In briefSchemaFill: Faster LLM Tool Calling via Parallel Speculative Decoding
SummarySchemaFill is a framework that speeds up tool calling in large language models. It generates candidate values for multiple argument slots in parallel, then has the target model verify them under the actual output prefix. Only verified tokens are kept, and the paper reports up to a 4.05 times end-to-end throughput gain over autoregressive decoding on Glaive and BFCL.
In briefStructured representation learning for safely controlling a nuclear reactor by imitation
SummaryThis paper studies whether a learned controller can safely imitate an expert model predictive control policy for load-following a pressurized water reactor. The authors encode variables from each timescale into separate latent spaces and train a controller on the resulting representation. On long rollouts, this improves accuracy and feasibility over a shared-embedding baseline, but standalone use still leaves several percent of trajectories infeasible. Using the method to warmstart the NMPC optimizer restores full feasibility and cuts computation time by about 15%.
Perceval Beja-Battais (CB), Alain Grosset{\^e}te, Nicolas Vayatis (CB)
In briefMoCAR: Coordinate-aware autoregressive model for continuous trajectory forecasting
SummaryThis paper presents MoCAR, a decoder-only model that treats trajectory forecasting as next-code prediction in a continuous latent space, where each motion code captures local trajectory geometry and the reference-frame change. It uses a single-stage architecture without trajectory queries, goal candidates, or proposal-and-refinement pipelines. On Argoverse benchmarks, MoCAR reports top-tier performance, transfers zero-shot from AV2 to AV1, and improves on turn-heavy scenarios.
In briefAI-Decision Checkpoints for Business Process Management: A Framework for Teaching AI-Augmented Processes
SummaryThis paper proposes "AI-decision checkpoints," explicit points in process development where developers identify AI-candidate sub-processes, assess their effects on time, cost, quality, and flexibility, consider legal and organisational constraints, and document whether to adopt, constrain, or reject AI components. The checkpoints are demonstrated through a fictitious customer onboarding case within a six-module lifecycle framework covering modeling, simulation, AI-agent workflow execution, and process mining. A preliminary formative reflection suggests the approach helped distinguish task-level automation from process-level value.
In briefCan Small Language Models Learn to Negotiate? A Study of RL-Trained Sellers
SummaryThe authors train four Gemma 4 checkpoints, from about 2.3B to 31B effective parameters, with GRPO reinforcement learning to act as sellers in bilateral multi-issue bargaining. Each model is evaluated on 1,152 negotiations against two frontier buyer models not seen during training. The RL gain over the base model grows with size at a shared learning rate, and tripling the learning rate improves results at every size.
In briefDistributional Alignment of LLMs and Its Generalization Properties
SummaryThis paper studies how to align the output distribution of large language models with a human reference distribution, which matters for simulating human behavior such as surveys. Using two large-scale benchmarks, the authors find that models can achieve this alignment and generalize to unseen questions in the same domain, while cross-topic generalization depends on how related the topics are. However, aligned models can degrade unexpectedly, such as losing creative ability, suggesting their generalization is unpredictable and unstable.
In briefEORestore-Agent: A fidelity-guided agent that restores remote sensing images with several stacked degradations
SummaryRemote sensing images often suffer several degradations at once, such as haze, cloud, noise, blur, low light, and low resolution. EORestore-Agent uses a fine-tuned vision-language model to identify residual degradations and choose restoration tools, while a learned relative quality scorer predicts PSNR, SSIM, and LPIPS changes for each candidate step. A step is accepted only when no predicted metric worsens and the predicted PSNR gain is positive. On a synthetic Landsat-8 benchmark, it improves PSNR by 2.3 to 3.2 dB over the strongest retrained all-in-one baseline, and examples on Sentinel-2 show transfer to real atmospheric degradation without retraining.
In briefLoss-invariant random projections as fixed observers for tracking how neural network representations change during training
SummaryThe paper introduces passive probes, which are fixed, untrained, property-independent projections applied to neural network representations during training. Motivated by prediction on the sphere S², the authors show that equivalent parameterizations reveal a loss-invariant trace coordinate, leading to a general construction of such observers. Experiments indicate that accessibility of eventual task difficulty rises during training for surface-normal estimation and image inpainting, but stays near its initial level for image classification.
In briefCorrect Code, Broken Contributions? Benchmarking Coding Agents on Repository Policy Compliance
SummaryThis paper introduces SWE-CC, a benchmark that tests whether autonomous coding agents follow open-source repositories' contribution policies, not just pass unit tests. The authors converted developer documentation from 12 repositories into 823 machine-checkable policies and evaluated agents on 500 end-to-end tasks extended from SWE-bench Verified. Even when patches were functionally correct, agents violated 43.1 percent of applicable policies, with nearly half of violations occurring during intermediate execution steps.
In briefCopies or Sources? How LLM Aggregators Count Restated Evidence in Multi-Agent Systems
SummaryThe paper examines whether LLM aggregators in multi-agent systems count independent sources or every restated statement when pooling repeated observations. The authors convert reported probabilities into units of independent readings and test four models across three settings. They find forwarded copies receive only 0.06 to 0.42 of a new reading' weight, often leading to early commitment. Declaring what a copy contributes, or having agents refer to readings rather than restate them, substantially reduces this effect.
WhyProposes a copy-weight framework measuring restated evidence in independent readings, tested systematically across three testbeds and several models, with clear relevance to multi-agent aggregation.
In briefJudged Useless, Queried Anyway: Tool Agents Seldom Act on Their Own Evidence Verdicts
SummaryThis paper studies whether tool-using AI agents stop relying on a source after judging its results useless, using a retrieval setup with controlled source failures. Seven tested agents labeled failing-source results as useless 97-100% of the time, yet most rarely stopped on that judgment. The authors show that requiring an answer after five consecutive results judged useless makes stopping follow the evidence, and a pre-registered replication on 300 new questions confirmed the effect.
In briefReflective AI Agent for Designing Multi-Objective Peptides
SummaryLarge language models can reason about design strategies from scientific literature but struggle to apply them reliably to biological sequences. The paper presents EASER, a reflective agent that links reasoning to sequence generation through an offline-trained property interface that steers a diffusion generator with evidence-based intervention hypotheses. On multi-objective antimicrobial peptide design, it achieved the highest mean hypervolume and lowest mean IGD+ over six trials compared with baselines, and ablations confirmed the contribution of each component.
In briefFeedback Dominance Analysis for Pursuit-Evasion Games on Graphs
SummaryThis paper identifies dominance regions in discrete, simultaneous-move pursuit-evasion games played on graphs. It uses a set-based dynamic programming approach to give necessary and sufficient conditions for the pursuer to win under worst-case behavior, and converts these dominance sets into feedback strategies that adapt to player positions. For states where neither side can guarantee victory, it derives upper and lower bounds on the pursuer's winning probability, and simulations confirm the results.
