In briefImage Generation for Newly Emerged Entities Using Web Search Agents
SummaryThis paper studies image generation for real-world requests about entities that appeared only days ago, where web visual evidence is limited and search results are confusing. The authors build an automated pipeline that creates hard prompts from recent headlines and release Savanna, a benchmark of 500 prompts across 16 subcategories. They propose a Locate-Corroborate-Isolate paradigm and train Ranger-8B on 15K trajectories, which reportedly achieves a search score of 67.2 and the best generation quality across the generators tested.
WhyProposes a new locate-corroborate-isolate paradigm and benchmark, with broad comparisons across generators and benchmarks, offering notable impact on the subfield.
In briefA framework for measuring creator communication in short-form video
SummaryThis paper presents a five-stage multimodal framework for measuring creators' themes and communication strategies on short-form video at scale. It combines platform metadata, speech transcripts, and vision-language descriptions of frames, using large language models to produce indicators and labels. Applied to 7,311 retained TikTok videos from sixteen creators, it finds differences in education, humor, and responses to current events. Agreement with expert annotators averaged Cohen's κ = 0.40, close to the experts' 0.42, though it varied across indicators.
In briefRényi-based concentration metrics for Kolmogorov-Arnold geometry in MLPs
SummaryThis paper addresses the interpretability of multilayer perceptrons by studying how concentrated the minors of a hidden layer's Jacobian are. It proposes a family of α-RC metrics that compare normalized minor magnitudes to the uniform distribution via Rényi divergence. The metrics are bounded between 0 and 1 regardless of network size, and for α = 2 they reduce to an explicit transformation of the Participation Ratio. Experiments on single-hidden-layer MLPs trained on a synthetic dataset are reported.
Juan Carlos Fern\'andez, Asunci\'on Jim\'enez-Cordero, Jos\'e Francisco Paris
In briefBeyond Marginal Influence: Interaction-Aware Memory Selection for LLM Agents
SummaryThis paper studies how an LLM agent can choose a small subset of its memory that preserves full-memory behavior for each decision. The authors introduce JointCoreBench, a benchmark with nine interaction families and exact sufficient-set oracles, and JointCore, a cost-ordered group-deletion baseline. JointCore recovers minimum-cost sufficient sets in all tested exact conditions while using fewer evaluations. The authors note it is an efficient benchmark baseline and do not establish deployment-ready memory compression.
In briefADT: An all-atom diffusion tokenizer that turns 3D molecules into compact discrete tokens
SummaryThe paper proposes the All-Atom Diffusion Tokenizer (ADT), which separates token granularity from reconstruction resolution. It uses one token per atom for small molecules and one token per protein residue that encodes both backbone and side-chain geometry. A diffusion model recovers fine atomic geometry that is lost to quantization. Experiments report strong reconstruction from compact tokens and high-quality autoregressive generation of small molecules and all-atom protein structures.
In briefSCALE: Better guidance-free image generation via self-contrastive likelihood alignment
SummaryClassifier-free guidance improves visual generation quality but makes sampling slower. This paper proposes self-contrastive guidance, which contrasts a model's predictions on real and self-generated samples to steer generation toward the true data distribution. It integrates this signal into a likelihood objective to form SCALE, a guidance-free training framework that produces enhanced predictions in a single forward pass. Experiments on diffusion and autoregressive models for class-conditional and text-to-image generation show SCALE outperforms CFG-mimicking baselines and improves further when combined with DPO.
In briefExpand-and-Prune: Test-Time Candidate-Space Adaptation for End-to-End Driving
SummaryThis paper proposes Expand-and-Prune, a test-time framework that adjusts the candidate trajectory set of end-to-end driving planners without updating learned parameters. It expands candidates using semantic driving intent and temporally propagated predictions, then prunes them based on uncertainty in intermediate trajectory scores. The authors report consistent gains across four candidate-based planners and four benchmarks, including a Driving Score rise from 68.2 to 75.9 on LangAuto.
WhyTest-time candidate expansion and uncertainty-based pruning without parameter updates, validated across several planners and benchmarks with comparisons.
Fabian Schmidt, Karol Fedurko, Markus Enzweiler et al.
In briefReasoning models can control their thoughts better when instructing themselves
SummaryThis paper examines whether reasoning models can control properties of their chain of thought (CoT). Prior work, which gave constraints in user prompts, found very low controllability, and the authors suggest this reflects how constraints were specified rather than an inherent inability. They introduce Self-Instructed CoT Controllability, placing the constraint as a self-instruction within the model's reasoning, and find substantially higher controllability. The paper also shows this ability could be used for monitor evasion, posing a risk to CoT monitorability.
In briefBeyond Reward Scoring: Online Multi-Agent Preference Learning for Multimodal Response Generation
SummaryThis paper proposes PrefMind, an online dual-agent reinforcement learning framework with a response Generator and a preference Critic. The two agents analyze and discuss which of two candidate responses better matches human preference, with human labels supervising their judgments and corrections. Experiments across three model scales report gains in both standalone preference judgment and downstream response generation.
In briefSample, Mix, or Verify? Planning Inference-Time Compute When Verifiers Are Imperfect
SummaryThe paper studies how to split a fixed inference-time budget among drawing candidate answers, choosing models, and running verifiers, when real verifiers make mistakes. It introduces SPLIT, which fits a probability formula from a small labelled pilot to plan budgets on unseen tasks. Across seven benchmarks, SPLIT reportedly stays within 0.02-0.8 percentage points of the best allocation in hindsight.
WhyProposes an exact formula, fit from a small labelled pilot, for allocating inference compute under imperfect verifiers; validated across seven benchmarks, 26 models and 175 agent systems, with broad p
In briefEvolution of LLM agents in a simple simulated market
SummaryThe paper simulates markets of twenty LLM agents solving arithmetic problems, where agents differ only in their system prompts and bankrupt agents are replaced by mutated copies of successful ones. A market rewarding accuracy yields 94% accuracy, while a market charging per token yields 10% accuracy with low token use, even though only tokens were priced. The authors present this as a concrete empirical example of how AI systems may evolve in unintended ways.
Charlie Pilgrim, Christopher Krupenye, Dora Biro et al.
In briefPreventing data hallucinations in AI agents for fixing optical network faults
SummaryThis paper demonstrates an agentic AI system for automated troubleshooting in optical communication networks. It integrates data references into the agent architecture to prevent data hallucinations. According to the paper, this improves the interpretability and trustworthiness of root-cause identification and reasoning.
In briefFLARE: A physics-based benchmark testing whether language-model agents can restore power feeders when data is incomplete.
