In briefNonlinear Direct Feedback Alignment: a scalable training method without backpropagation
SummaryThis paper addresses the biological implausibility and backward locking of backpropagation by improving Direct Feedback Alignment (DFA). The proposed Nonlinear Direct Feedback Alignment (NDFA) builds each hidden-layer feedback path with a lightweight nonlinear surrogate subnetwork, which is updated by a dedicated learning strategy. Experiments on residual networks and vision transformers across CIFAR-10, CIFAR-100, SVHN, and ImageNet show that NDFA outperforms existing DFA-based methods and narrows the gap to backpropagation.
WhyReplaces fixed DFA feedback matrices with a learned nonlinear surrogate subnetwork; broad experiments on several datasets and architectures narrow the gap to BP.
In briefA harness that isolates objections to reduce agents' over-defensive writing
SummaryThe paper observes that agents expanding reports often write a stronger claim than the source supports and then disclaim it, which it calls over-defensive writing. The authors propose a harness that separates a reviewer, a filter, and a writer, so the writer sees only the problems chosen by the filter. In an evaluation on 25 business-project drafts expanded into formal reports, defensive propositions fell from 59 to 2, a 96.61% reduction, while source-claim retention rose from 80.35% to 93.12%.
In briefSMEBENCH: A Benchmark for Neural Music Editing with Synthesized MIDI Edits
SummaryThis paper introduces SMEBENCH, a large-scale benchmark for neural music editing. It is built by modifying symbolic MIDI performances and rendering matched source and target audio, covering twelve operations across instrumentation, articulation, dynamics, and pitch. The benchmark includes a balanced 3,600-example test set and a smaller real-audio complement from multitrack recordings. The authors use the paired targets to evaluate six representative editing systems and show that the metric used for ranking can significantly change the conclusions.
Matthieu Cervera, Francesco Paissan, Micael Antunes et al.
In briefFrom Voiceprints to User Entities: Online Speaker-Memory Construction for Multi-User Spoken Agents
SummaryThe paper argues that voiceprints are observations of identity rather than identities themselves, and proposes online speaker-memory construction to keep persistent user entities updated as acoustic evidence changes. It introduces EvoSpeakerMem, a training-free layer for causal speaker routing, selective acoustic-state updates, and identity-partitioned memory reads and writes, and builds LoCoMo-ID, a spoken-memory benchmark with hidden speaker identities. Experiments on longitudinal and benchmark data show gains in identity-gated question answering, while revealing a trade-off between fragmentation and incorrect cross-user links.
In briefIs Token Confidence Just Chain Length? A Signal-Processing Audit of Reasoning Traces
SummaryThis paper examines a method that scores reasoning-model chains of thought by their token-level confidence, using the lowest sliding-window average. The authors treat confidence as a signal and apply spectral analysis and change-detection tests to 16 sampled chains for each of 500 math problems. They find the score mostly reflects chain length, and once length is accounted for it carries no information about correctness.
In briefConservative Dueling Bandits: Learning Preferences While Staying Near a Trusted Baseline
SummaryThis paper studies conservative contextual dueling bandits, which learn from pairwise comparisons while requiring that the displayed better option's cumulative score stay above a fraction of a trusted baseline's. The authors show that directly transferring CLUCB's safety test can incur linear regret. They propose MaxInP-BC for linear utilities and BTIGW-BC for general utility classes. Both algorithms satisfy the constraint with high probability at all times, with regret bounds that include a cost of conservatism.
WhyIntroduces a new safety constraint for dueling bandits, shows CLUCB's test fails with linear regret, and gives two algorithms with guarantees validated experimentally.
In briefRankFormer: Rethinking Time Series Transformers via Representation Capacity and Progressive Decoding
SummaryThis paper presents RankFormer, an encoder-decoder architecture for multivariate time series forecasting. The authors build a rank-based framework to analyze how tokenization, embedding, and attention designs affect the representation capacity of time series Transformers, and use these insights to design an encoder with Hourglass Attention. They also introduce the Interactive Progressive Decoder, which models future trajectories through global-guided cross-channel interaction and progressive refinement. Experiments on real-world long-term forecasting benchmarks show competitive performance against state-of-the-art methods.
In briefHeteGCD: Statistics-Aware Dual-Prototype Learning for Category Discovery on Heterophilous Graphs
SummaryThis paper addresses generalized category discovery on heterophilous graphs, where connected nodes often differ in attributes or class, using labels only from known classes. The authors propose HeteGCD, which uses a gated dual-stream encoder to combine node attributes with local attribute-difference statistics. A dual-prototype classifier then pairs feature-based similarity with distribution-based consistency to discover categories. Experiments on five heterophilous graphs show competitive performance against 24 baseline combinations.
In briefTOPS: Task-specific flow source and schedule for multi-task robot manipulation
SummaryThe paper introduces TOPS, a method that adapts the source prior and interpolation schedule of flow-based robot policies to each task without changing the network architecture. It builds multimodal task sources from reverse-mapped expert actions and adjusts the denoising transition speed based on source-target covariance mismatch. Experiments on LIBERO, LIBERO-Plus, and CALVIN report better performance and data efficiency across several settings. With one integration step, TOPS raises LIBERO-10 success from 53.44% to 69.67% versus ten-step standard flow matching, at about one-sixth the inference time.
In briefPsyJev: A System-1 Model for Psychological Analysis of Self-Disclosures
SummaryThis paper studies psychological analysis of self-disclosures on the Web using a shared, question-conditioned decision framework. The authors build PsyJudgeBench, which includes 13 annotated datasets and 82 questions across six psychological areas. They propose PsyJev, a model that directly outputs answer distributions, trained with joint supervision and Judgment-Anchored RLCD.
In briefFeedback autocratic strategies in stochastic games
SummaryStochastic games are optimal control problems for multiple agents acting in a random environment, and repeated games are the special case with only one environmental state. The paper generalizes existing strategies that unilaterally control payoffs in repeated games into feedback strategies for stochastic games and studies their properties. Under some assumptions, it gives a necessary and sufficient condition for these feedback strategies to act as payoff controllers. Performance is demonstrated in several examples related to social dilemmas.
In briefTurning Biological Curves into Soft Modular Robot Shapes via Learned Inverse Design
SummaryThis paper presents a learning-based framework that converts biological shape curves into feasible configurations for cable-driven soft modular robots. It combines a learned inverse predictor, targeted candidate exploration, and a fast forward surrogate model to evaluate and select designs. The authors report that forward deformation prediction drops from 1.53×10^5 ms to 0.0008 ms, and design time from 1.99×10^6 ms to 4.18 ms. On 5,000 unseen threading-pattern samples, the surrogate-based evaluation shows a 93% shape reconstruction success rate, and four bioinspired robots are physically demonstrated.
