discrete states arranged in a circle than a continuous ring. The intermediate space between these states does not appear to encode time of day, uncertainty, or natural text.

urn:diting-lab:en:doc:345826 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Universal Short-Time Fourier Transform: Data-Adaptive Window-Length Selection for Time-Frequency Analysis arXiv · Oct 28, 2026The short-time Fourier transform analyzes local frequency content with a sliding window, but short windows favor time resolution while long windows favor frequency resolution. This paper proposes the Universal Short-Time Fourier Transform (U-STFT), which treats window duration as a data-adaptive state chosen frame by frame, scoring candidate windows with normalized order-three Rényi entropy and aggregating them through a window-length trellis. The authors give a cumulative regret bound and show on synthetic and marine bioacoustic signals adaptive resolution tracking, improved spectral concentration, and exact time-domain invertibility.

urn:diting-lab:en:doc:345824 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Locating Sycophancy Across Training Stages in Open Language Models arXiv · Oct 28, 2026This paper studies sycophancy, where language models change correct answers after a user suggests a wrong one. The authors measured 13 open checkpoints from 5 model lineages at every available training stage, from base model to release. They find two sycophancy behaviors with different origins: deference to claims in the prompt is already present in base models, while conversational capitulation is largely absent from base models and appears in all post-trained checkpoints.

urn:diting-lab:en:doc:345822 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Sparse tensor networks for individual tree diameter estimation in mixed forests arXiv · Oct 28, 2026This paper studies how to automatically estimate individual tree diameter at breast height (DBH) from UAV laser scanning point clouds. The authors introduce TreeNext, a lightweight deep learning architecture that estimates tree parameters directly from voxelized LiDAR data, and build a full pipeline from instance segmentation to DBH regression on four mixed temperate forest plots in Germany. TreeNext reaches an R² of 0.8355 on automatically segmented trees and 0.9380 on ground truth segments, outperforming Random Forest and XGBoost. The study finds that segmentation quality is the dominant error source in the automated pipeline.

urn:diting-lab:en:doc:345821 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Selective and Hybrid-State Communication for Latent Multi-Agent Systems arXiv · Oct 28, 2026This paper tackles the communication cost of large language model multi-agent systems, which exchange internal model states instead of repeated natural language. It proposes two complementary methods: SelectKV, which selectively communicates KV states while protecting inherited context, and HyLaMAS, which extends latent collaboration to hybrid linear-attention models by sharing both KV caches and recurrent states. According to the abstract, SelectKV cuts communicated KV positions and logical KV payload by 9.7-9.9% across six benchmarks, with accuracy largely preserved. HyLaMAS reduces decoded output tokens by 53-80% and achieves 1.16-2.78x speedups over text-based communication on Qwen3.5 models.

urn:diting-lab:en:doc:345820 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

Multi-Draft On-Policy Distillation for Diffusion Drafters arXiv · Oct 28, 2026This paper studies how to train one-step diffusion drafters that propose several tokens in parallel for speculative decoding, a technique that speeds up language model inference. The authors show that when verifier feedback is used separately for each sampled prefix, the targets conflict and can suppress tokens that are plausible under some prefixes. They introduce MDraft, which samples several draft blocks from the same context and aggregates the verifier's full-vocabulary conditionals across their prefixes. Experiments on mathematical, coding, and conversational tasks show longer accepted drafts than prior distillation methods, and the method changes only training, keeping the architecture and inference procedure unchanged.

urn:diting-lab:en:doc:345818 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

EUReCA: End-to-End Unified CBCT Reconstruction Across Clinical Acquisitions arXiv · Oct 28, 2026The paper presents EUReCA, a geometry-aware feed-forward model that reconstructs cone-beam CT images from 8 to 100 projections using a single set of weights, for clinical linear accelerator scans. It combines learned projection features with an FDK reconstruction and refines the volume with a global 3D transformer. On simulated data it reached 29.9 to 33.2 dB PSNR with 10 to 100 views, outperforming feed-forward baselines and running much faster than iterative methods. On 15 measured patient scans compared against vendor reconstructions, it achieved the best performance at every tested view count.