SummaryAfter faults in distribution grids, service must be restored using incomplete, conflicting information while respecting electrical constraints. The paper introduces FLARE, a benchmark of 800 AC-feasible cases across five observation settings, where agents gather evidence, evaluate switches, and take restoration actions under fixed limits. Five instruction-tuned models were evaluated, with mean normalized regret between 0.082 and 0.148, and P95 regret between 0.814 and 1.000, revealing a serious failure tail. Shadow replay found that 96.4% of 15,706 actions rejected by guard checks would have caused physical violations, so the authors stress the need for external safety checks.
In briefLonGaussian: Streaming 3D Scene Reconstruction from Long Videos
SummaryThis paper presents LonGaussian, a streaming feed-forward framework that builds compact, renderable 3D Gaussian scenes from long image sequences with known intrinsics and unknown camera poses. It is built on a pretrained streaming 3D reconstruction backbone, uses multi-frame point tracks to merge redundant Gaussian predictions of the same surface, and applies rendered-depth consistency filtering to remove incorrectly occluding Gaussians. On the DL3DV and RealEstate10K benchmarks, it improves novel-view rendering quality while using substantially fewer Gaussians than the evaluated baselines.
In briefWho Pays for Efficient Inference? Confidence-Based Routing Between Small and Large LLMs Across Tasks and Languages
SummaryThis paper systematically evaluates confidence-based routing, which keeps easier inputs on a small language model and escalates difficult ones to a large model to reduce inference cost. The study covers three model families, four languages, and three task types, spanning 72 configurations and 42,288 predictions. Confidence routing helps on Belebele and SIB-200 but offers no consistent gain over random routing on MMLU-ProX-Lite, and global confidence routing can create language escalation gaps up to 30.33 percentage points, which equal-quota routing largely removes.
Md Sabbir Hossen, Anichur Rahman, Pabon Shaha et al.
In briefCan Parts Fool the Test? Counterfactual Pairing Cycles for Relational OOD
SummaryThe paper proposes counterfactual pairing cycles, an evaluation method that keeps two images and two captions fixed and changes only which image is paired with which caption, cancelling image-only, text-only, and additive shortcuts. The same contrast also yields a training loss for relational outlier exposure and a diagnostic for bilinear vision-language scores. Experiments on synthetic controls, GQA, and COCO show cycles separate coarse image-caption matching from directional role-binding failures. In a shortcut-contaminated COCO stress test, cycle training raised CycleAcc from 0.882 to 0.927.
In briefThink More or Read More? A Token-Budget Theory for Shared-Context Reasoning
SummaryThe paper studies reasoning systems such as multi-agent debate and self-refinement, where reading peers' or one's own drafts consumes tokens from a shared window and leaves less room for reasoning. The authors propose a token-budget theory that models accuracy against the remaining reasoning budget and uses a Condorcet jury model to predict the optimal number of agents. Across five open models and four tasks, self-consistency matches or beats debate in 93% of 980 matched settings, with the gap largest under tight windows.
In briefMGE-PO: entropy-aware credit assignment to improve math reasoning in language models
SummaryThe paper notes that GRPO gives equally rewarded responses within a question the same group-relative advantage, ignoring their generation uncertainty. The authors propose MGE-PO, built on four-quadrant (4Q) advantage shaping that combines verified correctness with rollout generation entropy: correctness sets the sign of the correction and entropy sets its size. After eight training epochs on Qwen3-8B-Base, adding 4Q raises six-benchmark macro avg@32 by 3.23 to 3.69 absolute percentage points.
In briefA mathematical study of how attention gating can make small differences inside Transformers grow instead of shrink
SummaryThis paper splits self-attention in Transformers into value transport, query gating, and key gating, and studies how these parts change small differences between internal representations. It defines gating-induced regime switching, where a difference shrinks under fixed attention routing but grows under full attention, and derives a spectral criterion that decides when such switching exists. Experiments across four Transformer architectures are used to validate the theoretical predictions.
In briefRIVET: Using past responses to train cost-aware routing for multi-turn LLM use
SummaryThe paper studies multi-turn LLM routing, where a model must be chosen at each turn before its response exists and success is only known at the end. It proposes RIVET, which trains two value estimators jointly during training, using logged responses as extra information, while routing at deployment uses only pre-call information. On the MINT benchmark, RIVET solves more held-out tasks than the strongest multi-turn baseline while using 74.7% less total call cost.
In briefKingfisher: Fast Personalized Image Generation Without Model Tuning
SummaryThis paper presents Kingfisher, a method for personalized image generation that does not require tuning the model. It decomposes the early residual scale codes of a visual autoregressive model into content and style components, then recombines the parts from two images into a compact semantic prefix. A frozen generator predicts the remaining scales to complete the image. On CSD-100, the authors report strong content preservation, the highest style and text alignment among the compared methods, and much shorter pair-specific preparation time than diffusion and VAR-based alternatives.
Dinh-Khoi Vo, Akihiro Sugimoto, Minh-Triet Tran et al.
In briefX-Splat: Unified Gaussian splatting for joint RGB and X-ray 3D reconstruction
SummaryThis paper proposes X-Splat, a method that reconstructs both RGB appearance and X-ray internal structure of objects using one shared set of Gaussian geometric parameters. Only surface-near primitives carry appearance attributes. Cross-modal alignment uses multi-view stereo and X-ray geometric proxies, and a surface-depth consistency loss from X-ray geometry guides the RGB reconstruction. The authors report on the InsideOut and their Flower & Insect datasets that the approach gives more stable alignment, better RGB rendering, preserved X-ray reconstruction, and a more compact representation than two independent geometries.
In briefMultiAgent-OPD: Training Hierarchical Multi-Agent Systems End to End with On-Policy Distillation
SummaryThis paper addresses how to jointly train a hierarchical multi-agent system, where a Leader assigns tasks and subagents carry them out. The authors propose MultiAgent-OPD, which combines on-policy distillation with reinforcement learning to train a shared-parameter Leader-Searcher-Answerer team end to end, using a frozen Teacher to supply local process supervision for the Searchers. Across seven datasets, the system reaches 38.68% macro avg@4, beating Prompted Critic by 4.06 points and the trained C-MAS + RL baseline by 2.61 points.
In briefRAP: A dataset that turns papers into posters using markup from human-made examples
SummaryThis paper introduces RAP, a dataset and benchmark for generating conference posters from research papers as a visual coding task. It contains over 25,000 verified paper-poster pairs from machine learning conferences held between 2022 and 2026, with a three-level content taxonomy and a held-out test set of 150 papers. The authors also present PosterRAG, a training-free multi-agent pipeline, and an SFT+RL training recipe built on the same data.
In briefA Framework for Characterising and Analysing RAG Techniques
SummaryThis paper addresses retrieval-augmented generation (RAG), a method used to reduce hallucinations in large language models. Because many RAG techniques exist, choosing the right one for a given problem is difficult. The authors develop a conceptual framework of main and sub-dimensions for characterising and analysing RAG techniques, built through qualitative analysis of attributes reported in the literature. The framework is intended as a pre-empirical filter to shortlist candidates before costly benchmarking.