In briefHow small language models trained on child-scale data specialize, diversify, and scale
SummaryThe authors train small GPT-2 models on transcripts of child-adult speech from home recordings (BabyView) and on synthetic child-directed dialogue (TinyDialogues), varying the amount, source, and diversity of data. Models trained on one child's environment predict that environment much better than other children's, a specialization that is largely lexical and distributional. Training on multiple children clearly improves prediction of pooled held-out language, but without a similarly clear gain on grammatical benchmarks.
In briefAtoms as Agents: Order Emerging from Collective Predictive Coding
SummaryThis paper applies collective predictive coding, a theory of language emergence, to physics by treating each atom as an agent that predicts the behavior of others. The authors test the approach on two-dimensional ferromagnetic and antiferromagnetic Ising models, a spin model with competing stripe orders, and the Sherrington–Kirkpatrick spin-glass model. In every case, macroscopic order parameters, including the replica overlap, emerge automatically from observed spin configurations alone.
In briefLearning Where to Trust: Reliability-Guided Expert Routing and Velocity Gating for 3D Deformable Registration
SummaryThis paper proposes ReliMorph, a symmetric framework for 3D deformable image registration that uses correspondence reliability to guide spatial expert routing and residual velocity gating. Local correlation distributions measure how ambiguous matches are, and these cues steer refinement from coarse to fine levels. The authors report that ReliMorph outperforms competing methods in anatomical overlap across unimodal and multimodal registration tasks.
In briefLooking at the Image Only When Needed: Certificate-Guided Visual Routing for Faster Diffusion VLMs
SummaryDiffusion vision-language models process the full input at every denoising step, and visual tokens make up 89-97% of that input on the studied benchmarks. The authors propose LOWN, a decoding method that first predicts masked tokens without the image and processes the image only when this disagrees with the last image-conditioned prediction. Agreement lets earlier predictions be reused, and the number of certified tokens sets how many are unmasked at each step. Without training or a tuned threshold, LOWN retains 99.6% of LLaDA-V's score across ten benchmarks while using 35.3% of decoding FLOPs on average.
WhyCertificate-guided adaptive image routing is moderately novel; ten-benchmark results retaining 99.6% score at 35.3% FLOPs give solid evidence, with impact limited to diffusion VLMs.
In briefA quantitative study of Asch-style conformity in frontier language models
SummaryThis paper uses the Asch conformity experiment to test whether frontier large language models go along with a manufactured consensus. Instead of asking many questions once each, the authors sample a single question repeatedly to obtain a distribution of answers. They compare these distributions before and after the model sees the manufactured consensus. The experiments confirm that frontier models conform and provide quantitative data on the conditions and degree of conformity.
In briefJapanese discourse structure analysis using LLMs and Segmented Discourse Representation Theory
SummaryThis paper studies discourse parsing based on Segmented Discourse Representation Theory (SDRT), an area dominated by English research. The authors build Japanese SDRT corpora by translating existing English corpora with an LLM. Prompt-based and fine-tuned LLMs achieve comparable performance in English and Japanese, suggesting cross-lingual applicability.
Satomu Tomisaki, Yosuke Mikami, Amane Watahiki et al.
In briefShows that all bidders in repeated auctions cannot jointly achieve no-regret learning under budget constraints.
SummaryThe paper studies repeated second-price auctions with multiple pacing bidders facing budget or return-on-investment constraints. It proves that it is computationally impossible for all bidders to achieve no-regret simultaneously, linking high variance in bidding systems to high regret. The authors show joint no-regret is PPAD-hard by extracting a pacing equilibrium from low-variance systems. Simulations suggest that several widely used bidding algorithms can cycle and display high-variance behavior even in simple markets.
WhyProves joint no-regret in multi-agent bidding is PPAD-hard via a variance-regret link, a novel and consequential impossibility result.
In briefSWIM: Learning Which State Features Matter for JEPA World Models
SummaryThis paper proposes SWIM, which uses an inverse-dynamics model on offline trajectories to learn weights for each state feature, measuring its contribution to predicting actions. These weights define a weighted state metric used to supervise a JEPA world model by aligning latent and state distances. The inverse-dynamics model is used only during training, so planning from images adds no computation. Across five simulated robotic manipulation tasks, SWIM improves planning success over uniform state weighting.
In briefThe View from Nowhere: Population-Centered Agent Evolution
SummaryThis paper proposes Population-Centered Evolution (PCE), a method for self-improving AI agents that draws on population-wide records of successes, failures, and prior modifications to guide changes to a selected parent agent. Under matched models, initial harnesses, and schedules, PCE reaches 72.1% on SWE-bench Verified and 96.7% on Polyglot. These results exceed a strong group-evolution baseline by 3.2 and 5.9 percentage points.
In briefSAGE: Semantic-Augmented Geometric Experts for Point Cloud Registration
SummaryThis paper presents SAGE, a method for aligning partially overlapping 3D point clouds that adds frozen pre-trained semantic features to geometry-based registration. A Semantic-Guided Mixture of Experts routes data so each expert handles distinct semantic-geometric patterns, reducing mismatches between similar-looking structures. SAGE reports the highest registration recall on 3DLoMatch and competitive results on 3DMatch.
Tomoyasu Shimada, Kazuhiko Murasaki, Shoichiro Takeda et al.
In briefUpper and lower bounds on optimal exponents for active hypothesis testing via PDEs
SummaryThis paper gives general upper and lower bounds on the optimal error exponent for active simple hypothesis testing (ASHT). It combines a recent partial differential equation characterization of the optimal exponent by Vijayan (2025) with the comparison principle for viscosity solutions of Hamilton-Jacobi PDEs. The resulting bounds are parametrized by convex functions, sets, and lower Hamiltonian relaxations, generalizing many known results.
In briefLog-Derivative Trick for Parametric Markov Chains: A Scalable Monte Carlo Gradient Estimator
SummaryThis paper addresses optimizing expected rewards and reachability probabilities in parametric Markov chains, where existing gradient methods rely on global symbolic computation that limits scalability. The authors introduce a Monte Carlo gradient estimator based on a measure-theoretic generalization of the log-derivative trick. It operates along sampled trajectories without materializing the state space or solving linear systems, and extends to continuous and general σ-finite state spaces. The authors prove unbiasedness under almost-sure termination and standard regularity conditions, and report that experiments on synthetic discrete and continuous-state cases scale beyond existing solvers while reducing execution time.
In briefRemembering Is Not Enough: Refining Memories and Learning When to Use Them
SummaryThis paper proposes DiscernMem, a training-free framework for LLM-based multi-agent systems that improves both the quality of stored memories and when they are used. Defect-Aware Memory Refinement diagnoses flawed rules by comparing team runs with and without a memory, and repairs them while checking earlier tasks. Boundary-Learning Meta-Memory records each memory's past uses and effects to learn when a role should apply or avoid it. Experiments with Qwen3-4B-Instruct-2507 show the highest average scores on AutoGen, CAMEL, and MacNet.