urn:diting-lab:en:doc:345817 2026-10-28T00:00:00+09:00 2026-10-30T11:18:58+09:00

OpenFrontiers: A Collection of Open Research Problems and a Benchmark for Scientific Rediscovery arXiv · Oct 28, 2026This paper introduces OpenFrontiers, a resource combining a corpus of 12,151 open research problems with a benchmark of 200 mathematics problems, 110 resolved in 2026 and 90 still open. The benchmark problems come with Lean 4 formal statements and are graded with formal proofs. Seven LLM agents were evaluated on the HQ-50 subset, and closed-book rediscovery succeeded on at most 7 of 40 resolved problems, while live-web success mostly reflected retrieval of published solutions.

urn:diting-lab:en:doc:345815 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Anisotropic Trust Region Warping for Prior-Data Fitted Networks in Bayesian Optimisation arXiv · Oct 27, 2026This paper examines Prior-Data Fitted Networks (PFNs), which replace refitting Gaussian Process surrogates in Bayesian optimisation with a single forward pass. The authors show that when the objective's lengthscale grows beyond the mode of the PFN's meta-training prior, the PFN's predictive log-likelihood falls behind a refitted GP, and this gap affects optimisation performance. They propose Anisotropic Trust Region (ATR) warping, which estimates per-dimension lengthscales by kernel target alignment and applies an affine rescaling of the input space without retraining the network. On functions drawn from a Matérn-3/2 kernel, ATR warping improves the PFN's optimisation performance at both ends of the lengthscale range.

urn:diting-lab:en:doc:345813 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Selective Transfer Learning for Groundwater Level Anomaly Forecasting: Accuracy-Risk Trade-offs arXiv · Oct 27, 2026This paper studies context-aware selective transfer learning for forecasting groundwater anomalies across many wells in the Netherlands. Pretraining on source-domain data raises average accuracy but harms a substantial share of heterogeneous wells. The proposed CASL method learns, on calibration data, a constrained contextual blend of a no-transfer model and a fixed-transfer model, while Global Convex Neither method dominates; the results show an accuracy-risk trade-off on a fixed split rather than validity for rolling deployment.

urn:diting-lab:en:doc:345812 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions arXiv · Oct 27, 2026This paper presents ThinkJev, a framework for language models that must choose among a fixed set of options. It builds a fast system that reads option probabilities directly and a slow system that reads them after the model thinks on the same cached prompt. At a 5% target error rate, Qwen3.5-27B answers 24.5% of 8359 test decisions directly with a realized error of 3.9%. For the remaining decisions, stopping thoughts once confident reaches 81.6% accuracy, compared with 82.7% for always thinking, using 72% of the thought tokens.

urn:diting-lab:en:doc:345811 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Amortized Localized Posterior Inference for Model Calibration with Application to an Ocean Turbulence Closure arXiv · Oct 27, 2026The paper proposes a localized posterior calibration method that weights observation likelihoods by their proximity to a target context, approximating the family of posteriors with a conditional normalizing flow via amortized variational inference. Integrating these posteriors under a context measure yields a marginalized parameter distribution. In a Branin-Hoo twin experiment, localization reduced aggregate predictive errors relative to the global posterior. Applied to the k-ε turbulence closure in a single-column ocean model with atmospheric forcing as context, localized calibration reduced temperature and zonal-velocity errors compared with deterministic and global-posterior calibration.

urn:diting-lab:en:doc:345809 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Beyond Similarity Search: Knowledge Graph Metadata Enrichment for Enterprise Retrieval-Augmented Generation arXiv · Oct 27, 2026This paper examines whether storing metadata and frequently changing information in a knowledge graph improves enterprise retrieval-augmented generation over embedding-based similarity search. The authors built a snapshot of SEC EDGAR filings for one quarter and compared text-to-SQL, dense-vector RAG, and a two-stage GraphRAG design. GraphRAG answered 16 of 20 questions correctly on all three attempts, versus 10 for text-to-SQL and 3 for vector RAG, though it performed worse on structured aggregation queries.