In briefReLaDrift: Iterative language generation via drifting for flexible test-time compute
SummaryThis paper proposes ReLaDrift, which uses a shared Drifting Model to draft text from noise and iteratively refine it in a continuous latent space. Training samples the number of rounds and applies the drifting loss only at the final round, with gradients stopped through earlier rounds. On Sudoku Hard and GSM8K, it reports competitive or superior results against iterative generation approaches across various inference budgets.
In briefProcess reward modeling using Bayesian beliefs that account for reliability and conservatism
SummaryProcess Reward Models guide large language model reasoning during test-time scaling, but they can give inaccurate rewards due to limited supervision and mismatched policies. The paper proposes BayesianPRM, which treats process reward as a random variable and maintains a Bayesian belief over an ensemble-discretized reward hypothesis space. Two belief updates address PRM uncertainty and policy uncertainty, respectively. Experiments on multimodal reasoning benchmarks show more accurate reward estimates and improved test-time scaling performance.
In briefqGAMMA: A Framework for Finding Optimal DNN Mappings on Multi-Precision Accelerators
SummaryThis paper presents qGAMMA, a framework for exploring and optimizing how mixed-precision deep neural network models map onto general-purpose spatial accelerators supporting multiple fixed- and floating-point precisions. It searches the design space with a domain-specific genetic algorithm and an analytical performance model based on hardware synthesis results. The framework identifies the mapping and accelerator configuration that minimizes target metrics. Extensive evaluations across DNN workloads are reported.
In briefTwo TRECVID 2025 Video Tasks: MLLM Reranking and Frame-Subset Ensembling for Question Answering
SummaryThis paper reports the CERTH-ITI team's runs in the Ad-hoc Video Search (AVS) and Video Question Answering (VQA) tracks of TRECVID 2025. For AVS, a two-stage method first retrieves shots with several vision-language models using LLM-rephrased queries, then reranks the top results with a multimodal LLM without any training. For VQA, an audio-visual MLLM is fine-tuned and answers are aggregated across different random frame subsets, using majority voting for open-ended questions and mean-pooled logits for multiple-choice questions. The authors report the highest score on three metrics for open-ended answer generation and second place on multiple-choice questions.
Andreas Goulas, Damianos Galanopoulos, Ioannis Patras et al.
In briefNot All Test-Time Compute Is Evidence: Model-Specific Bias in LLM Reasoning
SummaryThis paper argues that when one reasoning model is sampled repeatedly, reasons at different depths, or evaluates itself, its errors are correlated, so extra compute from the same model adds little independent evidence. The authors propose DIA, which triggers only when intermediate and final predictions disagree and asks a small panel of heterogeneous models to solve the question independently, then aggregates their answers. Across seven benchmarks and two base models, DIA beats full chain-of-thought in 12 of 14 settings, with a 7.00% gain on MuSiQue, while activating on about 17% of examples.
In briefMAS-Bai: Adaptive Multi-Agent Translation for Bowen-Script Bai Language
SummaryThis paper addresses translation between Bai and Chinese when Bai is written in the Bowen script, where Chinese characters are often borrowed for sound rather than meaning, a problem that leads large language models to produce literal, incorrect translations. The authors propose MAS-Bai, which routes each input to either a lightweight retrieval-based path or a deeper path using etymological and syntactic agents, and uses cycle-consistency with a reverse translator as a reliability check. On a 1,536-sentence test set, MAS-Bai reaches 16.23 BLEU and 62.15 COMET, improving over Standard RAG by 8.11 BLEU and 6.85 COMET.
Jingting Zheng (1), Hong Ren (1), Yuehan Cui (1) et al.
In briefNon-parametric optimization for scalable learning in changing random decision problems
SummaryThis paper studies time-varying stochastic optimization, where decisions must be made under uncertainty and the underlying distribution changes over time. The authors build a stochastic variational framework based on Malliavin calculus, which yields non-parametric optimality conditions. They introduce the Stochastic Path Follower (SPF), a scalable deep-learning algorithm whose performance is insensitive to the parameterization dimension. On least-squares recovery and logistic regression under distribution drift, the authors report that SPF outperforms state-of-the-art learning-based and gradient-based methods.
Mohsen Amidzadeh, Lauri Viitasaari, Mario Di Francesco
In briefChecking AI-generated political social networks against survey data on partisan contact
SummaryThis paper uses large language models to generate political social networks from personas drawn from the nationally representative ANES 2020–2022 Social Media Study. The generated partisan exposure is compared with respondents' self-reported exposure to Democrats and Republicans. The authors find that different models vary in how strongly they separate partisans and in their errors for specific social groups, even when network structure looks plausible. They conclude that validation must occur at both aggregate and individual levels, since aggregate outcomes alone can hide local errors.
In briefFree-energy framework for spatial hearing: a hierarchical model of perceptual learning
SummaryThis paper applies the free energy principle to spatial hearing, using sound localisation as an example of listeners inferring sound direction. A hierarchical generative model describes how listeners update beliefs across trials as visual feedback arrives. The model captures individual learning trajectories, showing that lateral precision and hemifield disambiguation improve, and that polar precision also rises, most for listeners who started with the lowest precision.
Roberto Barumerli, Nils Meyer-Kahlen, Lorenzo Picinali
In briefFuzzInsight: balancing efficiency and effectiveness in coverage-guided DNN fuzzing
SummaryThis paper studies how to configure coverage-guided fuzzing that generates defect-revealing corner cases for deep neural networks. The authors propose Equalized Sampling, a seed-selection strategy that raises mean seed queue utilization to 77%, 13 percentage points above prior approaches. Across 84 configurations on MNIST, SVHN, GTSRB, and CIFAR-100, the non-structural criteria LSCD, LSC, and MDSC achieved comparable coverage with substantially less computation. Retraining with the generated corner cases improved generalization robustness.
Vivek Vekariya, Fathiyeh Faghih, Alexander Pretschner
In briefOfficeTown: Simulating Organizations to Train Evolving AI Agents
SummaryOfficeTown is a framework that generates simulated professional organizations from a natural-language description, in which members coordinate work, delegate tasks to assistant agents, and review deliverables while artifacts and decisions persist across tasks. The authors built OfficeTown Bench with 406 tasks and 4,836 evaluation criteria across 23 organizations in eight professional domains. Fine-tuning and on-policy distillation on this data substantially improved Qwen3.5-9B on external benchmarks including GDPval.
In briefProportionality in Ranking Compression: how to shrink a set of rankings while keeping fair representation
SummaryThe paper studies how to compress a given collection of rankings into a smaller collection, combining ideas from ranking aggregation and multiwinner voting. It proposes three proportionality notions: positional, pairwise, and proportionality for solid coalitions (PSC). The authors show that positional proportionality and PSC can always be satisfied and computed in polynomial time, while pairwise proportionality cannot be guaranteed in general but is achievable for single-peaked or single-crossing inputs.