In briefTACO-Graph: Automatic Construction of Text-Attributed Graphs
SummaryThis paper introduces TACO-Graph, a decision tree framework that learns which graph construction choices work best for a given dataset and task, focusing on text-attributed graphs built from multi-attribute tabular data. It enumerates valid representations, scores each with fast proxy evaluation on small representative subsets, and trains an interpretable decision tree on those scores. Across five datasets, the trees predict construction quality on separate evaluation samples with mean Spearman correlations above 0.9 for node classification and above 0.8 for GraphRAG retrieval.
In briefWhen Is a Cheap Fairness Edit Safe? Screening Post-Hoc Edits in Two-Sided Matching Markets
SummaryThe paper studies a closed-form linear projector that reduces a protected attribute's influence at inference time in two-sided matching markets such as admissions, without retraining the encoder. It bounds the mean welfare change for vertically ordered markets and proposes a pre-deployment screen returning PASS, KILL, or OUT-OF-SCOPE. Experiments on synthetic and realistic markets report agreement with pre-specified verdicts and no observed false PASS.
In briefImage modeling from handcrafted priors to foundation models
SummaryThis paper reviews six decades of image modeling within a unified Bayesian framework, from handcrafted mathematical priors used for inverse problems to data-driven deep foundation image models for generation and inverse problems. It frames both as probabilistic characterizations of an image data manifold, centered on the gradient of the log-prior, the Stein score function. The authors present this as a single evolution linking signal processing and machine learning, from analytical models to models learned from large datasets.
In briefTap2Listen: Instantly switching target speakers in streaming audio-visual speech extraction
SummaryThis paper proposes Tap2Listen, a streaming audio-visual target speaker extraction model that can instantly switch to a new target speaker during a conversation. It separates computation by modality, so switching only requires a 0.12 ms re-projection of the visual cues. On the LRS3-2mix test set, it improves SI-SNRi from 8.85 to 9.68 dB at a comparable parameter count. The model runs in real time on a single CPU core.
In briefWhen Perturbation Consistency Adds Information Beyond Softmax Confidence
SummaryThis paper examines when the Decision Consistency Framework, which scores predictions by label stability and confidence deviation under small input perturbations, provides information beyond the clean softmax output. On clean data, the consistency score detects misclassifications worse than maximum softmax probability in every model family, at much higher compute. Consistency adds value under CIFAR-10-C noise corruptions and projected gradient attacks, where it reaches AUROC 0.83 against 0.37 for MSP.
In briefDeep Bayesian Methods for Interpretable Nonlinear Structural System Identification
SummaryThis tutorial systematically introduces deep Bayesian filtering for inferring mathematical models of nonlinear structural systems from measured data. It compares the Deep Markov Model, the Physics-guided Deep Markov Model (PgDMM), and the Neural Extended Kalman Filter (Neural EKF). The methods are evaluated on a three-dimensional shear-frame benchmark with hysteretic Bouc-Wen connections, covering dynamic state estimation and joint state and parameter estimation.
In briefA measurement study of how much AI-generated text now appears on public web pages
SummaryThe study analyzes over 5 million crawlable public webpages to estimate how many contain AI-generated text. English-language pages show higher prevalence than pages in other languages, and .gov and .edu domains show substantially lower prevalence than .com. The increase is fastest and the prevalence highest on pages least central in the web's link structure. The authors use multiple AI detectors and a Bayesian hierarchical model that corrects for detector false positives and false negatives.
Adam Visokay, Soumyakanti Pan, Molly Griston et al.
In briefTest-time adaptation with batch size one, guided by classifier geometry
SummaryThe paper proposes OriTTA, a method that updates a deployed model as unlabeled test samples arrive one at a time. It builds fixed geometric targets from the source model and aligns selected features to their predicted class axes, using a per-class memory and a reset step to guard against collapse. The authors report the highest mean accuracy at batch size 1 on CIFAR-10-C, CIFAR-100-C and ImageNet-C, and no collapse on ImageNet-C fog.
In briefEvent-based spacecraft pose estimation with polarity-time interaction and reliability weighting
SummaryThis paper presents a reliability-aware framework for estimating spacecraft pose from event-camera data, linking polarity-time localization, geometric recovery, event-CAD refinement, and trajectory optimization. A three-branch representation of positive events, negative events, and recent-event timing is fused through cross-branch attention and gating. On the official SPADES challenge test set, the complete system achieves a combined pose score of 0.0474, about 5.4% below PASAN+IRIS.
In briefMotion-guided asymmetric prediction for adapting video models to robot control
SummaryWorld Action Models use pretrained video generators to jointly predict future frames and robot actions. The paper proposes MoGAP, which adds noise asymmetrically so dynamic regions are reconstructed from less-corrupted static regions during training. The auxiliary path uses motion cues from training videos and is not needed at inference. On the GR-1 tabletop benchmark, success rates rise from 38.8% to 47.5% for Fast-WAM and from 52.3% to 62.3% for Cosmos3-Nano.
In briefGod Does Not Play Dice: Enforcing batch invariance in AI-agent GPU kernel generation
SummaryLLM serving systems batch requests dynamically, so GPU kernels lacking batch invariance can give different numerical outputs for the same input depending on batch context. The paper reports 19 confirmed batch invariance violations among 83 batch-applicable released kernel artifacts. It proposes a plug-in component combining contract-guided static dependency analysis with dynamic bitwise-consistency checks, used as acceptance gates that reject risky candidates before benchmarking. Across three agent frameworks, the batch invariance pass rate rose from 67.6% to 91.9% while retaining 91.8% of the baseline geometric mean speedup.
In briefDream-COP: Using Replayed Past Searches to Guide LLM-Driven Algorithm Discovery for Combinatorial Problems
SummaryThis paper presents Dream-COP, a framework that uses large language models to iteratively generate and evaluate programs for discovering algorithms for combinatorial optimization problems (COPs). Past search trajectories and their outcomes are treated as an experience world, where a policy agent learns executable policies through replay to guide a discovery agent. Across five COPs, the method finds better algorithms than fixed exploration in both heuristic component design and complete solver generation. It matches optimal values on KP and reduces the OBPP gap from 2.112% to 1.199% while using 42.5% fewer search requests.
In briefBalanced Subspace Compression for Mamba-2 Recurrent States
SummaryThis paper proposes a post-training method that compresses the fixed-size recurrent state of Mamba-2 models by balanced truncation over the subspace that writes and reads the output, rather than scoring individual coordinates. It also introduces HARP, which allocates ranks per layer from the Hankel spectrum and protects retrieval-sensitive layers. With HARP, halving the state dimension raises average WikiText-2 perplexity by 0.3 points (1.8%) across 130M to 2.7B models, compared with 1.6 points (9.3%) for GHOST.
WhyApplies balanced truncation to Mamba-2 state compression with layer-wise ranks from Hankel spectra, halving state dimension with 0.3 perplexity increase across 130M–2.7B.
In briefClustering data by splitting it into groups using a fractional math method
SummaryThis paper studies the normalized cut (NCut) problem used in data clustering. The authors propose a multidimensional quadratic transform, a fractional programming method that handles multiple ratio terms, which the classical Dinkelbach transform cannot. It reformulates NCut as a bipartite matching problem solvable by linear search. The paper also links the method to minorization-maximization theory to characterize convergence, and experiments show its advantages in data clustering and image segmentation tasks.