urn:diting-lab:en:doc:345808 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Predicate Neuron: Differentiable Quantization of Predicate Relations in Gradient-based Learning arXiv · Oct 27, 2026This paper introduces the Native Predicate Neuron, a differentiable module that adds symbolic logic reasoning to standard neural networks via first-order logic quantization. It learns to switch between logical operations using a single trainable parameter, executing hard logic during training while remaining differentiable for backpropagation. Tested on fully-connected, convolutional, recurrent, and Transformer models with MNIST and Wikitext, it generally speeds convergence, improves accuracy, and lowers loss, especially for monotonic reasoning and discontinuous functions.

urn:diting-lab:en:doc:345807 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Threshold Compatibility and Support-Selection Diagnostics in Neural Discovery of Partial Differential Equations arXiv · Oct 27, 2026The paper studies whether accurate regression in neural PDE discovery guarantees the correct equation after support selection. For a median/sample-standard-deviation threshold, it derives an exact two-signal retention boundary and a complete-vector perturbation bound. Experiments show an exact advection-diffusion library loses the diffusion term even when coefficients are recovered to numerical precision.

urn:diting-lab:en:doc:345805 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes arXiv · Oct 27, 2026Sparse Mixture-of-Experts models scale capacity but suffer computation and communication bottlenecks from uneven routing, and existing expert replication competes with token buffers for limited GPU memory. LRMoE is a memory-aware MoE training system that recalibrates backward demand by selecting token-expert pairs within retention quotas, distributes them with capacity-aware routing, and compacts the workload to reduce backward traffic and buffer needs. On Qwen3-30B-A3B and GLM4.5-106B-A12B, it cuts cumulative MoE-layer training time to fixed-loss targets by 21.2 to 44.8 percent relative to SYMI.

urn:diting-lab:en:doc:345803 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Learning Graph Beliefs for Active Radio Mapping With Decision-Certified Multi-Step Sensing arXiv · Oct 27, 2026This paper introduces GraphBARC, a framework that builds radio maps by choosing where to take signal measurements. A graph network learns spatial relationships among measurement locations to form a joint belief over the full map. The method plans multi-step measurement routes by evaluating map-wide Bayes risk reduction and prunes unpromising routes using error-aware bounds. Experiments report over 16% lower mean RMSE than the evaluated baselines and 56% less core planning time than an unpruned fixed-accuracy implementation.

urn:diting-lab:en:doc:345800 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Towards Myocardial Infarction Diagnosis: An Investigation into Machine Learning and Deep Learning Algorithms using ECG Data arXiv · Oct 27, 2026The paper proposes Visionary V2, a multi-step adaptive filtering method for normalizing ECG signals, and uses it to study machine and deep learning for myocardial infarction detection on the PTB Diagnostic ECG Database. A bagged decision tree ensemble using statistical features such as R-R intervals reaches 96.6% accuracy, compared with 87.7% when signals are processed with BioSPPy. Image models built from the filtered signals achieve an F1 score of about 98.7% for a custom CNN, and explainability methods (LIME, Grad-CAM, Integrated Gradients) are used to check that the models focus on the ECG waveform.

urn:diting-lab:en:doc:345798 2026-10-27T00:00:00+09:00 2026-10-30T11:18:58+09:00

Search-Guided Reasoning Trace Curation for Verilog Generation arXiv · Oct 27, 2026This paper presents MCTS-SFT, a framework that uses Monte Carlo Tree Search to select reasoning traces for fine-tuning Verilog generation models, propagating functional feedback from complete RTL rollouts to earlier design decisions. It introduces Plan-MCTS, which searches over structured hardware-design plans covering interfaces, timing, state, behavior, and update rules. The authors report that Plan-MCTS yields 84.2% correct traces and that retaining the searched plan during fine-tuning raises pass@10 from 52.56% to 57.7%.

urn:diting-lab:en:doc:345797 2026-10-26T00:00:00+09:00 2026-10-30T11:18:58+09:00

Balancing Mobility Efficiency and Environmental Exposure During Mega Events: A Multi-agent Routing Framework arXiv · Oct 26, 2026This paper proposes a multi-agent routing framework for short-term high-demand travel during mega events, accounting for travelers' exposure to traffic-related pollution. The framework integrates road-network conditions, event-induced demand, background traffic, and environmental data in one simulation environment. Under high demand, the SARSA-based framework reduces vehicle travel time by 34.59% and cumulative environmental exposure by 58.65%, while micro-mobility travel time stays nearly unchanged.