In briefVisual Causal-chain Bookmarking: credit assignment for reinforcement learning with delayed rewards
SummaryThe paper addresses credit assignment in reinforcement learning, where delayed rewards make it hard to tell which decisions caused an episode's return. It proposes Visual Causal-chain Bookmarking (VCBM), which concentrates credit on decision points that determine the return. In a low-dimensional, fully observed delayed-reward watch-repair task, VCBM reached 95.07% correct decisions in a mean of 1,532.3 episodes across 20 seeds, faster than the LSTM-based RUDDER baseline (mean 3,327.1 episodes; p<0.001). In a partially observed AppWorld language-agent experiment, it used a heuristic based on state-mutating tool calls with retrieval, but with only one seed per arm, so it is a compute-limited demonstration.
In briefLangNCO: Turning Natural-Language Rules into Inference-Time Controls for Neural Route Solvers
SummaryThis paper presents LangNCO, a framework that converts instance-specific natural-language requirements into inference-time interventions for a frozen neural combinatorial optimization solver, without changing its architecture or parameters. A pretrained LLM translates each requirement into schema-validated DSL rules, which a deterministic executor applies as hard action masks, while Critical-Offset Decoding reallocates part of the rollout budget to biased rollouts. Evaluated on a CVRP policy across road closures, time windows, vehicle assignments, visit-order constraints and their combinations, the authors report 100% feasibility on four primary datasets.
In briefA stronger internal teacher for MoE models via bidirectional distillation
SummaryThis paper proposes MoEvolve, a bidirectional distillation framework for Mixture-of-Experts language models. It first restores and strengthens a model with a larger per-token activation budget, then distills those enhanced capabilities back into the original low-budget configuration. Across diverse MoE models and benchmarks, it improves initial models by up to 22.2 domain-average points and outperforms same-budget RLVR baselines by up to 6.5 points.
In briefAMEND: Constraint-Guided Multi-Agent Text Detoxification with Isolated Critics
SummaryThis paper presents AMEND, an inference-time framework for text detoxification that separates role-based candidate generation, isolated multidimensional evaluation, and constraint-first selection. AMEND Lite allows at most one targeted repair round, used only when toxic spans remain. On ParaDetox, one-round selection across diverse roles improves the joint metric by 0.0508 over the best single role. On a frozen confirmation test set, AMEND Lite scores 0.6597 ± 0.0062, outperforming unconstrained multi-round loops while cutting inference requests by 28.4%.
In briefWhat LLMs Do in the Shadows: Reading Context Reliance from Dark Space
SummaryThis paper studies how to detect faithfulness hallucinations in large language models, meaning whether answers rely on the provided context. The authors project hidden states onto a "dark space" that weakly composes with the unembedding matrix, and define the Dark Matter Ratio (DMR) as a direct score of context utilization. Combined with a parametric knowledge score, their DarkLens detector needs only a single forward pass and matches or outperforms baselines on three QA and summarization benchmarks across six open-weight models.
Alexandra Kuleshova, Andrey Volodichev, Daria Kotova et al.
In briefTreeOPSD: Tree-based posterior self-distillation for long-horizon agent reinforcement learning
SummaryThis paper introduces TreeOPSD, a critic-free reinforcement learning framework for training LLM agents on long-horizon tasks. It branches alternative continuations from shared prefixes, propagates terminal rewards back as turn-level credit, and maps that credit to a bounded teacher distribution that limits policy drift across sequential mini-batch updates. The authors report improved agentic performance across several benchmarks and model scales, especially on long-horizon tasks.
In briefDo speech BCIs need bigger models? Rethinking neural decoding without scaling up
SummaryThis paper questions whether large models are necessary for accurate speech brain-computer interface decoding. It proposes BEST, a two-stage framework that uses hierarchical intermediate representations and a compact speech-oriented decoder to map neural activity directly to text. On the Brain-to-Text '24 and '25 benchmarks, BEST reaches word error rates of 6.19% and 1.62% using only 3.85% of the parameters of comparable methods. The authors also report that decoder scaling gives non-monotonic gains and that cross-modal alignment matters more than model size.
In briefStructural Affinity: a cage-based shape design algorithm for coherent deformations
SummaryCage-based deformation controls 3D shapes through a small set of handles and is used in animation, interactive modelling and aerodynamic shape optimization. This paper introduces Structural Affinity, which precomputes an affinity matrix with mean-value coordinates to couple an inner control cage to a coarser outer cage, so handle displacements propagate to neighbouring handles. The authors report that across robotic control-morphology co-design, neural deformation transfer and wing shape optimization, the method reduces cage self-intersections and mesh distortion and lessens reliance on hand-designed structural loss terms.
In briefKnowledge-guided distributional prompt learning with visual hierarchy for detecting anomalies without training examples
SummaryThis paper addresses zero-shot anomaly detection, which finds anomalies in object classes not seen during training. It proposes HiDist, a hierarchical framework that models each prompt token as a Gaussian distribution and defines anomalous distributions as deviations from learned normal references. These representations are specialized for image-level detection and pixel-level localization. Experiments on seven industrial and medical datasets show strong cross-dataset performance.
In briefChinese Decision: Language models that output probabilities over a few choices in one pass
SummaryThis paper presents Chinese Decision, a family of Chinese-language models adapted from pretrained causal language models for choosing among a small set of options. Each question is read in a single forward pass, normalizing the logits of candidate letters instead of generating an answer. Built on Qwen3.5 0.8B, 2B and 4B backbones, the 4B model reports 85.97% overall accuracy on the CJ-Bench benchmark.
Qiang Yan (Institute of Computing Technology, Chinese Academy of Sciences)
In briefLongForge: Scoring and Improving Long-Context Training Data Using Evidence-Assertion Reasoning Graphs
SummaryCurrent quality control for long-context instruction-tuning data gives only a single score, without showing which document parts an answer uses or how many reasoning steps it contains, so the data cannot be audited or controlled. LongForge treats data quality as the alignment structure among document, question, and answer, pairing an analyzer that measures such data with a synthesizer that generates it. Measuring seven mainstream datasets, it finds breadth barely grows with document length and depth is generally low. Its two data mixtures, LongForge-1 and LongForge-N, outperform the strongest baseline by 2.2 to 4.7 points on average across six benchmarks.
WhyFormalizes long-context data quality via evidence-assertion graphs, exposing two unmeasured defects; gains of 2.2-4.7 points across two backbones and six benchmarks.