In briefTwo-Point Local Optimality for k-Means via Boundary-Point Screening
SummaryThis paper studies k-means clustering and introduces r-point local optimality, in which no reassignment of at most r samples lowers the objective, with a focus on r=2. The authors show that improving two-point moves involve only boundary points and use this to propose BPS-2PLS, which reaches a two-point local optimum. Across twelve benchmarks, it attains the lowest available mean WCSS on ten.
In briefAnlu: In-context time series anomaly detection via counterfactual supervision
SummaryThis paper studies time series anomaly detection where a reference record supplies evidence about normal behavior while model parameters stay fixed. The authors introduce counterfactual supervision, pairing each query with two references that imply different normal rules and labeling it under each. Anlu adds a reference memory and zero-initialized gated adapters to a frozen time-series foundation model. On 350 TSB-AD-U sequences, mean VUS-PR rises from 0.542 to 0.607, and falls to 0.499 when the reference is replaced with zeros.
In briefAuditable patient timeline reconstruction using provenance-aware evidence graphs
SummaryThis paper studies whether a clinical timeline reconstruction system can be auditable by keeping the mentions behind each answer, recording how facts were revised, and abstaining when the text lacks evidence. The tests use a fully synthetic corpus of 1,000 patients, 3,353 notes, and 220 revision edges. The results describe implementation behaviour on synthetic data and do not indicate clinical performance.
In briefLearning to Correct Expert Answers When Deferring Queries Online
SummaryThis paper studies online learning to defer, where a learner picks an expert and a correction function before buying an answer and then applies that correction. The authors propose ORUCB, which pools shared and expert-specific polynomial response models and calibrates risk regression and exploration using a bound on response-learning error. Under stated assumptions it achieves high-probability pseudo-regret O(√T log(T+1)). On four test streams, the selected cubic policy has lower fee-inclusive cost than seven baselines that use answers unchanged.
Yannis Montreuil, Axel Carlier, Lai Xing Ng et al.
In briefAnosognosia in LLMs: Can Models Detect Their Own Quantization Degradation?
SummaryInspired by anosognosia, a condition where patients fail to recognize their own impairments, this paper tests whether LLMs can detect degradation in their computational substrate caused by quantization. Existing models fail to self-report their quantization state, even when given their own generated text as a cue. Linear probing finds that generated text carries almost no trace of quantization, while internal representations contain clear, method-specific fingerprints. The authors suggest internal representations may be a more promising route to LLM self-awareness, while noting fundamental limits to generalizing self-monitoring.
In briefFrom Target Selection to Condition Reuse: Checking Whether Edits to Tool-Use Models Carry Over
SummaryThe paper studies whether activation edits that make Qwen3.5-27B give a target answer on tool-use tasks also let the model apply the intended condition elsewhere. Target-selection accuracy rose sharply, but the gains vanished when familiar facts and identifiers were recombined. Internal edits can change how a source fact affects judgments, yet this does not guarantee correct complete actions.
In briefMS-Exam-Gen: An Automatic Benchmark Toolkit for Testing LLMs on MS MRI Knowledge
SummaryThis paper presents MS-Exam-Gen, a framework that automatically builds and audits a text-based multiple-choice benchmark on multiple sclerosis MRI knowledge, without assessing direct image interpretation. From a 66-source corpus, it produced a 3,058-item candidate benchmark across 16 topics and 53 subtopics. Twelve primary LLM endpoints scored between 46.9% and 89.7% accuracy. The authors note that expert adjudication and full option-order counterbalancing remain future work, so it is not a clinically certified exam.
Abdul Basit, Muhammad Abdullah Hanif, Muhammad Shafique
In briefBayesian optimization for finding neural network designs without training them
SummaryThis paper applies Bayesian Optimization to zero-shot neural architecture search, which otherwise often relies on near-random evolutionary mutation. It learns a latent space of architectures with a Variational Autoencoder and combines several zero-shot proxies into one search objective. The authors report state-of-the-art results on image classification, object detection, and semantic segmentation under given parameter count constraints.
In briefAdding Code-Switched Contexts to Bilingual Model Training Inspired by Cognition
SummaryThis paper draws on how bilingual children learn language from code-switched input and adds synthetic code-switched data during pretraining of bilingual models. The authors test two typologically distinct language pairs while controlling where code-switches occur and how often they switch across training stages. For typologically close languages, training with code-switched data improves cross-lingual alignment.
In briefTIGER: Time-series classification via an in-context-learning gated ensemble of representations
SummaryThe paper proposes TIGER, which applies three general-purpose classifiers to four feature representations of time series, stacking the twelve resulting predictions into a meta-feature matrix. A final prediction is chosen per dataset, adaptively, between a weighted majority vote and the pretrained tabular model TabICLv2 used in-context. On a 142-dataset UCR benchmark, TIGER reports the highest mean accuracy, balanced accuracy, and F1-score among six compared algorithms, including HIVE-COTE 2.0.
WhyCombines general classifiers over four representations with adaptive meta-classification, beating HIVE-COTE 2.0 across 142 datasets with thorough ablations.
In briefGraph-based recognition of simulated train-driver alertness from face and upper-body keypoints
SummaryThis paper presents a vision-based system that uses one front-facing RGB camera and a graph neural network to classify simulated train-driver states as alert, not-alert, or emergency-like behaviour. Comparing skeletal-only, facial-only, and combined features, the combination achieved 81% accuracy for the three-class model under light conditions and 99% for alert versus not-alert. The authors also released a controlled RGB video dataset recorded under three illumination conditions.
Olivia Nocentini, Marta Lagomarsino, Gokhan Solak et al.
In briefReinforcement learning for workload-aware chip power network optimization
SummaryThis paper proposes a reinforcement learning framework for optimizing workload-aware power delivery networks (PDNs) in VLSI chips. Power traces from system-level simulations are mapped to spatial power density, and a Deep Q-Network agent optimizes wire widths to reduce PDN area. Evaluated on 4-, 8-, and 16-core floorplans with PARSEC and SPLASH-2 workloads, it cuts average normalized area by 47% while satisfying EM and IR-drop constraints. It matches simulated annealing in quality and is about 26 times faster.
Oran Hayes, Maria Pantazi-Kypraiou, Athanasios Tziouvaras et al.
In briefFaster test-time adaptation for vision-language models by verifying candidate labels
SummaryVision-language models can lose accuracy when the target domain shifts at inference time. Most existing test-time adaptation methods adjust predictions, which adds computational cost. This paper proposes Test-Time Correction (TTC), which treats each top-k candidate label as a hypothesis, reconstructs the test feature in its memory-bank subspace, and selects the candidate with the smallest divergence shift. TTC is training-free and avoids iterative optimization, and it reports accuracy gains across 15 benchmark datasets and five settings, with up to 2x speedup.