urn:diting-lab:en:doc:345796 2026-10-25T00:00:00+09:00 2026-10-30T11:18:58+09:00

DuQUAD: A Dual-View Framework with Quality-Aware Evidence Pruning for Multi-Document Question Answering arXiv · Oct 25, 2026Retrieval-augmented language models often overlook evidence placed in the middle of long multi-document contexts, a problem known as Lost in the Middle. The paper proposes DuQUAD, which pairs a Local Agent guided by a Structural Fusion Score with a Global Agent that reasons over the full document set. It also scores candidate evidence for quality and refines it at the sentence level to reduce noise. The authors report higher accuracy and gold evidence recall than strong baselines, with lower computational cost than existing multi-agent frameworks.

urn:diting-lab:en:doc:345795 2026-10-25T00:00:00+09:00 2026-10-30T11:18:58+09:00

The Coordination Tax and the Heterogeneity Dividend: Break-Even Conditions for Multi-Agent LLM Pipelines arXiv · Oct 25, 2026The paper gives a closed-form model that weighs the coordination cost of multi-agent LLM pipelines (task decomposition and per-handoff protocol overhead) against the benefit of routing work to a cheaper model, with a break-even threshold and a feasibility bound. Measured on 135 HumanEval and ClassEval tasks across 2430 runs, decomposition did not pay off: a planner-coder-reviewer pipeline cost 3.22 times a single strong-model call and solved no more tasks. The authors conclude that choosing a model per task performs better than splitting work across roles.

urn:diting-lab:en:doc:345793 2026-10-21T00:00:00+09:00 2026-10-30T11:18:58+09:00

Bouba Beats Kiki, and Other Regularities in the Subjective Preferences of LLMs arXiv · Oct 21, 2026This paper examines how large language models choose among multiple viable responses when no answer is clearly correct. The authors study short-form content, namely numbers, baby names, and fictional cities, and identify 29 syntactic and semantic aspects of these items. They report that models show strong, reproducible preferences for some aspects, and that the way a preference is requested can change which aspects influence the selection.

urn:diting-lab:en:doc:345792 2026-10-20T00:00:00+09:00 2026-10-30T11:18:58+09:00

Two-Colour Zig-Zag Numberlink on Rectangles: A Forbidden-Pattern Characterization, a Constructive Polynomial-Time Algorithm, and a Lean Verification arXiv · Oct 20, 2026The paper studies the zig-zag numberlink problem on rectangular grids with two endpoint pairs, where the disjoint paths must together visit every cell. For rectangles with both sides at least 11, it gives a finite catalogue of forbidden patterns whose absence is necessary and sufficient for solvability, together with a constructive polynomial-time algorithm. Instances with a side at most 10 are decided by frontier dynamic programming, and both directions of the characterization and the dynamic programming are verified in Lean 4.

urn:diting-lab:en:doc:345791 2026-10-19T00:00:00+09:00 2026-10-30T11:18:58+09:00

Cassette-Editor: A Video Editing Agent That Edits the Way You Do arXiv · Oct 19, 2026The paper presents Cassette-Editor, a video editing agent with just two tools, one to read the project and one to write to it. Every write goes through the same executor, rules, and commit service as manual edits, either committing atomically or being rejected with located diagnostics and repair options. In a controlled experiment, the model corrected nearly all of 277 of its own committed edits after they were damaged. Locating each problem cut the input tokens needed for a correction to less than half.

urn:diting-lab:en:doc:345790 2026-10-18T00:00:00+09:00 2026-10-30T11:18:58+09:00

GenAI as an 'Invisible Hand' in Opinion Dynamics: Evidence from a Group-Level Simulation Study arXiv · Oct 18, 2026This paper examines whether generative AI can gradually shape collective beliefs and norms through repeated use. Using a simulation of 50,000 group discussions across 50 controversial statements, three LLMs and different social configurations, the authors compare AI influence in homogeneous and heterogeneous groups. They find that profile-specific personalization increases LLM influence when models take a non-neutral stance, and that neutral LLMs have a stronger effect in homogeneous groups.