In briefExplaining Chess and Go in Natural Language
SummaryThis paper proposes an automatic metric for judging chess explanations, based on chess-engine evaluations. Tests show that current language models perform poorly, often relying on surface board details rather than transferable chess concepts. The authors introduce BinProp, a training-free method that helps these models discover reusable concepts and explain moves in natural language, nearly tripling average scores compared with direct prompting. The paper also reports that the method finds useful concepts in Go, draughts, and shogi.
In briefAlgorithm Modelization: Turning Optimization Algorithms into Trainable, Evolving Models
SummaryThis paper introduces Algorithm Modelization, which converts iterative algorithms into trainable models by treating their execution trajectories as computational experience. The authors model iterative optimization as an autoregressive process and propose a dimension-independent population-temporal architecture. An imitation-to-surpassing framework first learns from heterogeneous teacher algorithms and then improves through optimization feedback.
In briefNear-linear private learning for Littlestone classes
SummaryThe paper studies differentially private binary classification, where Littlestone dimension characterizes approximate private PAC learnability. For classes of Littlestone dimension d, the authors prove a sample complexity upper bound of O(d log d), improving on the previous O(d^5) bound. The main technique is a structural exponential mechanism over data-dependent candidate sets with a suitable rank, which for subspace classes equals the codimension.
WhyImproves private Littlestone-class sample complexity from d^5 to near-linear O(d log d) via a structural exponential mechanism, a significant theoretical advance.
In briefCross-Centaur: Language Models Predicting Individual Decisions Across Psychology Tasks
SummaryThis paper introduces Cross-Psych-101, a dataset of 591 participants and 261,551 valid decisions across 12 tasks, in which the same participants appear across different tasks. The authors train Cross-Centaur to predict a participant's actions in a target task using the source-task description and that participant's behavior in the source task. It outperforms off-the-shelf LLMs and Centaur on held-out participants, and probing suggests its hidden states encode individual tendencies that transfer across tasks.
Hikaru Asano, Hidezo Suganuma, Ryutaro Mori et al.
In briefQuAD: A Multi-Agent Framework for Quantitatively Assessing Qualitative Data Quality via Distribution Alignment
SummaryThis paper introduces QuAD, a framework for quantitatively checking whether a set of open-ended survey responses is suitable for a researcher's chosen analytical lens before analysis begins. It compares an intrinsic cluster distribution rebuilt from the responses with an extrinsic lens derived from metadata, and reads four quadrant metrics together to signal fitness for purpose. A content-aware multi-agent system is proposed to improve scoring consistency, and the approach is validated on synthetic data and a real-world social survey case study.
In briefCIIR: Class-indexed interface learning to link datasets without paired samples
SummaryThis paper proposes CIIR, a modular encoder-decoder framework that links independently collected datasets through a class-indexed address space without requiring paired or synchronized samples. The authors learn the address space from handwritten digits, clothing images, spoken-digit spectrograms, and hand signs. After freezing the base space, they attach the SVHN dataset by training only lightweight adapters and report end-to-end Address Transfer Accuracy in both directions.
In briefCan AI Interview? Comparing AI-led and Human-led Text Chat Interviews
SummaryThis paper compares the data quality of interviews conducted by AI chatbots and human interviewers via text chat. In a randomized experiment with 41 university students, interviews were led either by trained non-expert humans or a locally hosted 7B-parameter AI model. AI interviewers asked over twice as many questions and elicited about 65 percent more text, with responses nearly uniformly on-topic in both conditions. The authors conclude AI interviewing is viable for large-scale qualitative data collection.
Semra Guven, Tobias G{\aa}rdhus, Andreas Bjerre-Nielsen et al.
In briefGrounded credit assignment for multi-turn search agents in reinforcement learning
SummaryThe paper notes that outcome-reward reinforcement learning such as GRPO gives every token of a search agent's rollout the same advantage, failing to separate finding evidence from using it. It reports that Search-R1 improves accuracy while the share of correct Bamboogle answers supported by evidence falls from 84.0% to 42.9%. The authors propose Grounded Credit Assignment (GCA), which uses verifiable sub-outcomes to assign search and answer credit separately. GCA raises average EM over Search-R1 by 5.7 and 6.3 points on two Qwen models and lifts the evidence-supported share from 76.9% to 93.4%.
In briefLocal proxy networks for memory-efficient transformer training
SummaryThe paper studies how to train transformers locally with accuracy and low memory use. Each transformer block gets a local proxy network trained by distillation to predict the global output, so gradients can be propagated independently while approximating backpropagation gradients. On CIFAR-100 vision transformers, the method gives a better accuracy/memory trade-off than other local learning algorithms. It is also shown to be useful for fine-tuning large language models under a small memory budget, training faster and performing better than memory-efficient zeroth-order methods.
In briefLexPilot: Benchmark testing whether legal AI can proactively question laypeople in Chinese civil consultations
SummaryThe paper finds that when laypeople give incomplete case details, current large language models often give irresponsible answers with limited usefulness. The authors introduce LexPilot, a benchmark built from 432 real Chinese civil judgments across 36 causes of action, pairing complete and underspecified consultation histories with court verdicts. Experiments show models struggle to interact proactively in underspecified consultations, with judgment prediction averaging 23.84% lower than in fully specified settings.
In briefMetaCurate: Automatically Evolving Filters to Select Better Coding-Agent Training Data
SummaryThis paper introduces MetaCurate, a training-free framework that automatically evolves executable filters for selecting agent trajectories used in supervised fine-tuning of coding agents. It uses disagreements between the filter's rankings and pairwise LLM judgments to revise the filter's rules, rubric, and aggregation policy. In controlled SFT experiments, MetaCurate achieved the highest mean SWE-Verified score among the compared methods, including hand-authored filtering and a training-loop baseline that costs 3.5x as much to search.
In briefBlind Spots of Conformal Test Martingales: Undetectable Distribution Shifts
SummaryThis paper studies a blind spot of conformal test martingales, which sequentially test exchangeability using conformity-score p-values: any distribution shift that leaves the score distribution unchanged is invisible to them. The authors define the A-equivalence class and the closest KL-divergence element, the A-cryptic I-projection, with a Pythagorean split into detectable and cryptic parts. In experiments, EMNIST letters are conditionally resampled to match MNIST digit scores, shifting the distribution to a disjoint support while the martingale remains blinded.
In briefReliability-consistent appearance modeling for RGB-infrared multi-object tracking
SummaryThis paper proposes RCA-MOT, a framework for tracking multiple objects using paired RGB and infrared video. It keeps separate RGB, infrared, and fused identity representations, and uses a simple training-free estimator to produce per-branch appearance reliability for each detection. This reliability regulates identity branch weighting, calibrates borderline ReID candidates, and controls how strongly appearance memory is updated. The authors report new state-of-the-art results on VT-MOT and UniRTL, with higher HOTA and IDF1 than previous best methods.