In briefEffect of Data Augmentation on Confidence Calibration for Melanoma Classification
SummaryThis paper examines how data augmentation affects the confidence calibration of neural networks that classify melanoma in medical images. The authors trained models on the SIIM-ISIC 2020 melanoma dataset with and without augmentation and compared AUC and expected calibration error (ECE). They report that different augmentation methods improved uncertainty calibration.
In briefDemo: VLM-Guided Online Calibration of an Electromagnetic Digital Twin
SummaryThis demo addresses mobile robots that rely on an electromagnetic digital twin for wireless awareness, whose material conductivities change with the environment. The framework uses a vision-language model twice: first to classify visible materials and derive conductivity priors via ITU-R P.2040, then to choose the next measurement waypoint online. In a real indoor scenario, it reached a normalized mean absolute conductivity error of 1.74×10⁻⁴ within 20 m of travel. Random initialization never converged, and random waypoints needed over twice the travel.
In briefAttention May Not Explain Forecast Peaks: Temporal Reference Versus Forecast Output in Attention-Based Forecasting
SummaryThe paper examines whether cross-attention maps in an attention-based load forecasting model explain when forecast peaks occur. Across 31 windows of the Panama load dataset, the forecast's peak timing was closer to the observed peak (median error 0 h) than the attention argmax (median error 5 h). The authors conclude that structured, reproducible attention need not provide a valid peak-selective explanation, and that it may instead serve an internal reference role for integrating future exogenous information, a role not causally established.
In briefHow Loss Function Choice Affects Neural Network Performance in Melanoma Diagnosis
SummaryThis paper examines how different loss functions affect deep neural networks used to classify melanoma, the deadliest skin cancer, from images. Because melanoma datasets such as SIIM-ISIC are class-imbalanced, the authors compare focal loss, logit-adjusted softmax cross-entropy, and weighted softmax cross-entropy. They evaluate both classification performance and uncertainty calibration. The results suggest that focal loss achieves a good balance of AUC and expected calibration error.
In briefCombining graph sparsification and approximate computing for low-power FPGA graph neural network inference
SummaryThis paper presents an FPGA-based graph convolutional network (GCN) accelerator on the AMD Kria KV260 that combines DSpar graph sparsification, 8-bit quantization, and approximate multipliers. Tested on Cora, LastFM Asia, and Amazon Photo, it finds that approximate arithmetic effectiveness depends on accumulation depth, and sparsification improves viability by reducing that depth. The combined approach reaches up to 9.88× speedup with 86.6% accuracy on Amazon Photo, while total power stays below 1 W.
Nathaniel Kaye Mellor, Shreejith Shanker, George Floros
In briefAnchorGen: Anchored Optimization for Customizable Generative 3D Design
SummaryThis paper introduces AnchorGen, a rectified-flow framework for generating 3D shapes, trained without conditioning on paired shape and sketch latents. At inference, the sketch latent is kept close to a target sketch with a token-wise cosine penalty, while gradient descent optimizes the shape latent to minimize a differentiable drag surrogate. According to the paper, one model supports design-preserving optimization, dimensionally explicit edits, and sketch-only synthesis.
In briefMachine learning for journal entry testing: type-aware evaluation of anomaly detectors under a review budget
SummaryThe paper argues that journal entry anomaly detectors are usually judged on full-population metrics such as ROC-AUC, precision and recall, which ignore the review budget and which anomaly types are found. The authors propose a type-aware evaluation using per-type recall, fair-share type recall, type coverage and first-hit rank. They compare nine unsupervised detectors, a supervised reference and feedback-driven DeepSAD on four real client ledgers and a public synthetic ledger. The results show that high hit rates can hide systematic blind spots, and that feedback may reinforce existing patterns without broadening anomaly coverage.
In briefFew-shot bioactivity prediction using meta-learning that accounts for assay differences
SummaryThe paper addresses predicting compound bioactivity in early drug discovery when each assay has few measurements. It finds that meta-learning performance declines as meta-training tasks become more heterogeneous. The authors introduce MetaHeta, which conditions predictions on auxiliary data from related assays and combines linear attention with exact attention. Experiments on ChEMBL and BindingDB show improved few-shot prediction and compound prioritization in retrospective Bayesian optimization.
Michal Kmicikiewicz, Tommy Rochussen, Vincent Fortuin et al.
In briefCombining Survival-Based Wear Models with Data-Driven Remaining Useful Life Prediction
SummaryThis paper proposes a probabilistic fusion framework that combines a wear-based survival model with a sensor-based 1D-CNN model, using failure probability distributions, to estimate remaining useful life (RUL). Evaluated on multiple N-CMAPSS datasets, the fused model improves point accuracy, preserves the C-index, and gives narrower, well-calibrated prediction intervals than either component alone.
Abhishek Srinivasan, Juan Carlos Andresen, Sepideh Pashami et al.
In briefBenchmarking Guardrails That Stop Jailbreaks in Embodied AI Agents
SummaryThis paper presents the first systematic evaluation of jailbreak guardrails for embodied agents, which are AI systems powered by language and vision models that act in physical environments. The authors built a pluggable framework to test six representative guardrails under identical conditions, using template-based and automated jailbreak attacks as well as safe instructions. Results across defense effectiveness, usability, and efficiency show a clear trade-off among these dimensions, and no single guardrail performs best in all settings.
In briefMind the Drift: Diagonal Linear Networks at Large Learning Rates
SummaryThis paper studies how large learning rates change training in diagonal linear networks, where two implicit biases compete to determine the sparsity of the recovered regression solution. Beyond the Gain, which measures average discretization error of Gradient Descent relative to Gradient Flow, the authors derive a related but overlooked quantity called the Drift, describing imbalance between discretization errors. They introduce an intervention that steers the Gain to recover sharper, sparser solutions, and conclude that large learning rates can be used to control implicit bias rather than always hindering sparse recovery.
In briefORCA: An Annealed Spectral Conditioning Optimizer for Faster, Better LLM Training
SummaryThis paper introduces ORCA, an optimizer that applies strong but temporary soft orthogonality regularization early in training and then removes it, so weights keep a broader spectrum early and adapt freely later. Experiments cover LLaMA, Qwen3, and fine-grained mixture-of-experts models from 130M to 8B parameters. ORCA reaches lower final validation loss than Muon, requires no architectural changes, and adds minimal overhead.
