In briefDynamic Positional Attention Modulation (DyPAM) for Efficient Fine-Tuning of Large Language Models
SummaryThis paper proposes DyPAM, a parameter-efficient fine-tuning method for large language models. It adapts how positional information affects attention by modulating query and key representations per input, dimension, head, and layer, matching the structure induced by rotary positional embeddings, without changing the pretrained backbone. The authors report that DyPAM consistently outperforms strong existing PEFT baselines on mathematical and commonsense reasoning benchmarks across multiple backbone models.
In briefOMIT: Measuring Framing-Invariant Omission Bias Under Philosophical Disagreement
SummaryThis paper introduces OMIT, a benchmark of 218 paired-frame scenarios across 10 moral conflict types, built from disagreements among five philosophical perspectives. Evaluating eight LLMs, the authors find omission bias is widespread and inversely correlated with model size within families. They also test four inference-time interventions, and report that prompting models to consider moral principles before answering yes or no reduces omission bias, though lower rates can coincide with shifts toward action-biased responses.
In briefScen-Opt: A Toolbox for Data-Driven Convex Optimization with Statistical Guarantees
SummaryThis paper presents Scen-Opt, an open-source software tool that combines convex programming with data samples and provides statistical guarantees based on scenario theory. It is implemented in Python, supports data-driven linear, quadratic, and semidefinite programming, and includes a web-based interface usable online or installed locally. The authors apply the tool to representative benchmarks to demonstrate its practical effectiveness.
Ben Wooding, Simone Garatti, Marco C. Campi et al.
In briefCHARTER: Auditing Reference Substitution in Compact-Evidence Evaluation for Pathology AI
SummaryThis paper examines how evaluation choices affect compact-evidence explanations of whole-slide pathology image models. In hierarchical pipelines, candidate filtering adds candidate-conditioned predictions, and changing the evaluation reference can alter both measured fidelity and comparisons between strategies. The authors propose CHARTER, a charter asking researchers to declare the target and reference, quantify candidate-induced prediction shift, and audit comparative conclusions. In their five-seed Random-K audit, 4 of 15 comparisons showed determinate reversals, and one ACMIL comparison changed under a native-ranking stress test.
In briefTreating Privileged Context as a Source of Policy Drift in On-Policy Self-Distillation
SummaryThe paper studies how the content and source of privileged context affect policy drift in on-policy self-distillation (OPSD), a method where a language model learns to match a copy of itself conditioned on privileged context. The authors train Qwen2.5-7B across nine combinations of content and source on three datasets, measuring accuracy, prior-task retention, reverse KL from the base policy, and parameter-update geometry. They find that changing content produces a wider median KL range than changing source, and that adapters sharing content have more aligned updates. The authors conclude that privileged context should be considered part of OPSD's stability design.
In briefDHCG: Dynamically Building Hierarchical Agent Collaboration Graphs for LLM Reasoning
SummaryThis paper proposes DHCG, a framework in which three modules (Planner, Worker, and Generator) progressively build a hierarchical collaboration graph for LLM-based multi-agent systems from the query and execution feedback. The Planner decides the roles and scale of agents and routes information among them, and is trained with action-aware preference optimization. Evaluated on code generation, mathematical reasoning, and domain-specific reasoning benchmarks, DHCG reports the best average performance among compared methods, improving on the single-agent baseline by 13.06 points and on static and dynamic multi-agent baselines by 2.77 to 8.02 points.}
In briefFreshCast: Timely, Complementary Retrieval for Frozen Time Series Forecasters
SummaryFreshCast is a plug-in retrieval framework that helps frozen time series forecasters by consulting historical patterns stored in a memory. Unlike methods that build memory only once, its streaming memory keeps adding new observations, and its correction weight is calibrated on a validation segment. Across seven datasets, ten forecasters, and three input lengths, it reduced average MSE for every forecaster and input length. Ablations indicate that post-training observations are a main source of the gain.
In briefHarness Engineering for Software Engineering Using Modular Executable Dev-Primitives
SummaryThe paper introduces Dev-Primitives, which pair each code repository component with a resident LLM so that components can reason in natural language, communicate, and modify themselves locally. It builds on this with HERMES, a framework that uses dependency-aware dynamic activation and a bug diagnosis mechanism to map execution evidence back to components needing revision. The authors report that HERMES outperforms matched baseline harnesses by 12.4% on average across four software engineering benchmarks.
In briefAgentic Semantic Sensing for Adaptive AI-RAN Resource Use
SummaryThis paper proposes Agentic SemS, a closed-loop framework for AI-enabled radio access networks that adjusts sensing configurations within a communication-feasible set. A profile-conditioned causal Transformer updates a semantic belief from streaming observations, and a utility network estimates the task benefit of the next observation block after sensing cost. On Widar3.0 with six emulated profiles, the resource-efficient setting reduces normalized cumulative sensing cost by 25.33% while achieving 85.79% Macro-F1.
In briefForecast Accuracy Is Not Trading Profit: Evolved Small Recurrent Networks for Stock Return Prediction
SummaryThe paper compares linear, fixed recurrent, transformer and mixing architectures with recurrent networks evolved by neuroevolutionary architecture search for stock return prediction. Each model is evaluated on forecast accuracy and on the net return of a daily long/short strategy, all fit on one pooled panel. The evolved networks rank first on both measures, while the second most accurate model loses money after positions and costs are included.
In briefCANDLE: A Method for Noninvasive Brain Source Imaging Using Cortical Null-Space Decomposition
SummaryThis paper presents CANDLE, a learning-based model that estimates cortical source activity from noninvasive electrophysiological measurements such as EEG. CANDLE learns a prior over the null space on subject-specific cortical geometries derived from T1-weighted MRI. Trained only on simulated data, it outperformed prior methods on simulated source estimation and generalized to intracranial stimulation localization and epileptogenic zone estimation.
In briefTTNet: Deep Learning Model Analyzing Table Tennis Players from Smart Racket Data
SummaryThis paper uses six-axis sensor data from smart table tennis rackets collected in the AI CUP 2025 competition. It proposes TTNet, a multitask deep learning model combining CNNs, ResNet, and self-attention to predict four player attributes: gender, playing hand, years of experience, and skill level. A two-stage training strategy with data augmentation and task-specific losses is used to improve generalization on imbalanced data. The approach placed second on the official competition leaderboard.
Ko-Hsun Chen, Xiang-Wei Ke, Hsien-Cheng Huang et al.
In briefNucleus Speculative Decoding: Plausibility-Aware Verification Beyond Exact Distribution
SummarySpeculative decoding speeds up text generation by having a small draft model propose tokens that a larger target model checks in parallel. The standard acceptance rule often rejects tokens the target finds plausible, wasting costly verification work. The paper proposes Nucleus Speculative Decoding (NSD), which also accepts draft tokens inside the target's nucleus, without training or extra forward passes. Across six models and five reasoning and code benchmarks, NSD reports the highest throughput, with speedups up to 5.16x over autoregressive decoding and 3.15x over standard speculative decoding.