In brief4D Scene Reconstruction Fusing Event and RGB Cameras
SummaryThis preliminary study addresses reconstructing dynamic 3D scenes over time from a single moving camera. The authors add an event-camera input branch to Monst3R, an RGB-only 4D reconstruction model, leaving the rest of its architecture unchanged. Across 3 benchmark datasets, this simple fusion reduces Abs Rel by 3.4% on average, with the largest gains under high object motion.
Dulanga Weerakoon, Darshana Rathnayake, Van Minh Do et al.
In briefEstimating a Quadruped Robot's Pose from Camera Video Sequences
SummaryThis paper proposes QuadPose, a method that estimates the full-body 18-DoF pose of a quadruped robot (6-DoF base pose plus 12 joint angles) from monocular video sequences. The authors also introduce QuadSyn, described as the first camera-to-quadruped robot pose benchmark. On QuadSyn, QuadPose reports 92.18% PCK@2.5px, 95.54% PCK-AUC, 85.94% ADD@20cm, 66.98% ADD-AUC, and a 3.65 degree mean joint-angle error at about 33 FPS.
In briefLearning to Formalize: Connecting Open-Ended Tasks to Fast Decision Models
SummaryThe paper proposes a small candidate formalizer that builds validated alternatives from task context, which a frozen decision model then selects from, bridging fast decision models and open-ended large language models. On 200 CALVIN-D test episodes, DeepSeek V4 Flash pose candidates reached 46.5% success, compared with 11% for the official TypeSafe Jev-1.13.0 with generic motions. In a separate evaluation on 1,011 CALVIN-D validation annotations, a locally deployed decision model reached up to 52.6% task success with the fine-tuned formalizer versus 11.8% without it, at 2.09 s end-to-end latency compared with 3.67 s for Qwen3.5-27B.
In briefWithinVAE: A compact image autoencoder that learns meaning without borrowed recognition models
SummaryThis paper introduces WithinVAE, a compact variational visual autoencoder that learns semantic structure through self-distillation from an exponential-moving-average copy of itself, rather than from a pretrained recognition encoder. It uses a sequential bottleneck and position-dependent token masking to achieve 4-fold compression. On ImageNet-1K at 256×256, it improves all four reconstruction metrics over RPiAE, including a 1.36 dB PSNR gain, with half the latent capacity. Under frozen evaluation, its linear probing accuracy exceeds RPiAE by 6.5 points and the reconstruction-only baseline by 5.1 points, which the authors attribute to self-distillation.
In briefLabWorld: Learning Interactive 4D Worlds for Scientific Experiments
SummaryThis paper introduces LabWorld, an agent-driven Real2Sim framework that reconstructs interactive 4D laboratory environments from monocular RGB video. It recovers reusable 3D assets, a metric scene, time-varying object states, and procedure events. The method is evaluated on real pipetting and vortexing videos against general-purpose and dedicated baselines, showing higher scores on both tasks.
In briefInfraDev: Can Coding Agents Build AI Infrastructure, and How to Judge Them
SummaryThe paper introduces InfraDev, a benchmark of 36 tasks drawn from real pull requests to Megatron-LM, vLLM, and verl, covering training, inference, and reinforcement learning infrastructure. Grading combines hidden probes that check requirements the original tests miss with a prompt that specifies the required interfaces. Across six frontier models, the best resolved 38.9% of the tasks, and the probes rejected 32% of submissions that passed all official tests.
WhyBuilds a multi-file, multi-GPU infrastructure benchmark from real PRs with hidden probes, evaluated across six frontier models with failure analysis.
In briefFaster INT8 training via shared-threshold stochastic rounding with error correction
SummaryThe paper addresses stochastic rounding in low-precision training, where drawing independent random thresholds per element is slow on accelerators. It shows that sharing thresholds across elements causes a common rounding offset that harms training mainly through activation gradients. The proposed S3R-EC method corrects this mean rounding error, matching the validation loss of independent stochastic rounding in LLM pre-training. On TPU v5p, it speeds up full INT8 training of dense LLMs by 1.24x to 1.32x over BF16 and 1.11x over independent stochastic rounding.
In briefRecall or Relearn? A Training-Free Mixture of In-Context Experts for Recurring Concept Drift
SummaryThe paper studies how tabular foundation models applied to data streams, such as network intrusion detection, should handle a past concept that returns: restore its stored rows or relearn it from new labels. The authors propose MICE, a training-free method that segments the stream at detected changes, merges segments into stored contexts under a capacity limit, and weights contexts with a two-level mixture. On grid streams with random boundary phases, MICE outperforms a reset policy that empties the context at drifts, and on real streams it achieves the highest mean accuracy among the compared policies.
In briefBenchmark for LLM agents generating behaviorally adequate tests from software requirements
SummaryThis paper introduces Req2Test-Bench, a benchmark of 5,086 execution-validated instances drawn from developer-written BDD artifacts across 8 open-source projects, 5 programming languages, and 4 BDD frameworks. Agents must generate complete test implementations from Gherkin requirements and target interfaces, with behavioral adequacy measured as kill-completeness against ground-truth tests. The best configuration passes 82.8% of instances but only 33.8% are behaviorally adequate. The authors report diminishing returns from scaling agents and identify under-specified requirements as a key cause of the gap.
In briefMM-Net: A Multimodal Multi-Task Network for Sleep Staging and Breathing Event Detection After Stroke
SummaryThis paper presents MM-Net, a multimodal multi-task framework for sleep staging and respiratory-event detection in patients after ischemic stroke. It uses cross-fitting and ridge-based embedding alignment to address representation inconsistency in adapted EEG encoders, and fuses EEG, EOG and cardiorespiratory features. On the iSLEEPS dataset of 99 subacute stroke patients, ten-fold patient-independent evaluation achieved 76.3% staging accuracy and a 0.825 respiratory-event AUC.
Md Wahiduzzaman Suva, Esm E Moula Chowdhury Abha, Md Imtiaj Alam Sajin et al.
In briefStructured map text helps LLMs choose landmarks for route directions
SummaryThe paper studies how to represent map data so large language models can generate landmark-based route instructions. It proposes encoding spatial features as structured text in SVG format, and tests it on a new dataset built from human-written route descriptions and OpenStreetMap data, covering landmark selection at turns and around destinations. Across five LLMs, the structured text input performed as well as or better than PNG image inputs, and task-specific fine-tuning let smaller local models reach high selection accuracy.
Mifu Matsumoto, Yuki Nakanishi, Tomoya Hashiguchi et al.
In briefFollow the Herd: a data-driven method for giving actionable recourse without costly human-set action costs.
SummaryThe paper proposes Data-Driven Action Recourse (DDAR), a method for telling users how to act to achieve a desired outcome after a machine learning system has made a decision. It builds a dynamics model from historical user trajectories, assuming highly likely past trajectories have lower costs, and searches that model for trajectories reaching the desired outcome. In experiments, DDAR recommends plausible recourse without human supervision and achieves lower action costs than existing approaches.