In briefWhispering Wisely: Smart Subset Selection for Efficient Speech Recognition Fine-Tuning
SummaryThis paper addresses the high cost of fine-tuning speech recognition models such as OpenAI's Whisper. It proposes an embedding-based method that uses Whisper's pretrained encoder and a greedy k-center algorithm to select diverse, representative training subsets. Experiments on Hebrew and Danish show k-center selection often outperforms random sampling, especially at smaller subset sizes, and lexical diversity analysis suggests it covers a broader range of linguistic content.
In briefFlash-OPD: Fast On-Policy Distillation for Faster Student Training
SummaryThis paper proposes Flash-OPD, a method that speeds up on-policy distillation, in which a teacher model supervises a student's own generated trajectories. Instead of fixing a rollout length in advance, it decides for each trajectory when to stop based on observed teacher-student compatibility events, interleaving cached generation with teacher verification. The authors report speedups of 2.2x to 7.5x over standard on-policy distillation across diverse datasets and teacher-student settings, while maintaining or improving accuracy.
In briefTesting CLIP for gender guessing from full faces and eye regions without extra training
SummaryThis paper tests the CLIP image-text model for zero-shot gender estimation on full-face and periocular (eye-region) images from the Adience dataset. Full-face accuracy reaches 95.54% without task-specific training, while periocular predictions are biased toward males due to a misaligned decision boundary. Threshold alignment raises periocular accuracy to 85.29%, and linear SVMs on CLIP features reach a best of 86.17%. The remaining gap with full-face performance shows periocular gender estimation is harder.
Fernando Alonso-Fernandez, Kevin Hernandez-Diaz, Jose Maria Buades et al.
In briefAgentic Language World Models: Simulating Interactive Environments from Past Interaction Traces
SummaryThis paper proposes agentic language world modeling, where a world model agent serves as the environment for a task agent instead of rebuilding an executable system. It introduces Trace2Env, a learning-free framework that, when the original system is unavailable, turns historical interaction traces into a reusable environment worldbook with schemas, evidence, and behavioral knowledge. At runtime the world model consults this worldbook and persistent episodic state to produce observations and state changes. The authors report improved next-observation fidelity and long-horizon consistency over prompt-based baselines across nine environments.
In briefWhy a Third Data Type Can Hurt: Geometry-Based Fix for Multimodal Self-Supervised Learning
SummaryThis paper studies why adding a third modality (image, text, tabular) can lower performance in self-supervised learning. Across nine datasets, the trimodal model underperformed its best bimodal subset in 55.6% of paired VICReg runs, a failure the authors call multimodal saturation. They attribute it to alignment geometry and propose Riemannian VICReg, which aligns views by geodesic distance on learnable negative-curvature factors and reduces to VICReg as curvature vanishes. Over the same 45 paired runs, the third modality helped in 64.4% of cases, up from 44.4%.
In briefTransmission factors: measuring how PDE training-data compression affects a trained operator
SummaryThe paper argues that the common practice of setting compression tolerance by stored-field reconstruction error measures the wrong quantity, since what matters is the accuracy of a model trained on compressed data. The authors propose measuring a transmission factor by passing compressed fields through a model trained at full precision and comparing outputs against those from clean fields, which requires no training on compressed data. Across six PDE families, this ranks downstream cost better than field error and can suggest per-dataset storage rates. Back-checks against trained models find these suggested rates off by up to a factor of two in stored bits.
In briefRefineCBCT: Reducing artifacts in cone-beam CT for radiotherapy without paired training data
SummaryThis paper presents RefineCBCT, a framework that improves cone-beam CT (CBCT) image quality for image-guided radiotherapy using unpaired training data. It uses pseudo-label guidance and short-step diffusion to reduce artifacts while preserving patient-specific anatomy. Evaluated on the LUNG TCIA and PELVIC TCIA datasets against GAN and diffusion baselines, it reports the best results on several metrics. The authors state that code is publicly available on GitHub.
In briefAdaptive Mean Flow: a fast, smooth method for closed-loop robot control
SummaryDiffusion and flow-based robot policies are widely used in imitation learning, but their iterative denoising causes prediction latency that makes high-frequency closed-loop control difficult. This paper proposes Adaptive Mean Flow (AMF), which uses an accelerated form of Flow Matching to reduce latency and conditions predictions on a corrupted version of the previous step's trajectory to keep successive predictions consistent. The authors evaluate AMF on a range of simulated and real robot tasks and report significantly improved performance over baselines.
In briefRethinking least-core computation in contextual-distractor games
SummaryThis paper studies how payoff selection and coalition sampling affect least-core attribution, a game-theoretic method for assigning credit to features or training examples and identifying harmful contributors. The authors propose entropic least core (ELC), a smooth approximation whose unique minimizer follows a continuous path to the nucleolus. Experiments report that ELC approximates the nucleolus faster than an LP-based solver with small payoff errors, matches the minimum-norm selector in identification accuracy, and ranks distractors by harm more accurately in full-coalition games.
Hiroshi Kera, Toshinori Yamauchi, Sai Ganesh Nagarajan
In briefStronger transferable attacks on deep reinforcement learning agents
SummaryThis paper studies black-box transfer attacks on deep reinforcement learning agents, where perturbations crafted on a white-box surrogate agent are applied to an unknown victim. The authors propose a trajectory-level attack that optimizes a sequence of observation perturbations over a receding horizon. On CartPole-v1 with ten DQN and DDQN agents, it outperforms per-step attacks and random noise in white-box, cross-model, and cross-algorithm settings.
In briefDo VLAs Adapt to Handled Objects or Just Replay Learned Motions?
SummaryThe paper asks whether vision-language-action (VLA) models truly understand object physical properties and adapt their motions, or merely replay learned behaviors that happen to succeed. Probing seven VLAs, the authors find physical properties such as mass and friction are less decodable than semantic or non-physical properties. In a LIBERO case study that increased an object's mass, most models lifted heavier and original-mass objects similarly, causing task success to decline.
In briefFinal-iterate convergence rate of normalized gradient descent under Hölder smoothness
SummaryThis paper studies how the last iterate of normalized gradient descent converges for convex objectives with Hölder smoothness. With a constant stepsize, it proves an upper bound containing an extra logarithmic factor compared with known best and averaged-iterate results. Numerical experiments based on performance estimation problems examine whether this overhead is intrinsic. The authors show that a linearly decreasing stepsize achieves the same order O(T^(-(1+ν)/2)) for the last iterate without knowing ν or M_ν.
In briefAttention Tax, Handoff Tax: A model of when splitting LLM work across agents helps
SummaryThis paper presents a simple reliability model explaining when multi-agent LLM systems outperform a single agent. It identifies two trade-offs: splitting a task reduces long-context burden but incurs a handoff cost when information is passed between agents, while repeated sampling gains depend on how strongly failures are shared. The authors estimate parameters from single-agent and handoff runs on a ledger-reconciliation task. The model predicts a crossover at depth 10 and that decomposition wins at depths 20, 50, and 100, and the measured results matched these predictions.