In briefLED: Lookahead Editing to improve masked diffusion language models
SummaryThe paper proposes Lookahead Editing (LED), which adds a rewritable proposal state to masked diffusion decoding, performing several lookahead updates per step before committing tokens. Fine-tuned from Qwen2.5-7B and Qwen3-8B into LED-7B and LED-8B, the method yields average gains of +11.4 and +7.7 points over standard masked diffusion model SFT across eight benchmarks. Using SFT alone, LED-8B outperforms LLaDA2.1-16B Q-mode, which was trained with continued pre-training, SFT, and reinforcement learning.
In briefScene-Designer: Learning design intelligence for agentic 3D indoor scene generation
SummaryThis paper presents Scene-Designer, an agentic framework that generates 3D indoor scenes from natural language while learning both scene design and quality critique. The authors build two datasets, Scene-OTA-35K for design and error correction and Scene-OTC-15K for critique, and propose Scene-GRPO, a Critic-guided training method that reduces unnecessary trial-and-error corrections. Experiments report better physical validity, functional correctness, and layout coherence, and the learned Scene-Critic outperforms general-purpose models in scene critique.
In briefAligning multimodal embeddings with domain language for explainable predictions
SummaryThis paper proposes a framework to align any embedding space linked to interpretable auxiliary data with natural language, without requiring large captioned image datasets. It converts auxiliary data into partial captions, uses a soft contrastive loss weighted by property similarity, and evaluates alignment with rank-based concept retrieval. Applied to multimodal Earth embeddings for species distribution modelling (S2BMS and SatBird), the aligned concepts outperform previous Earth observation vision-language models on average concept retrieval and recover species-habitat associations (ρ=0.44).
Gabriel\.e Tij\=unaityt\.e, Thijs L. van der Plas, Mohamed-Bachir Belaid et al.
In briefWebCoder: Training a General Frontend Development Agent via Artifact-Grounded Reinforcement Learning
SummaryThis paper presents WebCoder, a frontend development agent trained with reinforcement learning that scores each step by its effect on a shared executable application. It introduces Webdev-World as the training environment and Webdev-Bench for evaluating multi-requirement projects. The authors report that WebCoder-27B improves requirement satisfaction on Webdev-Bench from 64.8 to 72.8 and its average visual and functional score on Vision2Web from 41.1 to 62.4.
In briefA graph-based model for predicting how structures move over time
SummaryThis paper introduces GraphNARX, a graph-based nonlinear autoregressive exogenous model for structural dynamics. It represents a structure's geometry and models its dynamic behaviour as interactions between its elements, with the next-timestep displacement as output. The model is evaluated for system identification and virtual sensing, achieving good accuracy and generalising to structures of different sizes.
In briefSeparating How Language Affects Tool Access and Post-Retrieval Safety
SummaryThis paper examines whether the language of a request, retrieved page, or response affects the safety of LLM agents that use retrieval tools. Testing ten models across seven languages, it finds that non-English harmful requests raise tool admission rates. Requiring answers in the page language also increases harmfulness, while translating pages alone does not. The authors argue multilingual safety evaluations should measure tool admission and response harmfulness separately.
WhySeparates tool admission from post-retrieval generation risks, proposes SLIP, and validates across ten models and seven languages with controlled experiments.
In briefToward High Semantic Fidelity: Hyperdimensional Symbolic Messages for Multi-Agent Communication
SummaryThis paper addresses semantic fidelity in multi-agent communication over non-ideal channels, covering both transmission and conversion fidelity. It proposes Hyperdimensional Symbolic Messages (HSM) and a framework called Herm, which uses entity-attribute-value encoding to reduce semantic entanglement and a receiver-centric ego-binding mechanism with similarity-aware dual-branch fusion to align semantics. An iterative decoding scheme based on algebraic inverse operations enables explicit message-to-semantic conversion. Experiments reportedly show Herm outperforms baselines across various noisy scenarios while keeping messages structurally readable.
In briefRethinking GRPO for Making Large Reasoning Models Safer
SummaryThis paper studies why GRPO, a common method for aligning large reasoning models to be safe, can give weak safety learning signals. It identifies two failure modes, advantage collapse and unsafe diversification, and proposes Beacon, which combines Safety Signal Recovery and Safety-Utility Decoupled Alignment. Experiments on four model variants and eight benchmarks report the best average safety and reasoning results, reducing safety error by up to 32.83 percentage points with low over-refusal.
In briefStructural roles of latent factors and how they are jointly identified
SummaryThe paper formalizes the structural role of a latent factor in dynamical systems, describing where it is expressed, whether it receives influence from the past, and whether it passes influence to the future. It shows that under suitable structural and variability conditions, all three roles are jointly identifiable under a common permutation at each time point. Guided by this theory, the authors propose RoleTrace, a framework that traces these roles in a shared latent coordinate system, validated on synthetic data and multiparty meetings.
In briefPretraining a graph-based relational model on synthetic databases to create frozen-use representations
SummaryThe paper pretrains the graph-based relational foundation model Griffin on large collections of synthetic relational databases. Its latent representations are passed to a frozen in-context predictor, with no training on target tasks. On 16 real-world relational classification tasks, a synthetic curriculum starting with schema-guided data and continuing with schema-agnostic data performed best. The results offer initial evidence that PFN-style synthetic pretraining transfers to graph-based relational models.
In briefBefore Retrieval Begins: Safety of LLM Agents' Decisions to Use Tools
SummaryThis paper studies whether large language model agents should call external tools, a decision that happens before the tool runs and that prior safety work has largely overlooked. Across seven open-weight models, the authors use controlled changes to input requests and find that neutralizing the requested operation produces the largest change in whether the tool is admitted. They also find that written refusals accompany only part of the harmful calls, with a relationship that varies by model. The authors conclude that tool admission should be evaluated and supervised directly rather than inferred from text refusal.
In briefParetoChat: An LLM Chat Interface for Interactive Multi-Objective Optimization
SummaryThe paper presents ParetoChat, an LLM-based expert-agent interface that converts users' natural-language preferences into structured parameters for multi-objective optimization, invokes a solver, presents candidate solutions, and iteratively collects feedback. The authors benchmark it against DESDEO's interactive methods, including reference-point and NIMBUS, using a simulated LLM decision-maker on knapsack and set-partitioning instances. They also apply it to historical courier-order assignment data to show how users can explore objective extremes and trade-offs through natural language.
In briefSame Person, Different Advice: Auditing AI Financial Advice via Repeated Risk Answers
SummaryThe paper shows that AI financial advisers often set stock allocation from a single risk-tolerance question, yet people answer it inconsistently over time. Seven large language models gave allocation advice for 800 household profiles differing only in the risk answer, and recommended stock shares moved by 10 to 16 percentage points on average. The authors argue that same-person retest audits should be a standard check for AI advisers, and that handling unreliable answers is a matter of client autonomy as well as statistics.
Soovadeep Bakshi, Apurba Bose, Aishwarya Bose et al.