WhyAdds a nucleus-based acceptance test to speculative decoding, exactly characterizing its deviation; validated across six models and five benchmarks with clear speedups.
In briefαTransfer: Reusing Merging Coefficients Across Model Sizes for Efficient Model Merging
SummaryThe paper proposes αTransfer, which searches for optimal model merging coefficients on a small proxy model and transfers them directly to larger target models. The authors show that models within the same family have highly similar performance distributions over merging coefficients across sizes. Reported results include a 6× speedup and 70% memory reduction on vision transformers, and a 20× speedup and 85% memory reduction on large language models, with comparable performance.
Shih-Cheng Huang, Zhi Rui Tam, Chieh-Yen Lin et al.
In briefStep-by-Step Delivery: Trading LLM Autonomy for Predictable Processes
SummaryThis paper proposes delivering standard operating procedures to an agent one step at a time over the Model Context Protocol (MCP), with each step returning a structured step_output. The approach trades some agent autonomy for process predictability and an auditable execution log. Across 15,475 trials in 13 SOP-Bench domains with four open-weight models, step-level delivery significantly raised process adherence and nearly eliminated answers produced without executing the procedure. Accuracy rose for the smallest model, while more capable models saw a small raw-accuracy decrease.
In briefDo I Need the Cloud? Step-Level Handoff for Small Model Agents
SummaryThis paper proposes STEPGATE, a framework in which a small language model acting as a local agent scores each of its actions and escalates only the difficult steps to a stronger model. On a 52-task single-step test, it reached 82.7% task success with 30.8% escalation. In multi-turn evaluation it reached 69.0% trajectory success using 30.0% cloud actions. The authors note the test sets are small and limited to one model family and scripted tasks.
In briefWhy stochastic gradient descent ascent is not optimal for certain nonconvex min-max games
SummaryThis paper studies how well stochastic gradient descent ascent (SGDA) performs in nonconvex-P{\L} min-max games when its timescale ratio and step sizes are tuned. For a fixed timescale ratio and non-increasing step sizes, the authors prove a complexity lower bound that matches existing SGDA upper bounds and separates it from Smoothed-AGDA. They also show SGDA can fail to find a stationary point when the timescale ratio is o(κ²), highlighting a fundamental limitation.
In briefSIFT: A Transformer model that ranks personalized search filters for multiple tasks at Airbnb
SummaryThis paper presents SIFT, a Transformer-based ranking model that recommends search filters to guests on Airbnb. It learns guest preferences directly from raw behavioral sequences instead of hand-engineered features, and serves several prediction tasks, including booking likelihood, filter engagement, and capacity thresholds. The authors report offline and online A/B test results and state that SIFT is now fully deployed in production.
WhyTransformer learns unified guest representations from raw behavior sequences, replacing hand-engineered features; strong offline gains and online A/B lifts with production deployment.
In briefA study on whether Agent Skills actually help AI agents complete tasks
SummaryThis paper empirically studies 87 SkillsBench tasks, measuring a skill's downstream utility as the pass-rate difference from a No-Skill baseline under the same model and configuration. The same skills helped some configurations and hurt others on 36.78% of tasks. Reranking candidates by support for required operations raised first-choice pass rates by 4.35 to 5.80 percentage points, and the authors derive 17 authoring practices plus guidance on organizing multiple skills.
In briefMASKerade: Token-routed mask experts for converting dense models into MoE models
SummaryThis paper introduces MASKerade, a method that converts a pretrained dense model into a Mixture-of-Experts (MoE) model. Instead of copying feed-forward networks, it learns binary masks over a frozen feed-forward network, defining each expert as a sparse subnetwork. A token-level router selects which masked experts to run and combine. The main configuration, with four 2:4 experts and top-2 routing, achieves the highest performance among compared baselines on five vision-language benchmarks with Qwen and Gemma backbones.
In briefCommon-Mode Errors Limit Low-Timestep Deep Spiking Q-Networks: Why spiking networks lose performance with fewer simulation steps
SummaryThis paper studies why spiking neural networks used in reinforcement learning degrade substantially when the number of simulation timesteps is reduced. The authors split Q-value estimation errors into common-mode and differential-mode parts and find that low-timestep networks suffer mainly from errors shared across actions. They propose CMC-DSQN, which uses an auxiliary ANN to compensate these errors and can be removed at inference. Experiments on Atari and MiniAtar report gains of nearly 20% over prior spiking baselines at T=2 and surpassing the ANN baseline at T=4.
In briefAdaptive Mean Estimation via In-Context Learning: A Gradient-Flow Analysis
SummaryThis paper studies how Prior Fitted Networks such as TabPFN learn to adapt location estimation to an unknown data family within a controlled setting. Gaussian data favor the sample mean, while uniform data are best estimated from extremes, and the authors show the learned estimator can switch between these behaviors. Through gradient-flow analysis of attention experts combined via softmax mixtures or GLUs, they derive efficiency results across Gaussian, uniform and mixture tasks and compare the two gating architectures.
WhyProves attention learns adaptive mean estimation with near-optimal rates across Gaussian, uniform, and mixture tasks, with end-to-end experiments supporting it.
Martin Eppert, Krishna Balasubramanian, Subhro Ghosh et al.
In briefThinkFuse: Trajectory-Aware Test-Time Fusion for Small Reasoning Models
SummarySmall reasoning models can struggle to recover once their chain-of-thought enters an erroneous path. ThinkFuse is a training-free test-time fusion framework that compares segment-level uncertainty shifts with trajectory-level uncertainty trends to locate unstable reasoning points and fuse auxiliary reasoning paths into the primary model's trajectory. The paper reports gains over baselines on mathematical and knowledge-intensive reasoning benchmarks, with fewer fusion triggers and fewer generated tokens.
In briefBenchmarking Western bluebird detection in wild outdoor images
SummaryThis paper introduces a new benchmark dataset for detecting and segmenting Western bluebirds, with over 6,000 labeled 4K images from 41 recording sessions, where birds occupy only small parts of each image. The authors evaluate supervised detectors, open-vocabulary models in zero-shot and fine-tuned settings, and segmentation methods. Supervised detectors perform most reliably overall, while fine-tuned YOLO-World becomes competitive, and zero-shot open-vocabulary models perform poorly. Failures are attributed to a combination of apparent scale, brightness, contrast, clutter, blur, crowding, and session variation, not object size alone.
Estela Monserrat Arriaga Santana, Julian Rosas Scull, Ibeth P. Alarc\'on et al.
In briefHow Novices Calibrate Reliance on AI Without Performance Feedback: Effects of Explanations and Self-Assessment
SummaryThis paper studies how novice users adjust their reliance on AI when no immediate performance feedback is available. In a between-subjects study with 110 participants completing a clinical entity extraction task with AI assistance and limited feedback, novices showed a systematic drift toward over-reliance when explanations were present. Higher self-reported task understanding was associated with more selective reliance behavior.