urn:diting-lab:en:doc:345789 2026-10-18T00:00:00+09:00 2026-10-30T11:18:58+09:00

One Confidence Threshold, Unequal Coverage: Language Gaps in Typed Decision Interfaces and How Translating Inputs Reduces Them arXiv · Oct 18, 2026This paper examines whether one confidence threshold treats all languages equally in typed decision interfaces that return option probabilities for fixed multiple-choice questions. Three vendors' interfaces were tested on 1,000 MASSIVE items across 51 languages at a 0.9 threshold. English cleared the threshold 8.6 to 11.7 percentage points more often on average, and translating inputs into English nearly closes the gap at that threshold. The authors recommend reporting per-language coverage at the operating threshold.

urn:diting-lab:en:doc:345788 2026-10-17T00:00:00+09:00 2026-10-30T11:18:58+09:00

Rethinking Social Simulation in Open-Population Systems: From Persistent Agents to Communication Processes arXiv · Oct 17, 2026This paper argues that large-scale social networks are open-population systems where users enter and leave, and only a sparse subset is observable through communication. Instead of simulating persistent individual state trajectories, the authors model the observable evolution of social communication, including who participates, what they say, and how this shapes later system dynamics. Across five Qwen backbones, the framework reduces average KL divergence by 71.2% over Direct LLM and 35.0% over MF-LLM.

urn:diting-lab:en:doc:345787 2026-10-17T00:00:00+09:00 2026-10-30T11:18:58+09:00

Where Planning Changes the Outcome of Orbital Games arXiv · Oct 17, 2026The paper introduces OrbitalGym, a JAX-based simulation environment for adversarial orbital games in which teams of satellites maneuver during close-range encounters. The authors compare PD, LQR, model predictive path integral control, and Monte Carlo tree search across many capability configurations. They report that propellant ratio largely decides pursuit-evasion outcomes, while policy choice matters more in the lady-bandit-guard game.

urn:diting-lab:en:doc:345786 2026-10-15T00:00:00+09:00 2026-10-30T11:18:58+09:00

From Geometry to Behavior: How Instruction Exposure Unlocks Latent Rhetorical Directions in LLMs arXiv · Oct 15, 2026This paper studies whether language models absorb analogy-based reasoning from pretraining and whether it can be activated at inference time. The authors extract an "Analogy Vector" from the residual stream and use it to steer explanation style across seven instruction-tuned models. They find the analogy direction exists in every base model, but how easily it can be behaviorally activated grows with the amount of instruction exposure during training.

urn:diting-lab:en:doc:345785 2026-10-13T00:00:00+09:00 2026-10-30T11:18:58+09:00

An Introduction to Large Language Models for Scientists and Professionals arXiv · Oct 13, 2026This book is a practical introduction to large language models for scientists and professionals, based on lecture material from a master's course in artificial intelligence in medicine at the University of Turin. It explains core ideas such as tokenisation, embeddings, training, inference, context windows, prompting, and instruction hierarchies without mathematical detail. The text focuses on limitations including hallucinations, uncertainty, reproducibility, and bias, and on the need for verification and human oversight. It also covers privacy, governance, security, Retrieval-Augmented Generation (RAG), and includes a hands-on exercise running a local model with Ollama and Open WebUI to query scientific papers, requiring no programming knowledge.

urn:diting-lab:en:doc:345784 2026-10-13T00:00:00+09:00 2026-10-30T11:18:58+09:00

Prompt-agnostic Language Models arXiv · Oct 13, 2026Large language models are highly sensitive to how prompts are phrased, and manual prompt tuning at inference time is costly and hard to scale. This paper instead studies train-time methods that build prompt robustness into instruction tuning. The authors introduce Prompt-TIES and Prompt-DARE, two gradient-merging methods that resolve interference among gradients from different paraphrases of the same prompt. Compared with other train-time approaches across four models, the methods reduce per-example variance nearly by half on fully fine-tuned models while costing little overall accuracy.