In briefWhat Replay Must Keep: Separating Correctable Bias from Class Correspondence
SummaryThis paper studies logit replay methods such as DER and DER++ for class-incremental learning, separating classes known at storage time from those learned later and editing each group's stored scores before and after a task-level offset. On CIFAR-100 with DER++, fixed constants for later-learned classes stay within 1 percentage point of the original accuracy, and the offset reduces the cost of deleting their matching. Reassigning non-gold scores of classes known at storage still costs accuracy after the offset, and a similar cost appears in image distillation.
WhyProposes a controlled diagnostic separating correctable bias from class correspondence, with CIFAR-100 DER++ experiments offering clear decomposition evidence for replay design.
In briefTrustMI: Causally controlling how AI assistants trust users
SummaryThis paper studies whether language model assistants can be made to trust or distrust users they cannot verify, defining trust as willingness to accept vulnerability to another party's actions. The authors built 2,000 contrastive conversations across ability, benevolence, and integrity, and learned steering matrices with frozen model parameters. Tests on six instruction-tuned models from three families show trust decisions change monotonically in both directions, and the effect extends to harmful requests, prompt injections, and insider-threat settings.
Th\'eo Lasnier, Romain Froger, Maxence Lasbordes et al.
In briefVision Transformer ensembles for segmenting street panoramas
SummaryThis paper describes a first-place system in the PalmCity challenge for semantic segmentation of street panoramas. Nine pretrained segmentation models were compared under roughly equal computation budgets, and the two strongest were trained with three random seeds. Averaging class probabilities from three Encoder only Mask Transformer models gave 60.95% mean IoU on the public validation split and 57.08% on the hidden test leaderboard.
In briefPaying to learn and sharing to earn: incentive-based federated bandit learning for self-interested players
SummaryThis paper studies federated multi-player multi-armed bandits, where players share information through a central server to learn faster. It assumes self-interested players need incentives: sharing data earns a reward, while buying information incurs a cost. The authors propose a UCB-based algorithm, Buying-UCB, and derive upper bounds on group regret and buying cost. Numerical experiments validate the theory across different collaboration and pricing settings.
In briefObjective Evaluation of Modern Turkish Text-to-Speech Using Speech Quality Models
SummaryThis paper benchmarks four modern text-to-speech systems (Chatterbox, CosyVoice, OmniVoice, VoxCPM2) in fine-tuned and zero-shot setups, compared against a VITS baseline and natural speech, for Turkish, a morphologically rich lower-resource language. Eighteen objective metrics cover naturalness, intelligibility, signal quality, speaker similarity and distributional fidelity. The authors also study how quality varies with utterance length and measure speaker and naturalness drift over long chunked speech. The evaluation code is released.
Yunus Emre Ozkose, Alperen Kahraman, Ali Haznedaroglu
In briefROT: Rotating hidden states toward context vectors to reduce hallucinations in vision-language models
SummaryLarge vision-language models often describe objects that are not in the image. This paper proposes ROT, a training-free method that works layer by layer: it detects semantic drift in middle layers and uses a norm-preserving rotation to pull hidden states back toward the plane spanned by textual and visual context. A smoothing step then stabilizes the adjusted representations in later layers. The authors report that ROT reduces hallucinations across multiple benchmarks, architectures, and model scales.
In briefLocal2Mesh: Localized Contour-to-Mesh Reconstruction of the Left Ventricle from Sparse 2D Cardiac MRI
SummaryThis paper addresses reconstructing 3D left ventricle geometry from sparse 2D cardiac MRI contours, which suffer from inter-slice misalignment. It proposes Local2Mesh, which deforms a template mesh without requiring 3D mesh annotations, using geometry-aware alignment and a plane-aware Local Router. Experiments on the M&Ms-2 and ACDC datasets report better geometric and functional estimates than existing methods, along with zero-shot transfer between datasets. The reconstructed meshes are also reported to improve disease classification compared with sparse contours.
In briefRocketAgent: An AI Agent for Coordinating Multidisciplinary Design of Liquid-Rocket Thrust Chambers
SummaryThis paper presents RocketAgent, a long-horizon engineering agent for preliminary multidisciplinary design of liquid-rocket thrust chambers. A planning Coding Agent coordinates skills for performance sizing, subsystem optimization, geometry generation, and multiphysics assessment, using a knowledge graph and a shared Design Intermediate Representation to keep parameters consistent. Revision-aware checks invalidate affected results and block superseded inputs. In a representative simulation, the agent recovered from an infeasible cooling search through an engineer-authorized operating-point revision and then coordinated geometry generation and multiphysics assessment.
In briefGO-based clustering for learning causal gene regulatory networks at cluster level
SummaryThe paper presents GO4CD, an algorithm that uses Gene Ontology to group genes sharing a biological process into partitions at multiple granularities. It is integrated with the CLOC cluster-level causal discovery method. On multiple E. coli regulatory subnetworks, GO4CD partitions were inadmissible in 18.1% of cases, versus 65.3-82.3% for semantic-similarity baselines.
Azlaan Mustafa Samad, Wei Zhang, Ad\`ele H Ribeiro
In briefDisentangling Representations to Make Chest X-Ray Diagnosis Fairer
SummaryThis paper proposes a single-encoder framework that combines dual-level decorrelation with prototype-guided cross-group contrastive learning to reduce demographic dependence in chest X-ray diagnosis models. It also introduces DRAR, a new metric that measures how much demographic structure is reduced within disease representations. Evaluated on four classification tasks with 34,809 CheXpert test images across eight intersectional groups, the method lowers the mean equalized-odds gap from 15.41% to 10.86% and the AUC gap from 5.95% to 5.01% compared with ERM, with only a slight drop in mean AUC.
In briefMercerFlow: Flow Matching in a Kernel-Induced Latent Space for Forecasting
SummaryThis paper proposes MercerFlow, which performs conditional flow matching for probabilistic time series forecasting in a latent space defined by the Mercer eigenbasis of the prior kernel. Series are mapped into this latent space and an MLP learns the flow, avoiding sequential backbones like RNNs or Transformers. On five GluonTS benchmarks under a shared protocol with TSFlow, it matches or beats TSFlow on CRPS while using about 4.7 times less training memory and 3.5 to 4.4 times less time per epoch.
In briefCasual flash photos for Gaussian splat inverse rendering
SummaryThis paper recovers geometry, materials, and lighting from casual indoor photos taken with the flash on or off, without needing dark rooms or special hardware. It proposes a GS-anchored diffuse field, where a hash-encoded MLP is queried at the rasterized 2DGS depth, and renders static lighting with deferred shading. On five synthetic and three real indoor scenes, the method outperforms six recent baselines on diffuse color, albedo, roughness, and relighting, where PSNR improves by 4.17 dB over the next-best baseline.