In briefThought Blocks: Non-Autoregressive Latent Communication for Multi-Agent AI via Trajectory-to-State Compilation
SummaryThe paper proposes Thought Blocks, fixed-size latent key-value continuation memories built in one shot, so that LLM agents can communicate without generating latent trajectories step by step. Compared with autoregressive latent communication, it reports an average 2.5 times latency speedup. Compared with text-based multi-agent communication, it reduces text tokens by 45.7% and autoregressive steps by 84.8%, for an 8.6 times average speedup.
In briefA unified probabilistic framework linking embodied foundation models to robot policies
SummaryThe paper proposes an architecture-agnostic probabilistic formulation for embodied foundation models, such as vision-language-action and world-action models, using stochastic kernels. Conditional output distributions and action-realization kernels induce task-conditioned robot policies. The authors connect pretraining and post-training to policy deployment and instantiate the framework for OpenVLA and DreamZero.[]
In briefAgentic In-Context Learning: Which Algorithms Transformers Learn Through Closed-Loop Interaction
SummaryThe paper studies in-context learning in language model agents whose actions determine what evidence enters their context next. The authors represent an interactive learner by its state, update rule, and decision rule, covering active learning, bandits, and online reinforcement learning. They construct causal Transformers that approximate these algorithms over a finite horizon and bound the resulting return loss. Experiments show Transformers can learn three representative recursions, and that different reward signals yield distinct decision policies.
In briefTriFold: Compressing irregular 3D Gaussian scenes into structured latents for scene generation
SummaryThis paper introduces TriFold, a VAE that encodes native 3D Gaussian scenes into a fixed-layout triplane latent while jointly reconstructing geometry and appearance. It uses cell-wise learned queries to aggregate projected Gaussians and occupancy-guided set prediction to rebuild local Gaussian sets with varying counts. A latent flow-matching model built on this representation enables point-cloud-conditioned and single-image-conditioned scene generation. The authors report high-fidelity reconstruction and generation from a single partial-view image.
In briefDistill Until Divergence: Error-Truncated Hindsight Distillation for Agentic RL
SummaryThis paper studies reinforcement learning for agents making multiple decisions over long interactions. It notes that skill-based distillation can mismatch the current state after an early error, so it routes two hindsight skills by trajectory outcome and error boundary: positive skills reinforce successful workflows and pre-error decisions, while negative skills correct decisive errors. The post-error suffix is optimized only by GRPO. Experiments report competitive task performance and effective fine-grained credit allocation.
In briefSynerGest: Temporal-Spatial Synergistic Gesture Generation for Dyadic Conversations
SummaryThis paper presents SynerGest, a framework that generates speech-synchronized gestures for virtual avatars in two-person conversations. It uses a Contrastive Flow Matching objective that penalizes mismatched semantic pairs, and an Impulse-Trend Synergistic Attention module that models speech-driven rhythmic impulses separately from continuous motion trends. The authors report state-of-the-art results on two public datasets.
In briefDIFFLET: A fast inference engine for running diffusion transformer models on AWS Trainium chips
SummaryThis paper presents Difflet, a serving engine for diffusion models on AWS Trainium2 hardware. Five model pipelines share one compilation backend, joint attention is split across cores by attention head, and a host controller decides at each denoising step whether to rerun the denoiser or reuse its earlier output. Across five image and video model families, fused attention speeds denoiser steps by 1.2 to 4.3 times compared with PyTorch/XLA attention. On four of the five models, resident serving is 1.5 to 3.4 times faster than standalone runs.
In briefSpikeQ: a quantization method for spiking neural networks that targets spike errors rather than numerical errors
SummarySpiking neural networks use much less energy than conventional networks on neuromorphic chips, but these chips have limited memory that calls for low-bit weights and membrane potentials. The authors propose SpikeQ, a post-training quantization framework that minimizes spike errors instead of numerical errors such as mean squared error, choosing weight levels by simulating downstream neurons and using a reset-aware grid for membrane potentials. Tested on seven image classification, segmentation, and detection benchmarks, it stays within 2.59 points of full precision at 4 bits, with about 7.8 to 7.9 times smaller memory and 6.4 to 6.9 times lower energy use.
In briefWhere Performance Is Lost When Translating C to Rust, and How to Recover It
SummaryThe paper studies performance slowdowns in Rust code produced by the \crust{} transpiler from 20 real-world C projects. On eight analyzed projects, translated Rust was more than 3% slower than the original C on 12 of 19 operations, by up to 62.2%. The authors trace these slowdowns to nine recurring issues and present \ourtool{}, an LLM-based method that applies reusable optimization rules, validated for functional and performance correctness. On 41 operations from twelve held-out projects, it speeds up 40 operations over the \crust{} translations, 18 of which become faster than the original C.
In briefGR-OPD: Group-relative sequence credit assignment for on-policy distillation
SummaryThe paper observes that common token-level updates in on-policy distillation (OPD) assign credit locally and ignore how earlier actions affect later teacher feedback. The authors propose GR-OPD, which aggregates token feedback into calibrated sequence-level credit and uses same-prompt rollout groups to calibrate teacher guidance. Across five Qwen3 teacher-student settings and six math reasoning benchmarks, it reports higher macro Avg@K than vanilla OPD and more stable training under mismatched thinking patterns.
In briefWhen evidence-graph structure predicts answer correctness: a label-free abstention method for retrieval
SummaryRetrieval systems assign scores to answers, but a high score does not show an answer is trustworthy. This paper proposes a training-free abstention method that applies a boundary cut to selected evidence in an evidence graph, requiring no correctness labels. On the MemoryAgentBench multi-hop task, it raises the area under the precision-coverage curve from 0.736 to 0.785. The authors report that usefulness depends on graph construction, edge semantics, and relation type.|
In briefFixing frame and scale mismatches in two-stage 3D scene reconstruction from unposed images
SummaryThis paper studies fast 3D scene reconstruction from a few photos without known camera poses, using two frozen models: VGGT for pose estimation and TokenGS for Gaussian reconstruction. The authors find that the main failures come from a coordinate-frame inconsistency and a translation-scale mismatch, not from rotation error. These can be corrected deterministically, and the corrected pipeline reaches quality within 0.51 dB of the COLMAP reference without extra training, including in the sparse four-image setting.
In briefForward-citation supervision for generating actionable research limitations
SummaryThis paper focuses on the first step of automated research ideation: identifying actionable limitations from the literature. The authors build ActLim-Bench, containing 18,279 human-verified, evidence-linked supervision triples derived from forward citations, and propose a three-stage post-training pipeline combining supervised fine-tuning, verifier-filtered rejection-sampling fine-tuning, and reinforcement learning. The reported results show the highest recall among frontier models, preference in 64.9% of non-tied human comparisons, and improvements to downstream ideation systems.