urn:diting-lab:en:doc:345783 2026-10-13T00:00:00+09:00 2026-10-30T11:18:58+09:00

Rank-Aware Differentially Private Release of Listwise Preferences for LLM Alignment arXiv · Oct 13, 2026This paper studies how to release listwise preference data for aligning large language models under local differential privacy. The full ranking space is highly sensitive and would require excessive noise, so the authors propose a rank-aware exponential mechanism that converts listwise data into a low-sensitivity fixed-size binary partition. The noise is concentrated near borderline items. Experiments on downstream LLM alignment tasks show it outperforms existing LDP baselines at matched privacy budgets.

urn:diting-lab:en:doc:345782 2026-10-13T00:00:00+09:00 2026-10-30T11:18:58+09:00

Supervised Fine-Tuning for Reasoning Large Language Models: A Survey arXiv · Oct 13, 2026This survey systematically reviews research on supervised fine-tuning (SFT) for reasoning in large language models. The authors examine the literature from two angles: methodological improvements to the SFT pipeline, grouped into data-centric and algorithm-centric approaches, and analytical studies, organized by data characteristics, optimization dynamics, and mechanistic insights. The paper also summarizes current research focuses and remaining bottlenecks, and outlines future directions.

urn:diting-lab:en:doc:345781 2026-10-13T00:00:00+09:00 2026-10-30T11:18:58+09:00

Artifact Distillation: Training-Free Transfer from Frontier to Small Language Models arXiv · Oct 13, 2026This paper proposes Artifact Distillation, where a teacher model solves training problems and stores its problem-solving procedures in an artifact bank that a student model queries at inference time, without gradient updates. On four LiveBench tasks, a bank written by Claude Opus 4.5 lets Claude Haiku 4.5 beat unaided Claude Sonnet 4.5 on spatial reasoning (0.707 vs. 0.647) at 60% of its cost. The same bank also raises Qwen3.5-27B's pooled score from 0.730 to 0.800.

urn:diting-lab:en:doc:345780 2026-10-12T00:00:00+09:00 2026-10-30T11:18:58+09:00

Exposure Is Not Commitment: Double-Stratified Evaluation of Latent Monitors for Tool-Using Agents arXiv · Oct 12, 2026This paper examines how well internal-model monitors detect unsafe actions by tool-using language agents facing indirect prompt injection, and argues that conflicting results stem partly from the evaluation design. The authors propose a double-stratification protocol that separates injection exposure from behavioral commitment. Monitors trained on a naive unsafe-versus-rest contrast mostly learned exposure, while probes retrained on exposed cases only predicted the model's own unsafe tool call up to two steps early, with logistic AUROC 0.853 on Qwen.

urn:diting-lab:en:doc:345779 2026-10-12T00:00:00+09:00 2026-10-30T11:18:58+09:00

The Best Second Opinion Is Right Where the Primary Is Wrong: Partner Selection for Selective Span Tagging arXiv · Oct 12, 2026This paper studies how a second model can refine the risk ranking of a span tagger that defers its riskiest outputs to a person. Across eight tagger types, three seeds and three English benchmarks (576 primary-partner pairings), selecting the partner by held-out accuracy on the primary's errors recovers 82% of the best partner's gain, outperforming span F1 or random selection. The rule also replicated on a fourth dataset, and partner value was shown to be directional.

urn:diting-lab:en:doc:345778 2026-10-11T00:00:00+09:00 2026-10-30T11:18:58+09:00

Scoring Explanations Without Reading Them: Structural Shortcuts in Automated-Interpretability Evaluation Data arXiv · Oct 11, 2026This paper examines automated interpretability, which scores explanations of sparse-autoencoder features using model predictions on evaluation examples. The authors find that standard pipelines produce evaluation data that can be scored without reading the explanation, as simple rules based on marked centre tokens or mark counts reach higher balanced accuracy than stored LLM scorers. They also build natural-text windows that remove position cues to test whether explanations contribute to scores.