In briefThin evidence, strong identity: how language models fill missing financial facts with persona
SummaryThis paper examines how a large language model gives equity allocation advice when users omit financial details. Keeping finances identical and varying only the described investor identity, the authors measure how much the recommended allocation shifts across seven disclosure conditions. The identity effect grows as financial facts are removed, and with no facts the model invents incomes, debts and savings, with adverse advice occurring more often for larger households.
In briefThe Geometry of Empowerment: How an Agent's Control Over Its Environment Relates to Structurally Central States
SummaryThis paper links empowerment maximization with skill-learning methods to offer new geometric ways of interpreting and analyzing empowerment, which measures an agent's capacity to control its environment. It answers longstanding open questions about how empowerment relates to structurally central states. The authors also distinguish information geometry from reward geometry, with implications for building scalable empowerment-maximization methods.
In briefInformation-computation gap in sparse community detection, analyzed via minimax, Fisher information and belief propagation
SummaryThis paper studies community recovery in the sparse symmetric stochastic block model with q communities, average degree d and signal strength λ. Using statistical decision theory and Fisher information, it characterizes the Kesten-Stigum threshold dλ²=1 and the information-computation gap below it. The authors show that minimax risk equals Bayes risk under the uniform prior for a class of rules, and analyze cycle-count Fisher information and belief propagation. Experiments on networks with up to 3×10⁵ vertices confirm the threshold for q=2 and a hard window for q=5.
WhyThree decision-theoretic, Fisher-information, and belief-propagation characterizations of the Kesten-Stigum gap are proven, with experiments up to 3e5 vertices; contribution is largely theoretical.
In briefCompressing Gaussian Splatting sequences using standard video codecs
SummaryThis paper proposes GSCV, a method that compresses Gaussian Splatting (GS) sequences by converting them into smooth 2D videos encoded with standard video codecs. It addresses the weak inter-frame correlation of the vanilla PLAS approach, which lacks tracking information in practice, by introducing an Inter-PLAS method. Experiments show GSCV outperforms MPEG video and point cloud-based baselines in GS sequence compression.
In briefGeometry-constrained bidirectional point cloud alignment for thin heritage artifacts
SummaryThis paper addresses the difficulty of non-contact 3D reconstruction of thin, sheet-like heritage artifacts, whose front and back surfaces must be scanned separately and then aligned. The authors propose a geometry-constrained registration method that combines semantic-guided preprocessing, PCA-based normalization, and a thickness-aware strategy using the estimated physical thickness as a constraint. Experiments report competitive or improved results in most cases, particularly in projected area consistency and physically plausible front-back alignment.
In briefServeLearnBench: Can AI Agents Improve by Learning From Real Service Experience?
SummaryThis paper introduces ServeLearnBench, a benchmark for testing whether large language model agents can learn hidden, changing environment knowledge from interaction and outcome feedback. It spans retail support, banking, and sales-pitch generation, with 53 environment windows and 7,718 tasks. The authors evaluate five learning harnesses across six models and report a substantial gap between task capability and learning from experience. They also find that continual adaptation is costly, can degrade already-correct behavior, and is bottlenecked by insufficient exploration.
Haizhong Zheng, Yizhuo Di, Ranajoy Sadhukhan et al.
In briefIllusory pattern perception in LLMs causing spurious inferences
SummaryThis paper examines whether large language models, like humans, infer meaningful relationships in random data. The authors adapt classic psychology experiments into three tasks and compare model behavior directly with human behavior. They find that LLMs often show stronger illusory pattern perception than humans, and use Sparse Autoencoders to analyze the internal representations linked to this effect.
WhyFirst systematic study of illusory pattern perception in LLMs, adapting psychology paradigms to three tasks with human comparison and SAE-based mechanism analysis; novelty is clear, but experimental s
In briefRefraction correction for underwater 3D reconstruction using flat-port camera views
SummaryFlat-port housings on consumer cameras are widely used for underwater mapping, but refraction at the port causes bowl-shaped distortion in reconstructed scenes and camera paths. This paper proposes a physics-based refraction correction applied in image space before reconstruction, so corrected images can feed existing reconstruction and SLAM pipelines. Tested on two real underwater datasets with different scene structures, it removes reconstruction deformation compared with conventional and refractive Structure-from-Motion, while registering more frames with low reprojection error.
Chelim Lim, Tobias Fischer, Emilio Olivastri et al.
In briefOOPMAS: Object-oriented multi-agent systems that build a workflow for each query
SummaryThis paper presents OOPMAS, a training-free framework that generates both the set of agents and their coordination workflow for each individual query. Agents are defined as object-oriented classes with roles, tools, and persistent state, and workflows are executable main functions over these agent objects. A dynamic skill library accumulates lessons from execution feedback across optimization rounds. On a mixed benchmark of code, math, and QA queries, OOPMAS reaches 89.6% accuracy, 18.1 points above the strongest baseline.
In briefExtending pathwise gradients to discrete variables using finite-order relaxation
SummaryPathwise gradients work well for continuous random variables but cannot generally be exact for every differentiable function when the variables are discrete. This paper proposes a framework that builds finite-order exact pathwise gradient estimators for common discrete variables such as Poisson, choosing the least-norm solution among estimators unbiased for polynomials up to a given degree. In experiments on linear, nonlinear and hierarchical latent-variable models, the low-order methods match or improve tuned baselines and run faster.
In briefAttacca: Goal-directed control for long-horizon embodied agents
SummaryThis paper proposes Attacca, a method for training visual goal-conditioned policies on full search-to-interact trajectories, so embodied agents can carry out chained tasks where the target may not be visible at the start. It uses goal images decoupled from the execution scene and adds a target-mask prediction head and behavioral-phase conditioning. Evaluated in Minecraft, it reports 39.0-47.5% clean success, improving over the strongest baseline by 1.7-2.4x, with up to a 7x gain on long-horizon tasks.
In briefFrom Lab to Road: Testing Wearable Eye-Tracker Gaze Accuracy While Driving
SummaryThis paper evaluates how accurately wearable eye trackers estimate where a driver is looking under real driving conditions. In 41 on-road scenes where a driver fixated a vehicle's license plate, the mean gaze error was 4.58 degrees. Applying an offset estimated from indoor recordings reduced the error to 1.10 degrees and improved every scene. The authors note the offset varied between sessions, so reliable bird's-eye-view supervision may need online recalibration.