In briefDo knowledge and reasoning benchmarks create stable internal signatures in LLMs?
SummaryThe study tests whether a predefined knowledge/reasoning label corresponds to a reusable internal representation pattern in five 7B to 31B language models. Probes evaluated on held-out whole tasks recover MMLU subject groupings, but prompt text alone predicts them almost as well, and the small activation advantage is not significant. After lexical adjustment, significance disappears, and no transferable knowledge/reasoning signature is identified at the tested sensitivity, though moderate effects cannot be ruled out.
In briefDesign Principles for Looped Language Models, Part I: Topology, Input Injection, and Recurrent State
SummaryThis paper studies looped language models, which reuse the same Transformer blocks across logical depth. The authors isolate recurrence placement, input injection, and recurrent state organization by training models from scratch at matched scale and token budgets. A prelude-recurrent-coda design improves context extraction and reasoning, while input injection mainly helps knowledge retrieval. The Huginn-style design outperforms Ouro at 730M parameters and also transfers to Mixture-of-Experts models.
In briefWhether and how a sparse hidden signal can be identified in a doubly sparse spiked Wigner matrix model
SummaryThe paper studies a matrix model where both the hidden rank-one signal and the noise are sparse, and examines when the signal can be distinguished from noise and recovered. For sub-Gaussian noise and bounded-away-from-zero spikes, a leave-one-out quadratic thresholding method identifies the support, while sign synchronization gives exact recovery for Rademacher spikes. Extensions cover continuous noise, and together with prior PCA results the authors address nearly the full parameter space.
In briefASE: Attribution-Guided Semantic Evolution for Time Series Forecasting
SummaryThe paper presents ASE, a modular framework that builds textual factors, measures their conditional contribution beyond numerical models, and deploys validation-selected numerical-text predictors. On four text-associated TESS datasets, its implementations achieve test MSEs below the corresponding published TESS scores. On dynamic CSI 300 stock ranking, the text-fusion version improves mean RankIC over its numerical expert and over matched shuffled text.
In briefCausal Value Attribution of Marketing Labels in LLM-Generated Ads: A Deconfounder Approach
SummaryThis paper studies how structured expert labels affect the causal performance of LLM-generated advertisements and proposes D-VAE, a counterfactual framework that disentangles expert labeling logic and estimates each skill's causal contribution. Tested on 29,177 foundation-makeup ad creatives, D-VAE outperforms state-of-the-art benchmarks. A field experiment found recommended configurations raised engagement by 32.7%, statistically indistinguishable from human experts.
In briefSemantic-Critical Regions for Guiding Video-to-Audio Generation (SCR-V2A)
SummarySounds in videos often relate to small, localized visual details such as object motion and contact, which existing video-to-audio methods capture poorly. The paper defines Semantic-Critical Regions (SCRs) as areas containing this sound-relevant visual evidence and proposes SCR-V2A, which localizes these regions and encodes them with adapter-based dual visual encoding alongside global context to condition audio generation. Experiments on in-domain and out-of-domain benchmarks report strong distribution matching and audio quality, competitive cross-modal alignment, and few trainable parameters.
In briefUsing user manuals to adapt computer use agents to new software
SummaryThis paper proposes MAGE, a framework that uses user manuals as a source of procedural knowledge for GUI-operating agents working on unseen software. At execution time it retrieves relevant manual content and generates explicit guidance, and it adapts the policy through environment interaction. During training, manuals are converted into executable task-plan pairs to produce trajectories without human demonstrations. Experiments across six applications report higher task success rates than strong baselines.
In brief"Deep lifts": Neural networks generalize even with a naive random optimizer
SummaryThis paper studies how well a neural network generalizes when its parameters are drawn uniformly at random from those that fit the training data within a bounded ReLU architecture. When the data can be explained by a smaller network, the authors show this random choice still generalizes at a rate governed by the smaller network rather than the full class's VC dimension. They use a proof technique called deep lifting to construct many near-equivalent parameters, and note the results survive label corruptions.
WhyProposes deep lifting to explain generalization of random ReLU interpolators beyond VC bounds; theoretically novel but with limited empirical validation.
In briefA Minimal ReLU Classifier for Non-Separable Data at the Edge of Stability
SummaryThis paper studies gradient descent on a two-neuron ReLU network trained with logistic loss on one-dimensional, non-linearly separable data, a setting where the objective is nonconvex and nonsmooth. The authors prove a finite-time loss bound for constant learning rates and show the gradient direction converges to an explicit max-margin vector. They also propose an adaptive learning-rate schedule with a loss rate of O(1/(T log T)), faster than the standard O(1/T), and validate the results on MNIST.
WhyAnalyzes edge-of-stability and adaptive step sizes for a two-neuron ReLU network on non-separable data, proving an O(1/(T log T)) rate with theory and MNIST checks.
In briefCPU-GPU Collaborative MoE Inference with Adaptive Expert Assignment
SummaryThis paper presents a system for running Mixture-of-Experts language models when expert weights exceed GPU memory. It assigns experts that the GPU lacks to either the CPU or the GPU based on each expert's token load, sending high-load experts to the GPU and low-load experts to the CPU. The assignment is adjusted so both execution paths finish as early as possible. Experiments on three MoE LLMs report prefill speedups of up to 51.9% and decode throughput gains of up to 39.5% over the strongest baseline.
In briefBRICE: A Brain-Inspired Encoder That Fills in Missing Parts of 3D Point Clouds
SummaryMost point cloud completion methods stack attention modules and rely only on feed-forward connections. Inspired by illusory contours, a brain mechanism for completing visual information in mammalian V1, this paper proposes BRICE, a plug-and-play recurrent encoder built on Predictive Coding to infer what is missing from incomplete 3D point clouds. Experiments report that BRICE improves various completion baselines on MVP2K, PCN, and PlantPCom datasets.
In briefAdjusting how varied versus polished an AI language model's answers are by mixing its aligned and base versions
SummaryAligning a large language model makes its outputs more polished but less varied, while the base model is more diverse. This paper proposes a training-free method that merges the lower layers of an aligned model with the higher layers of its base model to tune this balance. Experiments show that changing the number of layers taken from each model adjusts both diversity and quality, with merged models often showing higher semantic diversity than the aligned model. In most settings, the approach gave a better diversity-quality trade-off than temperature sampling.
In briefA noise-robust method for graph labels using neighbor label distributions
SummaryGraph Convolutional Networks work well on graph data but depend on clean labels. The authors show mathematically that label noise smooths a node's neighborhood label distribution, raising its entropy and reducing class distinguishability. They propose LENS, which identifies noisy nodes by measuring the drop in neighborhood entropy when a node is removed, and relabels them to keep useful feature information. The paper reports state-of-the-art node classification accuracy with computational cost about an order of magnitude lower than existing noise-robust approaches.