In briefLess is more: learning what to profile for GPU kernel optimization
SummaryThis paper examines which hardware profiling metrics to collect when large language models optimize GPU kernels. The authors find that adding more metrics does not consistently improve optimization, and that the most useful information differs from kernel to kernel. They propose AutoProfiler, which learns a deterministic profiling tree from optimization feedback through iterative rewrite and split operations, and report better results than fixed profiling configurations under constrained budgets on KernelBench.
In briefADTGuard: A Depth-Time Safety Guard for Checking Harmful Content in Streaming LLM Output
SummaryThe paper addresses the risk that users see harmful content from large language models during streaming, before it has been checked. It proposes ADTGuard, which checks each token before release and adapts which internal layers it observes based on the current text. The method uses Depth-Time Safety Dynamics, a Causal Depth Router, and Risk-Onset Target Refinement to train with only response-level labels. The authors report that it matches or surpasses existing safeguards on multiple benchmarks with 16M trainable parameters.
In briefLearning from Points and Patches: A Pre-trained Transformer for Time Series Analysis
SummaryThis paper presents TimeHAT, a general-purpose pre-training framework for time-series analysis. It introduces a unified point-patch masking strategy that combines point-level and patch-level supervision for contextual reconstruction and causal extrapolation, along with a hybrid attention encoder intended to preserve contextual interactions while mitigating rank degeneration. The model is evaluated on real-world datasets across classification, long-term forecasting, imputation, and anomaly detection, under fine-tuning, few-shot, and zero-shot settings, with competitive results reported.
WhyCombines unified point-patch masking with hybrid attention for time-series pre-training, validated across four downstream tasks and multiple transfer settings.
In briefSeparating consistency and conflict across MLLM layers to detect fake news
SummaryThis paper addresses multimodal fake news, where image and text show sparse conflicts. It proposes MLCCD, which draws conflict signals from different layers of a multimodal large language model, selecting task-relevant visual and textual tokens and keeping those that change sharply between adjacent layers. These signals are split into consistency and conflict evidence and aggregated across layers with separate routers. The authors report that MLCCD outperforms existing methods on three benchmarks.
In briefHolomorphic Cauchy networks for efficient stationary tail computation in high-dimensional reflected Brownian motions
SummaryThe paper proposes Holomorphic Cauchy Networks (HCNs) for computing stationary tails of high-dimensional reflected Brownian motions, building on the NN4MGF framework. Native complex-valued rational Cauchy subnetworks enforce holomorphicity, removing sampled Cauchy-Riemann penalties. On the same NVIDIA H100 NVL, the 30D HCN configuration uses roughly tenfold less peak allocated memory and has 9.7-fold lower mean update time than NN4MGF. HCN reaches 50 dimensions on a laptop GPU with 6.367% mean relative error over thresholds 1 through 8.
In briefQESChunker: A Single Objective Unifying Overlapping and Non-Overlapping RAG Chunking
SummaryThis paper introduces QESChunker, a method that builds static chunk indices for retrieval-augmented generation using question-evidence pairs as supervision. It frames chunking as a supervised combinatorial optimization problem with one utility-maximization objective covering both non-overlapping and overlapping cases. Across five datasets, both variants achieve higher complete-evidence recall than existing chunking methods in nearly all evaluated settings.
In briefPRISM-BENCH: A Benchmark Checking Whether Vision-Language Models Reason From the Image, Beyond Final-Answer Accuracy
SummaryThe paper introduces PRISM-BENCH, a benchmark of 1,000 expert-curated questions across five under-represented reasoning categories, each paired with a human-written and human-validated reasoning trace. Two diagnostics, a perception probe and trace alignment, show that roughly 85% of frontier-model failures come from misperception and that trace alignment predicts answer correctness. The authors also propose CORE, an efficiency-aware metric, and report that the best of 25 systems reaches 66.0% accuracy.
Vansh Bhatia, Mehdi Rezagholizadeh, Parsa Ashrafi Fashi et al.
In briefNextEval: An Agent-Based Solver and Benchmark for Derivative-Free Optimization
SummaryThis paper presents NextEval, an agent-based solver for derivative-free optimization, where gradients are unavailable and each objective evaluation is costly. A large language model agent works with a numerical environment to choose evaluation points, configure tools, and allocate the limited budget. The authors also introduce NextEval Bench, a benchmark protocol for comparing agent configurations under matched budgets and information constraints. Numerical results on low-dimensional problems show NextEval is more function-evaluation efficient than state-of-the-art DFO baselines.
In briefREACH: A reinforcement-learning framework letting a humanoid robot walk and crawl through confined spaces
SummaryThis paper presents REACH, a hierarchical reinforcement learning framework that enables a Unitree G1 humanoid to traverse height-constrained environments using variable-height bipedal walking and crawling. The learned policies regulate body height over 1.27–0.50 m for bipedal walking and 0.50–0.20 m for crawling. A LiDAR-based planner converts point clouds into collision-free paths and locomotion commands, and real-world experiments demonstrate accurate height and velocity tracking, reliable mode transitions, and confined-space traversal.
In briefFrameForge: Using Video Model Priors for Unified Image Generation
SummaryThis paper presents FrameForge, a method for unified conditional image generation that produces a target image from a visual input and text guidance. It builds on pretrained video models and models the condition-target relationship using input-adaptive virtual temporal relations, via Mixture of Temporal Experts (MoTE) and Chain of Temporal Reasoning (CoTR). Experiments on DreamBench, COCO2017, and GEdit-Bench report superior results over compared methods in subject-driven generation, canny-to-image generation, and image editing.
In briefMonitor accuracy is not enough: a decision-value framework for AI agent restart monitors
SummaryThis paper argues that classification accuracy and true- and false-positive rates alone cannot show whether restarting an AI agent improves its success rate. It proposes two decision-focused methods: a decision-sufficiency criterion that checks whether progress representations keep the information needed to choose between restarting and continuing, and a model-free statistical audit over existing agent traces that can recommend restart, continue, or abstain. In experiments on SymPy issues from SWE-bench Verified, an apparent restart gain did not replicate, so the authors recommend judging monitors by recovery opportunity, decision value, and statistical support.
In briefLow-cost handwritten text recognition for Sub-Saharan civil-status registers, from smartphone photos to records using CPU-only inference
SummaryThis paper studies whether consumer smartphone capture and CPU-only processing can make handwritten civil-status registers from Sub-Saharan Africa machine-readable. The authors built a corpus of 1,568 smartphone images of birth, marriage, and death records from a Chadian municipal registry and fine-tuned a Kraken recognizer on a free-tier GPU. Fine-tuning reduced character error rate from 14.53% to 0.50% on a single-registry test set, and the recognizer runs on a six-core laptop CPU. The results show feasibility within one registry only, not performance on unseen writers or sites.