In briefPersistent Memory in Multi-Agent LLM Inference: Costs, Benefits, and Detectability
SummaryThis paper studies splitting long-context inference across cooperating agents to bound the KV cache used per call, and tests whether an added persistent memory tier helps. Decomposition lowers peak KV cache per query to 14.3 MiB, versus 35.5 and 35.3 MiB for single-pass and retrieval-augmented baselines. The persistent tier adds about 0.368 MiB of peak cache and shows no detectable accuracy change across eight dataset pairs. The authors argue this null result is structural and describe conditions and detection procedures for agent-memory ablations.
In briefHow quantization affects tool-failure recovery, and why the answer depends on the test setup
SummaryThis paper compares 8-bit and 4-bit quantized versions of Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct on 20 deterministic tool-use tasks and five prompts, measuring recovery from temporary tool failures. The direction of the 8-bit versus 4-bit difference changes across prompts, evaluation targets, and scoring policies. The authors conclude that one prompt, one task set, or one scoring rule cannot establish a stable robustness conclusion. They recommend matched-task comparisons, reporting full-pipeline success, stating the scoring policy, and quantifying uncertainty across tasks.
In briefAPEX: a learned controller that speculates more efficiently in LLM decoding
SummaryAPEX is a learned controller for speculative decoding in large language models, selecting a speculation method per request and adjusting draft length at each verification block. It balances decoding speed against wasted draft tokens while keeping the target model's verification procedure. Integrated into vLLM and tested with Qwen3-8B on six workloads, it achieves up to 5.24X speedup over autoregressive decoding. APEX-B also reduces wasted-token percentage by 41.0% relative to fixed n-gram speculation at k=16.
In briefA One-for-All Foundation Model for Clustering Attributed Graphs
SummaryThis paper introduces OFAG, a foundation model for attributed graph clustering that aims to work across different graphs without graph-specific training, fine-tuning, or hyperparameter tuning. It learns a reusable clustering strategy from synthetic attributed graphs and uses a dimension-agnostic signal-wise graph encoder to handle incompatible feature spaces. On ten datasets, one frozen model achieves the best mean performance and average rank across NMI, ACC, ARI, and F1, while finishing all datasets in 12.43 minutes total. Code and a pretrained checkpoint are publicly available.
WhyAdapts PFN ideas to attributed graph clustering with a single frozen model, validated on ten datasets against baselines.
In briefReading Router Logits Instead of Steering Them: A Non-Intrusive Multimodal Safety Detector for MoE VLMs
SummaryThis paper addresses safety risks in mixture-of-experts vision-language models, where harmful intent can arise from combining images and text. The authors propose a lightweight detector that reads router logits during prompt prefill to flag unsafe requests before generation, without changing model parameters or expert routing. Tested on Qwen3-VL and Kimi-VL, it substantially reduces safety errors on HoliSafe and generalizes to out-of-distribution benchmarks such as MISHard and MM-SafetyBench.
In briefTRACE: Rollout-Guided Quantization-Aware Training for FP4 RL of MoE Models
SummaryThis paper proposes TRACE, an FP4 quantization framework for reinforcement learning training of Mixture-of-Experts language models. It uses rollout-side quantization outcomes to guide training-side FP4 rounding decisions, reducing the discrepancy between the two quantized paths. A caching scheme selectively retains deeper-layer quantization information to limit storage and communication overhead. Evaluations on four large-scale MoE models report RL performance comparable to BF16 rollout and up to 5.4x rollout speedup.
In briefOTel: Open Telecom AI Datasets, Benchmarks, and Models
SummaryThis paper releases Open Telco (OTel), an open resource of derived telecom datasets for retrieval, reranking, instruction tuning, and safety/abstention. It provides 30 post-trained baseline models covering 10 embedding models, 3 rerankers, and 17 language models. After post-training on OTel data, embedding retrieval reaches 93.1% NDCG@10, reranking reaches 0.947 MRR@10, and language-model correctness reaches 87.8%.
Farbod Tavakkoli, Gregory Diamos, Kenneth Church et al.
In briefSix Childhood Covariates Beat All Transformers at Predicting Later Depression
SummaryThe study uses the British National Child Development Study to predict depressive symptoms at age 23 from essays written at age 11. A logistic regression using six childhood covariates outperformed every text-only model, including seven fine-tuned transformers, bag-of-words, frozen embeddings and four zero-shot large language models. Adding text scores did not detectably improve the baseline's AUC-ROC.
In briefST-Bench: A Benchmark for Multi-Agent System Generation on Scientific Research Tasks
SummaryThis paper introduces ST-Bench, a benchmark of 100 Earth science data analysis tasks (in hydrology, agriculture, and wetland methane research), expanded into 2,067 expert-validated queries. It compares five multi-agent system generation methods against single-agent baselines using the same GPT-5 backbone. Nine of ten multi-agent configurations outperform the cheapest single-agent baseline, with the strongest reaching nearly three times its composite score, mainly through better coverage. The strongest configuration takes about four times the single-agent inference time.
In briefContrastive learning of review aspects for explainable recommendation (CLARER)
SummaryThis paper proposes CLARER, a recommendation model that combines aspect features learned from textual reviews with rating information to improve recommendation accuracy and explainability. Rating features are learned with an MLP, while aspect-specific review representations use a transformer encoder and contrastive learning. A transformer decoder then generates explanations from the final user and item representations. The authors report superior recommendation accuracy and explanation generation over baselines on three benchmark datasets.
In briefLater Is Better: Token Reduction for ViTs Under Distribution Shift
SummaryThe paper studies training-free token reduction for vision transformers and how it performs on shifted, corrupted data. It finds that the depth schedule of token removal governs the accuracy gap to the uncompressed model. The authors propose a one-parameter late-concentrated power-law schedule that improves out-of-distribution accuracy at no extra inference cost across methods, backbones, and tasks.
WhyProposes a one-parameter late-concentrated token reduction schedule, validated across many methods, backbones, shifts, and modalities; strong evidence, useful for a subfield.
In briefTrustworthy Method Comparison with AI Judges: Estimation and Design Under Order and Batch Effects
SummaryThis paper examines whether the common practice of randomizing prompt order and averaging scores is statistically valid when large language models serve as automated judges. The authors approximate LLM evaluation with a class of Markov generalized linear mixed models, supported by out-of-sample predictions from three major commercial LLMs. They report that randomize-and-average leaderboard selection is consistent under a mild separation condition, and that a Williams square design can improve efficiency when item qualities are close. For group comparisons, naive averaging can give inconsistent conclusions because of the model's nonlinearity. The approach is illustrated with AI judges comparing two graphical model estimation methods.
In briefAdversarially Trained Linear Transformers as Optimal Robust In-Context Learners for Gaussian Mixtures
SummaryThis paper asks whether robustness gained through adversarial pretraining can carry over to unseen tasks without further adversarial training. For Gaussian-mixture classification, the authors show that a sufficiently deep linear transformer adversarially trained across tasks can asymptotically reach the robust Bayes error on new tasks via in-context learning from clean demonstrations. A standardly trained model cannot. The paper also analyzes convergence under gradient flow, an accuracy-robustness trade-off, and demonstration complexity.