In briefHow routine clinical and demographic data help lung ultrasound classify tuberculosis at the point of care
SummaryThis study combines lung ultrasound images with routinely collected clinical and demographic data to automatically screen for tuberculosis using deep learning. A simple average of the output scores from separately trained image and clinical classifiers matched or outperformed a more complex early-fusion approach, reaching an overall AUROC of 0.95, about 4 points above the image-only baseline. Greedy feature selection reduced the number of clinical inputs without hurting performance, while the point-of-care tests included in the study added little benefit.
Joshua M. Jansen van V\"uren, Christiaan M. Geldenhuys, Devendra S. Parihar et al.
In briefLightMTP: Lightweight Latent Multi-Token Prediction, a training method that predicts several future tokens with minimal added parameters
SummaryStandard language model pretraining predicts only the next token, which can encourage reliance on local patterns. Existing multi-token prediction methods often add many parameters and rely on external helper models to encode future tokens. LightMTP builds future token representations from the model's own hidden states, adding at most 1% extra parameters. It retains better performance on general language modeling benchmarks and shows similar gains in planning, coding, and reasoning.
Tamara Czinczoll, Julie Kallini, Gerard de Melo et al.
In briefDifferentiable Bit-Widths: Jointly Optimizing SVD-Based Pruning and Quantization for Ultra-Efficient LLM Compression
SummaryThis paper proposes a unified framework that jointly optimizes SVD-based pruning and quantization for compressing large language models, instead of applying them in two separate stages. The key method learns component-wise bit-widths in a differentiable way, so less important components can be assigned 0 bits and pruned. The authors report that the approach outperforms two-stage baselines even under extreme 1.61-bit quantization settings.
In briefA method that helps AI agents coordinate with unfamiliar partners by sharing visual focus
SummaryInspired by human joint attention, the paper introduces MATE, a multi-agent reinforcement learning method in which agents align their visual attention on scene-salient objects during interaction. This reduces reliance on conventions formed with a specific partner during training. The method is evaluated on three benchmarks: a Card Alignment Game, Level-Based Foraging, and OvercookedV2. The authors report that the joint-attention signal consistently improves coordination with unknown partners.
Giulia Benintendi, Constantin Ruhdorfer, Fabian K\"ogel et al.
In briefJoint Precision Neural Networks: learning task-specific data dependencies alongside prediction.
SummaryThe paper proposes PNN-Joint, which alternately optimizes a sparse precision matrix (the inverse covariance) together with graph neural network weights, so the learned graph is tailored to a specific task. It also introduces Precision Neural Networks, a broader class of graph convolutional networks built on precision estimators, and relates them spectrally to PCA and covariance neural networks, including stability to finite-sample errors. Experiments on synthetic, neuroimaging, and motion sensor data report interpretable graphs, robustness in low-data settings, and best or second-best results on real-world tasks.
Andrea Cavallo, Samuel Rey, Antonio G. Marques et al.
In briefCuratorMAS: Automating Dataset Curation with Multiple AI Agents
SummaryThis paper introduces CuratorMAS, a multi-agent framework that automatically evaluates and curates high-quality datasets. It breaks the curation process into five programmable stages, including dataset exploration, retrieving online domain knowledge, computing evaluation metrics, filtering, and an evolution module that updates skills. Experiments report that it reduces the noise rate by up to 36.03 percentage points and improves downstream model F1 scores.
In briefScalable Minimal-Change Learning for Controllable Image Editing
SummaryThis paper introduces ARRO, a reinforcement learning method for instruction-based image editing that aims to change only the requested attributes and keep everything else intact. An agentic vision-language reward model checks for unimplemented requested changes and unintended changes, without per-instruction human annotations. On FLUX.1 Kontext-dev, average EditScore rises from 5.21 to 5.88 across four benchmarks, and off-target pixel change drops by 8.4% on 600 evaluation examples.
In briefFinding Similar Cell Structures in Electron Microscopy Using Image Patch Queries
SummaryThis paper presents a patch-based retrieval framework for locating self-similar cellular structures in volume electron microscopy datasets. It uses self-supervised learning of local image descriptors, so only a few manual annotations are needed to retrieve similar structures across an EM volume. Experiments on real biological tissue images show the framework reliably identifies relevant structures and substantially reduces the search space for downstream analysis.
In briefAdaptive expert guidance to make on-policy reinforcement learning more sample-efficient
SummaryThis paper addresses the sample inefficiency of on-policy reinforcement learning methods such as PPO when training from scratch. The proposed method lets the learner and a suboptimal expert alternate control within each episode, with the expert's share of control being a single learnable parameter optimized jointly with the policy. The expert's share decreases as the learner improves and eventually vanishes. The authors evaluate the method on 34 tasks across two benchmarks and report improved sample efficiency over guided and unguided baselines.
In briefTesting whether spatio-temporal video grounding models ignore the query
SummarySpatio-temporal video grounding locates objects or events in a video in both space and time, based on a natural language query. This paper shows that current models can still produce plausible predictions when the query is unrelated to the video or removed entirely. The authors analyze the HCSTVG-v2 and VidSTG datasets for regularities that may encourage this query-insensitive behavior. They call for evaluation protocols and architectures that explicitly assess query relevance.
Eryk Ko{\l}odziejczyk, Alberto Presta, Karol Szurkowski et al.
In briefParallel Haggling or Giving In? Concurrency-Aware Procurement Negotiation for AI Shopping Agents
SummaryThis paper studies how an AI buyer agent should run several procurement negotiations with different sellers at once, since each extra thread consumes resources and simultaneous deals carry cancellation risk. The authors propose CANO, a deterministic optimizer that jointly chooses the number of parallel negotiators and a common price cap. They report structural results on how concurrency and concessions interact, and show that CANO outperforms common heuristic policies across several market configurations and stress tests.
In briefSpectral geometry of attention: from information routing to uncertainty, a study of how transformer attention routes information and an attention-based uncertainty estimator.
SummaryThe paper analyzes transformer attention heads as functional maps between Hilbert spaces and derives a Token Difference Operator whose spectrum describes how information is routed to the output. It argues that Euclidean spectra are biased by attention sinks and proposes a token difference spectrum that separates sink effects from routing capacity. Based on this framework, the authors develop an attention-based uncertainty estimator that complements probability-based scores, with the largest gains on long-context inputs.
In briefD-Loop: Looped Diffusion Drafting to Speed Up Language Model Decoding
SummaryThe paper identifies a "repetition trap" in block diffusion drafters for speculative decoding, where neighboring positions predict the same token, which shortens accepted drafts. It proposes D-Loop, which adds intra-block causal conditioning by reusing the same drafter backbone in a second looped pass that regenerates a suffix given a selected prefix, trained with a prefix–suffix objective. On eight math, code, and chat benchmarks, D-Loop is reported to outperform DFlash and DSpark on Qwen3-4B and Qwen3-8B.