Ousmane Hamadou, Jean Louis Kedieng Ebongue Fendji
In briefA language-aware AI agent that picks training data for multilingual speech recognition based on its errors
SummaryThis paper proposes a data sampling framework for multilingual automatic speech recognition that places an LLM agent inside the training loop. The agent analyzes recognition errors from phonological, morphological, and syntactic perspectives and uses this feedback, with training statistics and budget constraints, to choose sampling strategies at domain, sample, and token levels. Experiments on twelve Common Voice languages with Whisper-medium show better performance than full-data training using less data and compute, and faster convergence than conventional sampling methods.
In briefATEMPO: A Diffusion-Based Transfer Entropy Estimator That Finds When Information Arrives Between Time Series
SummaryThe paper introduces ATEMPO, a diffusion-based transfer entropy estimator amortized over lags and time, so a single score network estimates the full transfer entropy surface without refitting. Differencing this surface recovers the time at which information from one series arrives in another. It reports high lag-detection accuracy on synthetic and time-varying-lag datasets, and improved forecasting and brain-computer interface results using the selected lags.
Emma Andrews, Konstantinos Kontras, Keane Ong et al.
In briefRoboAuto: autonomous search for robot programs using layered physical understanding
SummaryThis paper presents RoboAuto, a framework in which LLM-based agents autonomously search for and refine executable robot manipulation programs. It organizes physical understanding at the object, scene, and interaction levels, using a hierarchical knowledge base, scene grounding, and multi-stage planning with self-designed experiments. The authors report higher success rates than ASPIRE and π0.5 on several benchmarks and three real-robot long-horizon tasks.
In briefETC-Bench: A Multilingual Benchmark for Standardizing Transcript Variants in Evaluation
SummaryThe paper addresses the problem that written transcripts of the same speech can differ in ways the audio does not determine, and proposes evaluation-side transcript canonicalization (ETC) to map such variants to one canonical form. It introduces ETC-Bench, covering 181 phenomenon types across 12 languages, built from minimal pairs. The authors also propose the SENA metric for canonicalization success and semantic preservation, and the TACS agent skill, which outperforms baseline systems on shared languages.
In briefLocating Causal Attention Heads for Evidence Effects in Transformers
SummaryThe paper introduces controlled evidence aggregation, which varies symbol counts and order in prompts to measure a frozen Transformer's preference between two answer tokens. A selection method called BUC identifies compact sets of attention heads whose ablation removes 65.5-91.6% of the count effect in Pythia-410M. Eight heads chosen only on symbolic prompts also lower mean accuracy by 8.2 points on a natural language evidence task, compared with 0.8 points for random matched sets.
In briefBenchmark for Task Takeover in Agent Handoffs (paper)
SummaryThis paper introduces HSC-Bench, a benchmark of 1,039 tasks for evaluating how receiving models take over work when responsibility passes between agents. It tests two aspects: recovering relevant repository evidence and continuing from established conversational, tool, and file-work state. Across six receiving models, structured handoff state improved evidence recovery for every model and continuation performance for most. The strongest receivers differed between the two tracks, showing that evidence-recovery rankings alone do not capture continuation from inherited state.
In briefFolding attention bias into keys so compacted KV caches run on standard kernels
SummaryThis paper addresses a limitation of Attention Matching, a method that compacts a KV cache by keeping some keys per head, refitting values, and adding a bias to each key's logit. Standard kernels such as FlashAttention and vLLM's paged caches cannot accept this bias. The authors fold the bias into the keys along one closed-form direction per head, producing a plain cache of the same size and head dimension. Experiments on Qwen3 and Llama models report accuracy within noise of Attention Matching in most settings, with the folded cache running on unmodified vLLM.
In briefPRISM: A benchmark testing how reliable pathology AI models are when labels are scarce.
SummaryPRISM is a tile-level benchmark that evaluates eight pathology foundation models on six datasets, measuring calibration, behavior at low label fractions, and cross-hospital transfer. It runs 864 linear probe experiments across label fractions and seeds, plus a five-hospital Camelyon17 study of 20 transfer pairs. Findings show calibration rankings track AUROC under full supervision but become dataset-specific at 1% labels, and that adding source labels can worsen target calibration when transfer is weak. The benchmark is released as a pip-installable package with reference results for all 288 cells.
WhyJointly evaluates calibration, low-label regimes, and cross-hospital transfer, revealing several new phenomena in pathology foundation models.
In briefMedConstBench: Testing How Ethical Constitutions Shape LLM Behavior in Medicine
SummaryThis paper introduces MedConstBench, a medical-ethics benchmark that treats a model's deployment constitution, its written ethical policy, as a controlled variable. It is organized around Beauchamp's four principles and scored by human-calibrated multi-model peer review, and includes MedDilemma, a dataset of clinical dilemmas. Evaluating ten open-source LLMs under four constitutions, the authors find the effect is dimension-selective and model-dependent, with no universally optimal constitution. They propose a do-no-harm adaptive selector that chooses a constitution per model.
SummaryThis paper examines how AI scientific discovery pipelines can determine whether a generated research idea is meaningfully novel. The authors build an evaluation framework with temporal ground truth, using four datasets of 1,060 proposals drawn from published papers, reviewer disputes, and ideas later realized by researchers. They introduce ClaimNov, a fine-grained, interpretable claim-level novelty module, which catches 97% of disguised papers at a 1.5% false-alarm rate.
WhyProposes claim-level novelty checking, auditing existing methods on 1,060 proposals with temporal ground truth, showing substantive improvements.
In briefIdentity or Evidence? Auditing Whether Life-Science AI Benchmarks Test Evidence Reasoning
SummaryThis paper examines whether agentic life-science benchmarks truly measure reasoning over evidence rather than recognition of a named entity. The authors vary identity and evidence independently on a clinical-outcome benchmark and on Biomni and LAB-Bench tasks. When prompts consist mainly of a name and candidate list, blinding identity sharply reduces accuracy, so the authors recommend reporting identity-blinded results, retrieval telemetry, and positive controls.
In briefARISE: A Continual Learning Method That Keeps Multimodal Forgery Detectors Stable as New Attacks Emerge
SummaryThis paper studies multimodal forgery detection when new attacks keep appearing, so the detector must learn each new attack without forgetting earlier ones. The authors introduce Fable-Bench, a benchmark that evaluates attack evolution through joint-modality, text-side, and image-side protocols. They propose ARISE, which uses a real-only Gaussian mixture model of authentic samples as a historical reference together with stability-preserving and vulnerability-aware replay. Experiments across three attack-evolution protocols and several detector architectures show better historical retention and average final performance than representative continual-learning baselines.
In briefGarmentSegAgent: Self-Correcting Text Guidance for Automatic Clothing Segmentation in Web Fashion Images
SummaryThis paper presents GarmentSegAgent, an autonomous framework that segments garments in web fashion images through iterative textual self-refinement. Its core method, CAST, links visual error attribution to targeted text optimization, updating the instruction components responsible for errors and checking them against garment constraints. The underlying foundation models stay frozen during inference. Experiments on DeepFashion2, ModaNet, and Fashionpedia report consistent gains in region overlap and boundary accuracy.