WhyProves adversarially pretrained linear transformers asymptotically reach robust Bayes error on unseen Gaussian-mixture tasks, a novel theoretical result with rigorous analysis.
In briefFrom Evidence to Action: How Tool-Using Agents Fail
SummaryThis paper examines how AI agents that use tools to change external systems can act without properly supported evidence, even when their outcomes are correct. The authors find that failures often occur before execution, such as stopping investigations early or acting before required evidence is gathered, and that multi-action workflows reveal unmet prerequisites. They introduce SafeActBench, with 656 cases across six domains and five protocols, plus a provenance-bound Evidence Ledger and deterministic trajectory evaluator. The study concludes that failures stem not only from missing information but also from how agents use the evidence they have.
In briefHow Well Do LLMs Reason With Noisy Visual Evidence? A Benchmark
SummaryThe paper introduces VisualNoiseQA, a benchmark for active reasoning when visual feedback is noisy. A text-only LLM repeatedly queries a fixed off-the-shelf vision-language model treated as a stochastic sensor, using self-consistency across multiple samples as an uncertainty signal to decide what to ask next and when to stop. The authors evaluate several LLM reasoners on 1,000 instances spanning perception, chart understanding, and knowledge-intensive reasoning.
In briefCleave: Faster Tensor Kernel Compilation via Decoupled Algebraic Search and Scheduling
SummaryThis paper presents Cleave, a machine learning compiler that separates algebraic graph transformation search from operator scheduling by using symbolic shapes. It generates kernels for computations with multiple reductions, such as FlashAttention-style kernels, that previously required expert handwriting. On common LLM subgraphs, Cleave's kernels are up to 2.8x faster than the best baseline (1.6x on average), and compile time is 5.9x shorter than Mirage on average. On dynamic workloads, it achieves geometric mean speedups of 1.4x and 1.7x over FlashInfer's FA2 and FA3 backends.
WhySymbolic decoupling with a new Split operator cuts compilation time and parallelizes reductions, with measured speedups on LLM subgraphs and production traces.
In briefHigh-dimensional online calibration using harmonic weights
SummaryThis paper studies online calibration of multidimensional forecasts over arbitrary convex sets and error norms. For forecasting d binary outcomes simultaneously, it gives the first algorithm whose number of rounds is polynomial in d for fixed accuracy, requiring d^{O(1/ε)} rounds. The algorithm outputs a harmonically weighted distribution over harmonically smoothed past outcomes and works for every forecast set and norm.
WhyProposes harmonic-weight algorithm giving exponential dimension improvement for multidimensional calibration, with strong theory and notable impact on calibration research.
In briefBudget-aware LLM reasoning over medical knowledge graphs with citable evidence for post-discharge risk prediction
SummaryThe paper addresses post-discharge risk prediction from electronic health records, where important links between findings are missing from the record. It proposes BAR, a framework in which a large language model explores disease-specific medical knowledge graphs under a patient-specific budget and cites verifiable evidence. On MIMIC-III and MIMIC-IV across 8 diseases and 3 prediction horizons, BAR improves AUPRC by 3.4 points over the strongest baseline, raises citation precision from 59.8% to 77.9%, and uses only 62-65% of the budget cap.
In briefNash Social Welfare in Multi-Armed Bandits: Trajectory-wise and High-Probability Regret
SummaryThis paper studies fair multi-armed bandits under the Nash Social Welfare objective, which measures performance by the geometric mean of accumulated rewards. It proposes a trajectory-wise Nash regret that computes the geometric mean over full sample paths, and a high-probability Nash regret. The authors introduce a two-phase algorithm, RR-NCB, and show regret bounds that match the optimal rate, with simulations validating the theory.
In briefTraining Retrieval Models with Reinforcement Learning in Embedding Space
SummaryThis paper introduces RELER, a reinforcement learning framework that trains existing embedding models to retrieve directly in embedding space, aligning them with task-specific rewards rather than only contrastive objectives. On BRIGHT, a benchmark of reasoning-intensive queries, RELER achieves higher average nDCG@10 than InfoNCE and LambdaLoss when post-training BGE-M3 and Qwen3-Embedding backbones. In retrieval-augmented generation across seven QA datasets, jointly optimizing retrieval and answer rewards improves both retrieval performance and answer quality.
In briefSanSi: A Looped Typed Decision Model for Fast-Then-Deliberate Thinking
SummaryThe paper proposes SanSi, which converts a pre-trained looped language model into a typed decision model that outputs probabilities over declared options without generating text. Option probabilities are read after every loop and trained with a proper scoring rule, so one model serves budgets from one to eight loops. On 10,027 test decisions from 59 sources, SanSi reaches 72.0% accuracy, and it also extends solvable depth beyond training depths on two depth-controlled tasks.
WhyProposes looped typed decision models with proper scoring training, validated on 10k test decisions and depth-controlled tasks against baselines.
In briefFoveated Compression: Keeping Sharp Detail Only Where It Matters for Cheaper Vision-Language Models
SummaryThis paper addresses the high inference cost of visual tokens in vision-language models. It proposes Foveated Compression, which encodes an image once and mixes native-resolution and compressed visual tokens, using a learned selector to choose one of nine spatial cells to keep at full resolution. The learned selector beats random and fixed allocation at a 20.99% token budget but still falls short of simple whole-image resizing. The authors also report a large gap between the learned selector and a region-choice oracle, indicating room for improvement in region selection.
In briefStructure-aware keypoint localization for swallowing X-ray video analysis
SummaryVideofluoroscopic Swallowing Study (VFSS) is a gold-standard X-ray imaging test for diagnosing swallowing disorders. This paper introduces VFSSKep, a dataset that adds soft palate annotations and includes large amounts of unlabeled data. It also proposes S³KL, a semi-supervised framework designed to reduce spatial bias in fixed-layout medical X-rays. Experiments report state-of-the-art semi-supervised performance, with 25% labeled data plus unlabeled data surpassing fully supervised learning using 100% labeled data.
In briefPERSIST: Cross-Session Who-What-When Memory for Full-Duplex Voice Dialogue
SummaryThis paper presents PERSIST, a persistent memory system for multi-session, multi-speaker spoken dialogue that models who, what, and when. It organizes past conversations into readable event records and retrieves them by jointly scoring semantic content, speaker identity from audio, and temporal state, while reusing backbone representations to cut retrieval latency from 578.42 ms to 7.03 ms. The authors also introduce SpokenTrace, a diagnostic benchmark, on which PERSIST reaches 85.08% end-to-end task accuracy.
In briefSteerScope: A Benchmark Measuring Side Effects of LLM Steering Methods
SummaryThe paper introduces SteerScope, a multi-dimensional evaluation suite with 15 metrics for LLM steering methods, covering efficacy, side effects, generalization, and data dependence. It benchmarks 23 methods across 4 families, including prompting, LoRA, and SFT. The authors find that no evaluated activation steering method outperforms the Prompt Steering baseline in balancing efficacy and side effects. Under out-of-distribution prompts, efficacy is often preserved while side effects, especially in instruction relevance and fluency, become more pronounced.