In briefUltrasound guidance via world modeling and retrieval-based action planning
SummaryThis paper presents a method for guiding ultrasound operators by treating guidance as multi-step planning and retrieval within a world model. Images are encoded with a V-JEPA 2.1 backbone, and similar views are retrieved from a database of physically executed probe transitions, requiring no probe tracking hardware. On carotid ultrasound, the planner reached the target view in 86% of retrospective closed-loop episodes and 83% in a prospective feasibility study on unseen volunteers.
Noortje I.P. Schueler, Hans van Gorp, and Ruud J.G. van Sloun
In briefLearning While Scheduling Jobs With Unknown, Context-Dependent Service Rates: A Rate-Optimal Algorithm
SummaryThe paper studies contextual queueing bandits, where a learner schedules jobs while learning unknown service rates modeled by logistic functions of job-server features. It proposes WISE, which achieves queue length regret of order Õ(t^{-1/2}) at every sufficiently large time without knowing the horizon, under capacity slack and with no covariance lower bound. The authors also provide sharpened and new lower bounds, and simulations show small regret even when context-wise slack fails.
In briefLRCC: Token-wise routing to make low-rank compression adaptive
SummaryThis paper introduces Low-Rank Conditional Computation (LRCC), which adds a lightweight router to each Transformer block of a pretrained model to choose among nested low-rank paths for each token. Only the routers are trained while the low-rank factors stay frozen. Experiments on Llama and Qwen models show better predictive performance than static low-rank compression under the same average active-parameter budget.
Thomas Vaitses Fontanari, Maximo Eduardo Rulli, Federico Alvetreti et al.
In briefLearning Decision-Stump Thresholds from Context: Softmax Attention Dynamics
SummaryThe paper studies how gradient-based pretraining teaches a two-parameter softmax-attention model to estimate a decision threshold from context. It shows that, under a specific initialization and constant-step gradient descent, the learned estimator's error for each fixed threshold decreases with the number of tasks, examples, and fresh-context size. The mechanism involves the attention scale growing over training time, and the authors also discuss a limitation of the one-head model.
In briefP3: A Persistent Particle Method for Constrained Diffusion-Based Control
SummaryThis paper introduces Persistent Particle Planning (P3), a sequential Monte Carlo framework that keeps a weighted set of candidate trajectories from a diffusion model across successive replanning steps. At each step, the candidates are shifted, partially re-noised, refined under the latest observation, and reweighted and resampled to respect constraints, without retraining the model. Experiments show that reusing the population reduces route switching and improves success without constraint violations. On maze navigation, P3 also plans faster than regenerating populations or correcting a single plan.
In briefAgentSpy: Observing What AI Agents Actually Do
SummaryAI agents built on large language models run shell commands, access files, and use the network, but their real behavior is hard to see from test results or their own reports. This paper presents AgentSpy, which runs agents in an isolated environment and records system calls and network traffic from the agent and its subprocesses. It supports conformance analyses of required behavior and safety analyses of prohibited behavior, and it was evaluated on 77 tasks and for detecting attacks. The security rules detected four of five attack categories with no false positives across 50 runs.
Christoph B\"uhler, Matteo Biagiola, Luca Di Grazia et al.
In briefColor Alignment in VAE Latent Spaces and Its Applications
SummaryThis paper studies how the variational autoencoders behind text-to-image models represent color in their latent space. The authors find that a range of VAEs, from SD1.5 to FLUX.2 and Z-Image, share a color subspace aligned with brightness and opponent-color axes. Building on this, they propose three applications: ColorTuning for precise numerical color generation, saturation control, and color transfer to match a reference palette.
Julian D. Santamaria, Kai Wang, Jes\'us Malo et al.
In briefA BEMD-based quaternion filter that converts sharp-kernel CT images to soft-kernel-like images
SummaryThis paper proposes a method that makes CT images from sharp reconstruction kernels look like soft-kernel images, reducing noise while keeping anatomical structures. It combines Bidimensional Empirical Mode Decomposition with Quaternion Bilateral Filtering (BEMD-QBF) and works on images after reconstruction, without raw projection data. Tests across six kernels report higher SSIM and PSNR than Non-Local Means, Anisotropic Diffusion, and Bilateral Filtering.
Mahmoud Nasr, Jan K. Argasinski, Krzysztof Brzostowski et al.
In briefLangevin Flow Maps: Fast Molecular Dynamics Simulation with Larger Timesteps
SummaryMolecular dynamics simulations integrate Langevin equations in tiny femtosecond steps, making long-timescale properties and transitions hard to estimate. This paper introduces Langevin Flow Maps, which extend machine-learned force fields to also learn the stochastic Langevin integrator. The authors report that the method allows large-timestep simulations while recovering accurate dynamical properties, running about an order of magnitude faster than current machine-learned force fields. Training on a diverse molecular dataset is presented as a route toward transferable models.
Sam McCallum, Niklas Rindtorff, Alexander Tong et al.
In briefEpicWorldModel: An exploration-driven planning method using stochastic latent world models
SummaryThis paper introduces EpicWorldModel, a framework for training stochastic JEPA-based latent world models in partially observable environments with uncertainty. When goal-relevant scene content is missing from the history, the model uses flow matching to predict multiple possible future states. Its predictive variance guides exploration within Cross-Entropy Method planning, and experiments report up to 22% higher success rate than LeWorldModel.[]
In briefTurnslide: A Lightweight Method for Scalable Multi-Turn Training Data
SummaryThis paper presents a fully automated framework that synthesizes multi-turn tool-calling data by modeling each API as a finite-state machine. Tool sequences are generated to be valid under these states and converted into complete examples with a single LLM call. Fine-tuning small language models on the generated data significantly improves accuracy over an unmutated baseline, reaching 70.7% full accuracy and using 3.6 to 6.6 times fewer tokens than existing methods.
Aaron Fainman, Gabriela Kadlecov\'a, Maciej Gryka et al.
In briefHow to Use Data Augmentation in Vision-Language-Action Model Post-Training, and How Not To
SummaryThis paper systematically studies how image augmentation affects reinforcement learning post-training of vision-language-action (VLA) robot models. It finds that augmenting only the critic during RL updates, while keeping the actor's input clean, works best, whereas augmenting the actor makes training collapse entirely. The authors report gains on LIBERO-Plus out-of-distribution tests and offer practical recommendations.