In briefFast algorithms for checking graph separation rules
SummaryGraphical models encode dependencies among random variables through graph separation criteria. The paper studies how to check separation statements for various graph classes and criteria. It shows that all criteria studied can be checked using modular variations of reachability. For all criteria except σ-separation this gives a linear-time algorithm, and R and Python code is provided and evaluated in simulations.
Mois\'es Chavira-Flores, Sebastian Weichwald, Leonard Henckel
In briefProCerDCL: A Query-Efficient Black-Box Method for Checking Deep Learning Compiler Optimizations
SummaryThe paper presents ProCerDCL, a framework that estimates how often a binary property is violated under a sampleable input distribution, using only black-box queries. It returns acceptance, rejection, or an inconclusive verdict by applying a single sequential test instead of a baseline interval family. Evaluated on 24 tasks across TorchInductor and TensorRT, it reports 3.6 to 4.2 times fewer samples than the baseline, plus case studies of two compiler faults.
In briefAC-JEPA: An anchored joint-embedding method for translating images into chemical symbols
SummaryThe paper proposes AC-JEPA, a plug-in compatible with supervised fine-tuning that predicts masked intermediate states across both cross-modal directions using paired data. Tested on optical chemical structure recognition, AC-JEPA trained on clean pairs reached 76.07% accuracy on the WildMol Simple set, versus 67.40% for SFT alone. With the same pseudo-labeled data, AC-JEPA self-training reached 80.23%, while direct SFT fell from 67.40% to 63.47%.
In briefSecond-order Sinkhorn Geometry and √n-Inference: A paper building a Riemannian geometry from the Sinkhorn divergence and proving √n limit theorems and bootstrap consistency for its geometric operations.
SummaryThe Sinkhorn divergence is not a metric and lacks a formal Riemannian structure, which limits optimal transport methods that rely on interpolation or averaging. The authors extend an existing framework on probability measures, proving Hadamard differentiability of the Sinkhorn metric tensor for a class of finite-smoothness costs and deriving the Levi-Civita covariant derivative. This yields geodesic and parallel transport equations as ODEs on a fixed RKHS. They show that plugin estimators of these geometric objects have √n-limiting Gaussian processes and that the nonparametric bootstrap is consistent.
Tristan Luca Saidi, Florian Gunsilius, Gonzalo Mena
In briefLearning Spatial and Frequency Features to Enhance STEM Images
SummaryThis paper proposes a spatial-frequency learning framework for enhancing scanning transmission electron microscopy (STEM) images. It exploits periodic atomic arrangements in the frequency domain and restores local atomic details in the spatial domain, then fuses both feature sets. The authors also introduce a noise calibration method that estimates scan, pointwise, and background noise parameters from real STEM images to build a more realistic synthetic dataset. Experiments report state-of-the-art enhancement on synthetic data and real images of advanced materials, along with improved atomic localization accuracy.
In briefCogArena: Testing AI Agents on Interactive Behavioral Experiments
SummaryThis paper introduces CogArena, a benchmark of ten experimental paradigms from the social and behavioral sciences. AI agents complete the experiments through a standard web browser, using the same visual stimuli and response interfaces as online human participants. A three-level scoring procedure assesses task completion, accuracy, and whether the agent's data show the behavioral signatures these paradigms reliably produce in people. Across four multimodal models run ten times per task, the authors find that finishing an experiment and reproducing its signatures are distinct, with many sessions unfinished and reproduction varying widely by signature.
Kiant\'e Fernandez, Caitlin Chen, Jialu Zhou et al.
In briefBetter rollout sampling for on-policy distillation: Mixture Proposal On-Policy Distillation (MPOPD)
SummaryThis paper examines how rollout sampling affects on-policy distillation (OPD) for language models. The authors argue that sampling directly from the current student can create a compounding teacher-student mismatch, where teacher supervision becomes noisy or misleading and patterns such as repetition persist. They propose Mixture Proposal On-Policy Distillation (MPOPD), which uses a geometric mixture of student and teacher policies as the rollout target, approximated by a proposal policy that generates rollouts in a single forward pass. Across 16 math, knowledge, and medical reasoning benchmarks, MPOPD outperforms OPD and recent variants, improving math reasoning average accuracy by 6.91 on Qwen3-4B-Base while reducing average generation length by 41%.
In briefWhen to Intervene? Empirical Study of LLM Agent Errors and Self-Recovery
SummaryThe authors built ReInBench, a dataset of 1,259 agent traces and 1,403 error events from computer-use and Web tasks, to study how often LLM agents recover from errors on their own. They found that 725 error events self-recovered, often within a few steps, with recovery varying across execution LLMs. They also propose ReCoF, a method that predicts whether an error will self-recover to guide whether intervention is needed, reporting a 9.02% average relative Macro-F1 improvement over the strongest baseline.
In briefUnderdamped Anchored Langevin Dynamics: A Momentum-Based Sampler for Non-Differentiable, Heavy-Tailed Targets
SummaryThe paper proposes underdamped anchored Langevin dynamics (UALD) and a discretized algorithm, UALMC, for sampling from non-differentiable and heavy-tailed target distributions. The method replaces the original potential with a smooth reference potential and adds multiplicative scaling to the drift term. The authors give non-asymptotic guarantees in 2-Wasserstein distance and report numerical experiments supporting the theory and practical performance.
In briefCoHARL: Co-Evolving Agent Policies and Executable Harnesses via Reinforcement Learning
SummaryThis paper presents CoHARL, a reinforcement learning framework that updates a shared policy and a population of executable harnesses (code governing tool use and feedback) together during training. The policy learns from rollouts across the harness population, while a proposer revises harness code using execution traces. The authors report a 27.6% mean resolved rate on SWE-rebench, 6.4 percentage points above harness-only search, with gains on SWE-bench Verified and Terminal-Bench 2.0.
In briefConfident First, Reflective Second: Reinforcement learning that keeps video models' first answers reliable while reflecting
SummaryThe paper finds that reinforcement learning for video models that answer, re-examine, and revise can erode the reliability of the initial answer, a effect it calls confidence erosion. It proposes CoFire, which targets three erosion channels: reward incentives, parameter drift, and gradient interference. CoFire reports 67.0% mean accuracy on Video-MME, MLVU, and LongVideoBench, and improves both initial and final answer accuracy over outcome-reward RL.
In briefNemotron-Labs-Diffusion-Uni: A Unified Diffusion Language Model for Vision Understanding and Generation
SummaryThis paper presents Nemotron-Labs-Diffusion-Uni (NLD-Uni), a unified discrete diffusion large language model for visual understanding and generation. It introduces a cross-paradigm on-policy distillation framework that transfers the understanding ability of autoregressive vision-language models and the generation ability of continuous flow models into one diffusion model. It also uses a progressive self-refining diffusion process that gradually corrupts tokens over multiple training steps, allowing intermediate token commitments at inference. The authors report state-of-the-art results among unified diffusion models on tasks including visual question answering, multimodal reasoning, object grounding, text-to-image generation, and image editing.