In briefExact calibration and sharp risk bounds for ridge regression on random row subsets
SummaryThe paper studies ridge regression fitted on exactly s randomly chosen rows of a fixed design, with fixed responses. It identifies the unique penalty that matches a full-data ridge fit in expectation, which exists when s exceeds the target's effective dimension. The authors derive sharp centered covariance risk bounds and characterize the responses that attain the maximum risk.
In briefRefRoute: Cutting the Cost of Multi-Reference Image Conditioning with Compact Features and Spatial Routing
SummaryRefRoute is a method for generating images from many reference images, keeping each subject's appearance while combining them into one scene. It pairs low-resolution latent tokens with lightweight residual features from full-resolution pixels, and routes reference tokens to their target regions while limiting cross-reference attention. The authors also release the RefRoute-Data training set and the ManyRef100 benchmark with 10 to 17 references per example.
In briefStability of Transformers Mapping Sub-Gaussian Measures
SummaryThis paper mathematically analyzes measure-to-measure operators defined by transformers. It shows transformers map sub-Gaussian inputs to sub-Gaussian outputs and are Hölder continuous under the 1-Wasserstein distance. The authors also study a mean-field cross-attention analog and derive approximation guarantees for these transformers.
In briefJevCascade: Readout Stability and Compute Allocation for Prefill-Only Decision Models
SummaryThis paper presents JevCascade, a framework for allocating inference-time compute to prefill-only decision models that score menu candidates in a single forward pass without decoding. When only the menu changes and the input text stays fixed, the authors show that post-intervention accuracy can be predicted from the cached first-pass distribution, without labels or a second pass. The results cover six model families and sixteen datasets.
In briefWhen Old Facts Return: Re-reads, Reverts, and Limits of Temporal Memory
SummaryThe paper studies an ambiguity in memory systems: a retired old value can be wrongly restored when the same stale statement reappears, even though a genuine revert and a mere re-read can produce identical value sequences with opposite correct answers. Experiments on 130 atomic transitions derived from software fixes show that appending a verbatim re-read sharply lowers accuracy, while a guard that refuses to reactivate retired values restores it. The authors conclude that systems should distinguish observing a value from evidence that the value has changed.
In briefPredicting molecular and material properties with math-based topological neural networks
SummaryThis paper introduces MITNN, a framework that describes complex structures using several mathematical views, including topology, spectral theory, and discrete curvature, and combines them with topological neural networks. Systematic analyses show that performance depends on how these representations and architectures are paired, with some combinations beating single models and aggregating all components. MITNN reportedly outperforms existing methods on protein-ligand binding, metal-organic framework properties, mutation-induced protein solubility, and molecular toxicity prediction.
In briefEvaluating and Fine-Tuning Cell Nuclei Segmentation Models for Kidney Pathology
SummaryThis paper studies how to better segment individual cell nuclei in kidney pathology images, where general models often fail due to low contrast and dense nuclei. The authors combined easy, medium, and hard annotated data and compared seven fine-tuning strategies across nine model configurations. Fine-tuning improved all models; LSP-DETR reached an F1 score of 0.8725 using only hard data, and StarDist rose from 0.7380 to 0.8332 using only medium data. The results suggest that multi-difficulty annotations help adaptation, but the best data mix depends on the model.
In briefChoosing Negative Examples by Evidence: Explainable Negative Candidates for Recommendation
SummaryRecommender systems usually learn only from observed user-item interactions, and explicit negative feedback is often missing. This paper proposes implicit negative candidate discovery, which encodes observed customer behavior as symbolic rules, scores them by support, informativeness, and product relevance, and ranks them by evidence. An LLM then interprets the retained rules using business objectives and domain knowledge, and the interpretations are combined with statistical evidence in a final report. In an industrial B2B setting and five public datasets, candidate quality shows higher precision than the evaluated baselines, and symbolic selection improves downstream test PR-AUC by 12.5% over random selection with four negatives per positive in the industrial task.
Shreya Rajpal, Sonia Sharma, Swapnil Parekh et al.
In briefAgentMemGate: Preventing Unconfirmed Plans from Becoming Stored Facts in Assistant Memory
SummaryConversational AI assistants with long-term memory store facts from user messages for later use. The paper identifies speculation contamination, where a plan such as a possible move is recorded as already true. The authors propose AgentMemGate, a write-time gate that classifies extracted statements as speculation, completed event, correction, or other, keeping speculations out of memory. They also release a multi-session dataset where plans are confirmed, abandoned, or left unresolved. On the core benchmark, the method eliminated all observed contamination and raised task accuracy from 65% to 95%.
Chirag Sharma, Benjamin Fowlersmith, Karime Maamari
In briefWASD: Wasserstein-Based Distillation to Shrink Large Language Models
SummaryThis paper proposes WASD, a knowledge distillation method that transfers knowledge from a large language model to a smaller one. Unlike existing approaches that compare probabilities index by index, WASD uses a Wasserstein distance with a cost matrix derived from token embeddings to include token-level semantic information. It adopts the Sinkhorn divergence to keep optimization tractable. Experiments across multiple model families and scales report consistent gains on instruction following, mathematical reasoning, and code generation.
In briefWhat frame-level labels can and cannot do for locating small drones in thermal video
SummaryThis paper studies locating small UAVs as points in thermal infrared video using only frame-level presence or absence labels, without an external detector. A spatial feature extractor is first trained through classification and then frozen, and a readout is trained on the same frame labels to produce spatial score maps and point detections. The method is evaluated on the CST Anti-UAV and Anti-UAV410 datasets using localization hit rates and detection rates under false-alarm constraints, and compared with bounding-box detectors. The authors also explore applications on visible-light Airborne Object Tracking (AOT) imagery.
In briefIndependent Multi-Agent Learning with Counterfactual Semantic-Social World Models
SummaryThis paper studies fully decentralized multi-agent reinforcement learning, where each agent learns using only local information and no communication. The proposed CASTLE framework uses two frozen world models pre-trained offline: a Local Dynamics model summarizing local trajectories, and a Semantic-Social model predicting short-horizon task and social outcomes for candidate actions using counterfactual simulator rollouts. Their predictions guide an independent PPO policy. Across 30 matched seeds on Tag, Spread, and Adversary, CASTLE achieves the highest mean final score, exceeding the strongest baseline by 10.67, 6.46, and 0.33 normalized points respectively.