In briefFrom Transformation to Target State: Rethinking Query Representation for Zero-Shot Composed Image Retrieval
SummaryThis paper studies zero-shot composed image retrieval, which finds a target image from a reference image plus a modification text. It argues that the text describes a change while candidate images show completed states, creating a representation mismatch. The authors reconstruct a static target-state representation with a frozen multimodal large language model before retrieval, implemented as the training-free ASAP-CIR framework. Experiments on FashionIQ, CIRR, and CIRCO show the clearest gains on the multi-target CIRCO benchmark.
In briefForgetting Across Modalities When Post-Training Unified Multimodal Models
SummaryThis paper studies a risk specific to unified multimodal models, which handle both image understanding and image generation with one set of weights: fine-tuning one capability can harm another. The authors apply LoRA post-training to Janus-Pro-7B and find that a single training run leaves untrained capabilities intact, but chaining three runs erases earlier learned abilities and sharply raises general-text perplexity. They attribute this to an objective that constrains only the trained modality, and report that anchoring the other output head prevents the damage.
WhyIdentifies cross-modal forgetting in unified multimodal post-training, with chained-run experiments and an anchoring fix supporting the claims.
Francesco Colasurdo, Ivan Slobozhan, Joaquin Vanschoren et al.
In briefLabel-free coreset selection using pathology foundation models to reduce annotation effort
SummaryThis paper presents GCcore, a label-free coreset selection method for computational pathology. It embeds each image with a pathology foundation model and greedily selects samples that maximize global coverage of the embedding space, reducing the need for expert annotation. The method is hyperparameter-free and deterministic, with a lower-bound guarantee on global coverage for any coreset size. It outperforms 14 baselines across 10 tasks, ranking first on six.
WhyProposes a hyperparameter-free, deterministic greedy coreset method with a coverage lower bound, validated on 10 tasks against 14 baselines for pathology annotation.
In briefBreaking the Tie: A Cluster-Aware Routing Framework for Large Language Models
SummaryAs many large language models exist, choosing the best one for each user query is a challenge. This paper argues that when several candidate models answer the same query correctly, existing routing methods are misled, causing a sharp drop in generalization to unseen tasks. The authors propose CASLR, which trains a lightweight router using continuous soft labels based on cluster-level utility scores instead of one-hot labels. The paper reports a 7.80% higher overall average than Llama-3.3-70B-Instruct and a routing latency of 1.13s.
In briefByte-level language models: scaling, learned abstractions, and where uncertainty lies
SummaryThis paper studies language models that skip tokenizers and model text directly as bytes. As model size grows, byte-level Transformers consistently outperform subword Transformers when trained with token-superposition and hash embeddings. The models learn local text abstractions resembling external tokenizers, and restricting up to 25% of intermediate layers to these representations preserves downstream performance. Exploiting the induced structure for speculative decoding yields 3.4 times more accepted tokens than subword Transformers.
In briefStagQ: A Multi-Precision Weight Format for LLM Quantization
SummaryThe paper presents StagQ, a multi-precision weight format that stores one data stream in which lower-precision versions are valid prefixes, so several precisions can be served without storing multiple copies. The main stream is a 2-bit group-wise affine base followed by configurable 1-bit refinement planes. On Llama-3.1-8B, Phi-4, and OLMo-2-7B, the cheaper 2-bit configuration beats the strongest multi-precision baseline by 3.1 to 7.0 MMLU points at a slightly lower logical rate. In synthetic-weight tests on an NVIDIA A100 GPU, the kernel is faster than both baseline kernels in most shape-precision cases.
In briefCan language models learn to reject their own faulty reasoning steps?
SummaryThis paper asks whether a language model can reject its own bad reasoning steps without an external verifier. The authors propose Self-Step Rejection (SSR), which trains a lightweight LoRA acceptance gate on a frozen generator to accept candidate steps or resample them. Across three reasoning models and five math benchmarks, SSR improves macro-average accuracy over single-pass decoding while using about 1.21 to 1.40 times as many generated tokens.
In briefTransfer-Stratified On-Policy Distillation for Reinforcement-Learning-Improved Reasoning Teachers
SummaryThis paper studies whether gains from reinforcement learning (GRPO) on a reasoning teacher model carry over when the teacher guides a smaller on-policy student, using mathematical reasoning tasks. It proposes Transfer-Stratified On-Policy Distillation (TS-OPD), which routes training problems by joint student-teacher success, uses gated forward or reverse KL objectives, and adds an entropy brake. The authors conclude that transfer is structured rather than scalar, so a stronger teacher helps mainly when supervision direction and token budget match the student's ability.
In briefDenoising CT images by combining decomposition methods with curvelet thresholding
SummaryThis paper proposes a hybrid denoising framework that combines several decomposition methods, including EMD, VMD, MEMD and BEMD, with curvelet transform thresholding. It is tested on CT images reconstructed with kernels B50, B46, B41 and B36. VMD achieves the highest PSNR and SSIM, and the study compares soft and hard thresholding trade-offs between detail preservation and noise reduction.
Mahmoud Nasr, Jan K. Argasinski, Krzysztof Brzostowski et al.
In briefReachability-Aware Diffusion Policy Optimization (RADPO): A Model-Free Method for Safe Reinforcement Learning
SummaryThis paper proposes RADPO, a model-free method for optimizing diffusion policies in continuous-control reinforcement learning. It learns a discounted first-hit reachability value to estimate the risk of cost events and combines it with cumulative-cost budget feedback to shape rewards. Across ten continuous-control safety tasks, RADPO achieves competitive reward-cost trade-offs and substantially reduces constraint violations on several tasks compared with the baselines.
In briefFaster last-iterate convergence for two-player zero-sum games learned from bandit feedback
SummaryThis paper studies two players in a zero-sum discounted Markov game who learn independently from bandit feedback, without seeing each other's actions. The authors propose Adaptive Regularized TD Learning (ARTD), which under a uniform hitting time assumption achieves a duality gap bound of roughly t^(-1/4) for the current policies. This improves on the prior rate of Cai et al. (2023), and the algorithm requires no knowledge of the hitting time bound, time horizon, or confidence level.
WhyImproves the last-iterate convergence rate for bandit-feedback zero-sum Markov games with a new adaptive regularized TD algorithm and analysis.
In briefJLD: A Perceptual Image Distance Built from a Frozen Encoder's Jacobian
SummaryThe paper introduces the Jacobian Lens Distance (JLD), a perceptual image distance derived from a frozen vision encoder instead of human judgment labels. It fits a fixed metric tensor from the encoder's Jacobian using only 100 unlabeled images, in about 35 seconds. JLD reportedly outperforms LPIPS, DISTS, PieAPP, and DreamSim across four perceptual databases and stays stable when image resolution doubles. The authors also present JLD-fast and an extension to video.
WhyDerives a fixed Jacobian-based metric without human labels, outperforming baselines on four perceptual databases and extending to video with solid evaluation.