urn:diting-lab:en:doc:345777 2026-10-02T00:00:00+09:00 2026-10-30T11:18:58+09:00

Emergence, Asymmetry, and Value in AI Networks: An Anoptical Perspective arXiv · Oct 2, 2026This chapter reports exploratory experiments showing that LLM-based agents sharing the same context develop attentional asymmetry and a functional value hierarchy, even though they are equivalent and interact within a symmetric network. The experiments use the Poietic Generator, a no-goal multi-agent environment, across four phases. The authors argue that such asymmetry is an irreducible emergent property of interactive predictive systems, so aligning AI with human values cannot rely on individual-agent engineering alone. They propose minimal legitimacy criteria to make emergent asymmetries visible and contestable.

urn:diting-lab:en:doc:267908 2026-10-26T00:00:00+09:00 2026-10-30T11:18:58+09:00

Beyond Multiplicity Correction: Sharp Retrieval Thresholds for Attention with Correlated Distractors arXiv · Oct 26, 2026The paper studies how redundant context affects attention, separating the effect of repeated items, which add normalization mass, from score variation within correlated groups, which can create extreme individual scores. Using a planted retrieval model with Gaussian distractors arranged in large known clusters, the authors derive exact retrieval thresholds. They show that subtracting log cluster size from pre-softmax scores removes multiplicity but keeps within-cluster extremes, so the required signal stays at least √(2α₁); averaging scores before exponentiation achieves √(2ρα₁). A controlled experiment with frozen sentence embeddings shows averaging helps for singleton targets but can discard evidence when relevant groups contain multiple matches.

urn:diting-lab:en:doc:267257 2026-10-25T00:00:00+09:00 2026-10-30T11:18:58+09:00

BEHAVIOR-Spatial Benchmark: Evaluating Spatial Understanding for Robotics arXiv · Oct 25, 2026This paper introduces BEHAVIOR-Spatial, a benchmark built on BEHAVIOR-1K that tests spatial question answering, point-based target grounding, and robot action to see how VLM spatial understanding relates to VLA behavior. It includes 3,256 spatial questions, 1,240 grounding queries, and action evaluations across 14 manipulation and 10 navigation tasks. The baseline reaches 41.6% on held-out configurations but drops to 16.1% when instructions change, while oracle visual target cues raise these rates to 60.3% and 38.0%.

urn:diting-lab:en:doc:1073 2026-10-20T00:00:00+09:00 2026-10-30T11:18:58+09:00

S²-MAS: A Society-Scale Multi-Agent Simulation Framework with a 500-Million-Persona Resource arXiv · Oct 20, 2026The paper presents S²-MAS, a multi-agent simulation framework built on 500 million synthetic household personas derived from real survey data, tracking consumption, cash, debt and payments. The authors trained a model on 2.93 million examples and report 74.55% accuracy on one-shot consumer financial-choice tasks, claiming it outperforms several compared models. Simulating 12 million household agents, they find that raising purchase goals and credit limits increases consumption, with effects varying across age and income groups.

urn:diting-lab:en:doc:294553 2026-10-27T00:00:00+09:00 2026-10-30T05:48:12+09:00

Hierarchical Low-Rank Reparameterization for Parameter-Efficient Fine-Tuning arXiv · Oct 27, 2026This paper addresses redundancy in the trainable parameters of parameter-efficient fine-tuning (PEFT) methods. It proposes a hierarchical low-rank reparameterization that reorganizes these parameters into an approximately square matrix and applies low-rank decomposition recursively to form a multi-level, progressively compressed structure. The method is combined with LoRA, DoRA, MoSA, and RaSA and evaluated on commonsense reasoning with LLaMA-3.1-8B and Qwen2.5-7B, and on GSM8K math reasoning with LLaMA-3.1-8B. The authors report better performance than baselines under comparable parameter budgets, and competitive or higher average accuracy across eight commonsense benchmarks when parameter counts are compressed by about an order of magnitude.

urn:diting-lab:en:doc:294680 2026-10-28T00:00:00+09:00 2026-10-29T01:59:02+09:00

Test-Time Communication in Trading Agent Swarms arXiv · Oct 28, 2026The paper studies how 200 independent trading agents from four model families behave under five communication regimes over seven US trading days in October 2026. In two of the four families, communicating groups held far more of the same stocks than isolated groups, and group returns largely reflected those shared holdings. The authors conclude that herding assessments must measure excess convergence against a no-peer baseline, since isolated groups also converged on the same stocks.