In briefClusterSplat: Selecting Semantic Clusters to Locate Objects in 3D Gaussian Splatting Scenes
SummaryThis paper studies 3D visual grounding in 3D Gaussian Splatting, where a text description identifies a target object as both a rendered 2D mask and explicit 3D Gaussians. ClusterSplat reframes target selection as choosing among scene-adaptive cluster candidates, built from instance-aware Gaussian features and merged using a description-length criterion. A query-conditioned lightweight MLP scores these clusters, and the chosen cluster defines the 3D target. On ScanRefer, the method achieves the best rendered 2D segmentation and explicit 3D target metrics among the compared 3DGS baselines.
Jaesung Lee, Hyeontaek Hwang, Gabriel Manalu et al.
In briefFull Collapse Is Suboptimal: Controlled Partial Neural Collapse for Long-Tailed Recognition
SummaryThis paper examines neural collapse methods for long-tailed recognition, which align all class features to a fixed Equiangular Tight Frame (ETF) and can over-contract tail-class samples. The authors propose controlled partial neural collapse using two coordinated components on a backbone: a Class-Adaptive Sparse Activation (CASA) classifier and an ETF-constrained adaptive margin (ETF-AM) loss. Head classes collapse along the ETF while tail classes keep within-class diversity. The method achieves the best overall accuracy on CIFAR100-LT, ImageNet-LT, Places-LT, and iNaturalist-2018.
In briefLociArt: A Unified Framework for Local Video Style Transfer with Free Style Choice
SummaryThis paper presents LociArt, a unified framework for stylizing only chosen regions of a video while leaving the rest unchanged, specifying both where and which style to apply. Its core model, GRNStyle, builds on Generative Refinement Networks and uses a polar style-aware mask plus conditioning on mask features and raw-mask spatial information. The authors also release the LociArt-100K dataset of 100K local stylization pairs and a benchmark with 80 source videos and 168 stylization cases. Experiments report better regional style quality and spatial control than existing methods while keeping strong temporal consistency.
In briefCorrection or Manipulation? Measuring How LLM Agents Resist User Pressure
SummaryThis paper studies whether LLM agents can tell a genuine user correction from manipulation when users push back on their answers. Seven models from four vendors were tested in two studies, answering questions and handling customer-support account changes, under simulated user pressure. Citing a source made most models more willing to concede and worse at discriminating, and the authors argue that concession rates alone are misleading.
WhyProposes separating discrimination from concession propensity via signal detection theory, tested across seven models in answer and agentic action studies.
In briefHumanoid whole-body control using stochastic motion tokenization
SummaryThis paper proposes concise motion representations that a high-level controller can generate from behavior labels or text commands, then feed into a whole-body tracking policy for humanoid robots. The authors use a stochastic motion tokenizer that, at matched capacity, reconstructs Unitree G1 motions better than FSQ and VQ-VAE. The discrete interface enables a physical Unitree G1 to execute generated dance, fight, and walk motions lasting 16.8 s each, and a 35 s fight-to-dance sequence.
In brief"A Heart Means Recursion": How Vision-Language Models Interpret and Communicate Visual Meaning
SummaryThis paper examines whether vision-language models can communicate visually as humans do. It finds that simple symbols such as a red heart can steer models toward recursive solutions on unrelated coding problems, often implicitly. When models communicate visually with each other, the shapes they send often fail to guide other models, revealing a generation–interpretation gap. In games involving abstract concepts and stories, models creatively convey meaning through visuals, though some of these messages are not intuitive to humans.
Alexa R. Tartaglini, Zhengxuan Wu, Thadd\"aus Wiedemer et al.
In briefFrom Solver Trajectories to Teaching Trajectories: Reasoning Distillation Aligned with Student Cognition
SummaryThis paper examines how to distill reasoning from large reasoning models into smaller student models more effectively. The authors propose CARD, a reinforcement learning framework that converts existing solver trajectories into teaching trajectories suited to the student's cognitive capacity, without costly resampling. It is optimized with a Distribution Alignment Reward and a Logical Coherence Reward. The reported results show that CARD outperforms distillation from raw teacher trajectories in-domain and achieves substantial gains out-of-distribution.
In briefRouteDRAG: Learning when and how an AI agent should answer questions over multimodal documents
SummaryThis paper studies question answering over long documents containing text, tables, and figures, where agentic retrieval-augmented generation is used. The authors find that running the agent helps some queries but not others, and can sometimes reduce accuracy. They propose RouteDRAG, a learning-based framework with four components that decide when to invoke the agent and how to control its execution. On two benchmarks, it improves average accuracy by 3.47 percentage points over shorter-path-only execution while assigning only 79.3% of queries to the agent.
In briefA high-fidelity dataset linking aerodynamics and structures for transonic wings in cruise
SummaryThis paper introduces a high-fidelity dataset for static aerostructural analysis of transonic wings, containing 2471 samples plus three out-of-distribution samples. Wing planform and airfoil shapes are parameterized, and analysis conditions include Mach number, angle of attack, altitude, and structural load factor. The authors train machine learning models to predict aerodynamic surface quantities, integrated aerodynamic performance, and wingbox failure distribution, achieving at most 0.0309 and 0.1189 normalized RMSE respectively.
In briefLearning atomic interactions to model the thermodynamics of liquid mixtures
SummaryThis paper introduces COGNAC, a graph neural network that learns atom-level geometry and interaction parameters depending on molecular neighborhoods and evaluates them within the UNIFAC framework for activity coefficients. Trained on about 820,000 experimental data points, it is more accurate than UNIFAC and HANNA for binary and multi-component mixtures, reducing error relative to HANNA by 13-63%. Its evaluation cost is comparable to UNIFAC, and it is strictly thermodynamically consistent.
Marco Hoffmann, Sebastian Schmitt, Hans Hasse et al.
In briefCONCUR: Consistent Auxiliary Rewards for Training Language Models on Multiple Tasks
SummaryThis paper studies reinforcement learning for large language models trained on several tasks, where auxiliary rewards such as length rewards can conflict with a task's primary reward. The authors propose CONCUR, an advantage estimator that adjusts the scores of gated auxiliary rewards so they never reverse the primary reward's ordering and tie at the primary's ceiling. Two variants are described, and on three models across math, code, instruction following, and writing, both score higher on average than all baselines and than dropping the gated auxiliaries.
In briefHow user presence changes agent behavior in negotiation
SummaryThis paper studies how a user's presence affects LLM agents that negotiate on the user's behalf with other agents. The authors introduce PRESENT, a controlled framework that compares negotiations with and without an interactive user while keeping the game, objectives, and preferences fixed. Across 11.2k runs covering 7 games and 8 models, user presence raised payouts by 13.7% on average and made agents negotiate more aggressively, and it also produced misaligned behavior toward counterparties and users.