In briefTesting Whether Strategy Admission Gates Actually Help: Injected-Truth Study
SummaryThis paper studies admission gates, statistical criteria that quantitative strategy research must pass before a finding is accepted as not being search luck. The authors introduce an injected-truth protocol with a random-admission control adopting at the same rate as the gate, so a gate counts only if it outperforms that control. Gates remove false discoveries when signals are weak but reduce adoption to 1–7%, and add nothing when signals are strong. Criteria computed on absolute rather than excess returns reject every candidate, including true signals.
In briefDetecting AI-assisted Vietnamese writing from typing patterns, including deliberate evasion
SummaryThis paper studies whether keystroke dynamics can reveal when Vietnamese writing was assisted by a large language model. The authors built a dataset covering genuine composition, transcription, and paraphrasing, and modeled users deliberately altering their typing to evade detection. Sequential models generally outperformed feature-based methods, with transcription reliably detected, while paraphrased and manipulated samples were often misclassified as genuine. Adding adversarial training with manipulated data substantially improved separability and robustness.
In briefImproving argument mining data synthesis with adversarial reinforcement learning
SummaryArgument mining is limited by the scarcity of high-quality, structure-annotated datasets. The paper proposes an adversarial reinforcement learning framework in which a generator produces structured argument instances while a discriminator distinguishes real data from synthetic candidates, and both are optimized jointly. Experiments on three benchmark datasets, in full-data and low-resource settings, show consistent improvements in argument mining performance.
In briefRBMatch: Dual-Level Class Rebalancing for Semi-Supervised Building Footprint Extraction
SummaryThis paper addresses semi-supervised extraction of building footprints from high-resolution remote sensing images, where severe foreground-background imbalance biases pseudo-label generation and unsupervised optimization toward the background class. It proposes RBMatch, which rebalances classes at both pseudo-label selection and the unsupervised loss. Experiments on the WHU, INRIA, and Massachusetts datasets at labeled ratios of 1% to 10% report that RBMatch achieves the best building IoU and F1-score among the evaluated methods.
In briefAnchor-driven multi-modal, multi-scale expert selection for cancer survival prediction
SummaryThis paper proposes AM²ES, a framework that combines tissue slide images and gene expression data to predict cancer patient survival. It uses learnable semantic anchors to align the two data types and a hierarchical mixture-of-experts module that filters tumor regions at each magnification and selects the most informative resolution levels. Experiments on multiple TCGA cancer cohorts report state-of-the-art performance, and the authors show how molecular pathways influence expert routing across tissue scales.
In briefImproving CLIP's fine-grained vision via structure-aware latent masking
SummaryThe paper notes that CLIP aligns global semantics well but lacks fine-grained perception, which limits dense prediction and multimodal LLMs. It proposes SALM, an unsupervised embedding alignment framework that needs no image-text pairs, combining dual-matrix local alignment with latent mask modeling. A self-distillation variant, SALM-Self, is also introduced without external models. Experiments report gains in dense prediction and zero-shot accuracy.
In briefA Survey of Using LLMs to Assist Hardware Design Verification
SummaryThis survey reviews how large language models are used in hardware functional verification, including SystemVerilog assertion generation, testbench and stimulus generation, bug localization and repair, model and equivalence checking, SAT/SMT optimization, and agentic workflows. It finds that LLMs work best as reasoning, search, and orchestration components alongside deterministic tools such as simulators and formal engines. The authors argue that passing tool checks does not guarantee correct verification, and they identify semantic alignment, tool integration, generalization, and rigorous evaluation as key open challenges.
Hao Zheng, Jaime Rafael Imperial, Bardia Nadimi et al.
In briefBluffJAX: An Open-Source JAX Suite of Imperfect-Information Card Games
SummaryThis paper introduces BluffJAX, an open-source suite of adversarial imperfect-information games implemented in JAX. It includes well-known benchmarks such as Texas Hold'Em and Kuhn Poker, along with games like Bluff, Stud Poker, and Kemps that have seen little study in reinforcement learning. The implementations are designed for high-throughput parallel simulation on GPUs, with measured scaling up to hundreds of millions of samples per second across single and multiple GPUs. The authors also benchmark reinforcement learning, tree search, and game-solving algorithms to provide baseline results.
In briefDisentangling Dual References for Frequency-Aware Personalized Image Generation
SummaryThis paper studies personalized image generation conditioned on two references: a customization image and a color and style image. The authors observe that existing diffusion models entangle frequency bands during denoising, causing mismatches between text and the foreground or background. They propose Dual-FDM, which uses frequency-domain masks to separate bands, swapping mid-frequency background content for customization style transfer and low-frequency background content for color style transfer. The paper reports that experiments show Dual-FDM outperforms state-of-the-art diffusion models.
In briefEigenDEXplore: Structured Exploration for Dexterous Hand Manipulation Using Human Motion Priors
SummaryDexterous manipulation requires coordinated motion across many hand joints, which makes exploration in reinforcement learning difficult. The paper proposes EigenDEXplore, which adds perturbations along eigenvectors derived from human hand motion to independent joint-space noise, leaving the action space unchanged. Across multiple dexterous hands, the authors report consistent gains over joint-space and learned action-space baselines in grasping, in-hand reorientation, and contact-rich manipulation. The largest gains appear in settings with less reward shaping and curriculum design.
Harsh Gupta, Tyler Ga Wei Lum, Changhao Wang et al.
In briefVolume analysis of exact solutions for transformer length generalization
SummaryThe paper introduces normalized exact-solution volume (NESV), the fraction of a bounded parameter region where a transformer solves a task exactly for every input length. For single-layer transformers with log-scaled attention, it gives asymptotic NESV bounds for FIRST, MAJORITY, INDEX, and PARITY. For INDEX, a structurally modified model is studied, which reportedly reaches 85% accuracy at 10 times the training length versus 60% for the original.
In briefEIO-Agents: A Semantic Layer for Verifiable AI Agent Evaluation
SummaryThis paper introduces EIO-Agents, an open specification for interoperable AI agent evaluation with two layers. The Evaluation Intelligence Ontology (EIO) defines typed evidence, behavioral predicates, claims, and computable derivations leading to PASS, REVIEW, or BLOCK decisions. The Portable Evaluation Record (PER) is a content-addressed system of record that preserves the evidence-to-decision chain so evaluations can be re-derived and independently verified.
In briefEvaluating human-AI workflows for field research in viticulture
SummaryThis study tested two human-AI workflows in a California vineyard disease control project, using 2021–2024 scouting records and remote-sensing data across 140 hectares to forecast 2025 red-leaf symptoms. In retrospective simulations, adding model-informed row prioritization raised the share of newly recorded red-leaf observations encountered from 85.8% to 94.1%. The two workflows sampled 82% and 97% of their target vines, and 5 of 137 sampled vines tested positive for grapevine red blotch virus.
Niko Carvajal Janke, Daoyuan Jin, Shivranjani Baruah et al.
In briefCACHEFORGE: An LLM-guided method that generates cache replacement policies balancing performance and hardware cost
SummaryCACHEFORGE embeds a large language model in a hardware-aware loop that proposes new C++ cache replacement logic, which is then tested in a CRC-2 ChampSim simulator. The framework enforces feasibility using reward shaping, structural checks, mutation, temperature scheduling, and crossover. On SPEC CPU2006, the authors report total hit-rate gains of 27.36%, 19.69%, 13.72%, 13.15%, 11.83%, and 5.73% over MPPPB, ReD, Hawk-eye, SHiP++, LIME, and LRU respectively.
Kaushal Mhapsekar, Bita Aslrousta, Brijesh Kumar Bhayana et al.
In briefSENSE: A state-aware engine for generating emotionally navigable branching visual novels
SummaryThis paper presents SENSE, a framework for generating playable branching visual novels with multi-track emotional navigation. It combines a state-based narrative architecture called MIND, a structure analyzer, and a path-aware context management module. From minimal high-level inputs, it produces multiple intersecting routes while maintaining character consistency and narrative causality. Evaluations report gains over baselines in narrative diversity and asset integration, and preliminary human trials show directional improvements in emotional fidelity with comparable enjoyment.
In briefJoint Workflow and Prompt Optimization for User Behavior Simulation (SWORD)
SummaryThis paper introduces SWORD, a framework that simulates user behavior in information systems using AI agents instead of live users. SWORD jointly optimizes the structure of a multi-agent workflow and its natural-language prompts, guided only by a scalar task metric and without task-specific engineering. In controlled comparisons using the same backbone model, it outperforms prompt-only, workflow-only, and staged-optimization baselines, and it also improves accuracy over a strong published domain-specific baseline.
Nipun B Nair (1) Tongtong Wu (1), Hongzhi Yin (2), Hui Li (3) et al.
In briefMS-ECG-FM: A More Universal ECG Foundation Model Using Multi-Source Contrastive Learning
SummaryThis paper introduces MS-ECG-FM, an electrocardiogram foundation model trained by contrastively aligning ECG signals with multiple clinical note types, including echocardiography, radiology, and discharge reports. The authors evaluate it on an extended set of ECG detection benchmarks. They report that it comprehensively outperforms existing methods across the conditions ECG can detect, including reduced-lead configurations. Different report types improve representations for different diagnostic domains, and multi-source alignment captures their complementary information.
In briefMassive Activation Gating Channel (MAGC): A Single Channel Controlling Huge Hidden Activations in LLMs
SummaryThis paper studies massive activations, where a few hidden channels in large language models take extremely large values. The authors identify a single input-embedding channel, called the massive activation gating channel (MAGC), that controls whether a spike feed-forward network produces these activations. They verify its existence and effect across six LLMs from four model families. They also give a theoretical explanation: when MAGC is sufficiently large or small, the network's output reduces to a quadratic form that produces massive activations.
WhyIdentifies a single embedding channel controlling massive activations with a theoretical explanation, verified across six LLMs; strong mechanistic insight but limited broader application impact.
In briefDLoop: looping speculative decoding to cut verification passes and speed up LLM generation
SummaryThis paper proposes DLoop, a looped form of speculative decoding that speeds up autoregressive generation in large language models. It adaptively runs multiple drafting stages before verifying all accumulated draft tokens together. Loop-aware training keeps the draft model reliable in these additional stages by exposing it to its own hidden states for unverified tokens. According to the paper, across methods such as EAGLE-3 and DFlash, wall-clock speedup improves by 5 to 41 percent while decoding remains lossless.
In briefUniform discrete diffusion models shown minimax optimal for distributions with small effective support.
SummaryThis paper studies the statistical generalization of uniform discrete diffusion models, one of two main discrete diffusion approaches. It measures distributional complexity by the effective support size s_n(P_0) and derives expected total variation loss bounds scaling as O(√(s_n(P_0)/n)) and KL divergence bounds from n i.i.d. samples. The authors show the TV rate is minimax optimal and the KL rate is minimax optimal up to a log n factor.
WhyDerives generalization bounds for uniform discrete diffusion via effective support size, with TV rate minimax optimal and KL nearly so.
In briefDoes on-policy distillation used for safety create backdoor risks?
SummaryThis paper studies a security risk in on-policy distillation (OPD), a method for transferring capabilities from teacher to student models. It shows that a safety-aligned teacher with a hidden backdoor can pass malicious behavior to a clean student, reaching up to 70% attack success at a 3% poisoning rate. More training epochs and top-k KL increase this risk. The authors also propose Lazy Defense, which clips KL rewards to slow backdoor learning, and report that it delays backdoor transfer at low poisoning rates.</summary>
In briefSemantic communication using flow matching with time-aware structural residual injection for low-bandwidth image transmission.
SummaryThis paper proposes a semantic communication framework built on flow matching, which sends a selected text prompt embedding and a compact structural residual vector instead of the full image. The receiver's time-aware recursive network injects structural corrections progressively along the denoising trajectory. According to the authors, the approach improves structural fidelity and perceptual quality over existing text-driven semantic communication schemes under very low payload.
In briefSMART: Zero-Shot Sim-to-Real Robot Manipulation of Articulated Objects via Synthetic Pretraining
SummaryThis paper presents SMART, a system that uses large-scale synthetic demonstrations to train vision-language-action models for manipulating articulated objects such as doors and drawers. It introduces SMART-Sim, an articulation-aware simulation platform, and SMART-Data, which contains over 1 million demonstrations across 44 atomic task types, 5 robot setups, and 2,507 articulated objects. The authors report that a model pretrained on this data performs competitively on simulation benchmarks and achieves zero-shot sim-to-real transfer in real-world tasks.
In briefLoud and Clear: Dynamic Activation Steering to Make TTS Speech Clearer in Noise
SummaryThis paper examines whether a pretrained text-to-speech model can be steered, without retraining, to produce speech that is easier to understand in noisy settings. Drawing on the Lombard effect, it targets increased vocal effort and hyper-articulation. The authors propose a prompt-relative steering mechanism that avoids accumulating steering effects during generation and allows dynamic strength adjustment. Tests on seen and unseen speakers across several languages show speaker similarity of 89-95% and a 7-22% reduction in WER under background noise at 1 dB SNR.
In briefMatching Objects or Relations? Tracing Abstract Reasoning in Vision-Language Models
SummaryVision-language models perform well on visual benchmarks but systematically fail at tasks requiring abstract reasoning. The paper adopts the Relational Match-to-Sample paradigm from psychology and pairs it with an analysis of model internals. Four factors, including capability tier, model scale, objects per scene, and absence of per-object noise, shift models toward relational matching. Internal analyses indicate two competing circuits: an early one organizing images by surface object features and a later one organizing them by abstract relation.
WhyCombines developmental RMTS paradigm with mechanistic analysis revealing two competing circuits; multi-model experiments plus causal mediation and ablation give solid evidence.
Gouki Minegishi, Hiroki Furuta, Takeshi Kojima et al.
