In briefLearning from Failure: Brain-Inspired Self-Evolving Agents for Embodied Spatial Intelligence (HiMeco)
SummaryThis paper proposes HiMeco, a self-evolving agent for embodied spatial intelligence inspired by metacognitive monitoring and hippocampal memory consolidation. Through Hippocampal-Metacognitive Reflection, it diagnoses failures in perception, exploration, and planning, and compares execution attempts to derive reusable strategies. It also abstracts interaction programs into parameterized skills, validated by execution before reuse. The authors report state-of-the-art results on ESI-Bench and a 3.94 percentage point accuracy gain on EmbSpatial-Bench without evolution.
In briefEvent vs. Frame Cameras for Seeing Through Occluders: A Simulation Study
SummarySynthetic-aperture imaging removes occluders like foliage by combining many views along a camera path, and has mainly used frame cameras. This simulation compares event and frame cameras viewing the same scene and motion, with a new method that projects each event onto the focal plane. In bright high-speed scenes, the frame camera records only three images and cannot remove the occluder, while the event camera reconstructs the target. At 1 m/s with matched data budgets, the event camera shows no reconstruction advantage, and its data rate depends on the scene.
Jakob Hinterholzer, Mohamed Youssef, Oliver Bimber
In briefSuper-resolving turbulent flows from a single high-resolution sample using patch-based learning
SummaryThis paper trains a wide-activation super-resolution network (WDSR) only on overlapping patches taken from one high-resolution two-dimensional homogeneous isotropic turbulence realization. Unseen flow fields are reconstructed by patchwise prediction and Hann-window-weighted overlap-add assembly at super-resolution factors of 32 and 64. The method reconstructs coherent vortical structures and recovers fine-scale features better than conventional interpolation, as shown by spectral and statistical analyses. The authors suggest that exploiting spatial redundancy within one realization is a promising data-efficient approach when high-resolution training data are scarce.
In briefSpectrumQuant: Spectrum-aligned orthogonal rotations for LLM quantization
SummaryThis paper proposes SpectrumQuant, a closed-form method for choosing orthogonal rotations to improve post-training low-bit quantization of large language models. It estimates activation covariance from calibration data and builds a covariance-aware transform, with Hadamard matrices as a special case. Experiments on Llama and Mistral models report competitive accuracy with state-of-the-art rotation methods, using less than one-third of the GPU memory and much less time than learned rotation approaches like SpinQuant and DartQuant.
Dwaipayan Haldar, Indranil Patra, Azhar Yousuf et al.
In briefMMLA-RSI: A paper on recursive self-improvement via RTT and RTMA
SummaryThe paper proposes MMLA-RSI, a recursive self-improvement scheme with two routes in a Memory-Mediated Learning Architecture. Reasoning-Time Training changes the policy state, while Reasoning-Time Memory Adaptation changes authoritative memory, and each route also improves its own adaptation method. The authors give bounded rule languages and a shared algorithm, and report partial experimental results. They state that whether the recursive chains are beneficial and how efficient discovery is remain to be evaluated.
In briefCASI: Assigning Credit Within Agentic Retrieval Action Sets Using Set Interventions
SummaryThe paper introduces CASI, a method for assigning credit among multiple retrieval actions that a language model generates jointly while receiving only one reward for the whole set. It fixes replayable snapshots and evaluates alternative subsets under common conditions to build counterfactual signals, then uses Shapley aggregation to allocate within-set credit for policy optimization. In a five-action sponsored-search setting, policies trained with CASI achieved higher held-out auction-replay set value than baselines, and a seven-day production A/B test on 10% of traffic showed an observed 3.51% lift in the primary advertiser-value metric.
WhyIntroduces snapshot-grounded set intervention with Shapley credit; strong replay experiments and a production A/B test support it, though impact is mainly within sponsored search.
In briefTiming Retrieval in Medical Reasoning: Adaptive Evidence Lookup Using Internal Model Signals
SummaryThis paper proposes PUlSE, a lightweight probing method that uses a language model's internal layer representations to decide, at each reasoning step, whether external medical evidence should be retrieved. The authors observe that incorrect steps show higher inter-layer volatility, and train a compact probe on these features to detect errors and trigger retrieval and regeneration. Across four medical benchmarks, PUlSE matches exhaustive per-step retrieval in accuracy while reducing retrieval calls by 72%, and it transfers to an unseen legal domain.
In briefWhen correctness feedback hides evidence: phase-separated completeness auditing in Verus
SummaryThe paper studies how correctness feedback suppresses evidence that specification-completeness auditing should find, namely programs that meet a written contract but violate intended behavior. On 12 Verus/Rust tasks, no-feedback discovery yielded 59.0% certified witnesses versus 27.1% for IntentShield's corrective-feedback loop. Preserving these witnesses led to better specification repair, with 81.3% implementation success versus 62.5%. The effect did not reproduce in the available Dafny and Lean evaluations.
In briefPredicting Unseen Interactions: Flow Models Hallucinate Non-Additive Effects
SummaryThis paper studies whether conditional flow-matching generative models can recover interactions between two interventions when predicting unseen conditions. Experiments use a physiological time-series simulator, drug dose-response data, combinatorial cell perturbations, and molecular conformer generation. Individual effects are learned earlier than interactions, and models can hallucinate interactions where none exist. Interactions are recoverable only when condition structure continues from observed to held-out targets.
In briefFrom Plato to Aristotle: Toward a General Theory of Representation Hypotheses
SummaryThis paper examines the Platonic Representation Hypothesis, which claims that increasingly capable models converge toward a common representation, and why studies reach different conclusions. The authors propose the Lattice Representation Hypothesis, arguing that different alignment criteria sit at different points of one lattice. They support this with proofs in a linear autoencoder and a finite contrastive Gibbs model, plus experiments on dozens of pretrained vision and language models.
WhyUnifies conflicting representation hypotheses via a lattice framework, with proofs and large-scale experiments supporting the claims.
In briefECHOES: a method that compresses GUI agent visual history into text and is trained with privileged teacher guidance
SummaryThis paper introduces ECHOES, a method for long-horizon GUI agents that compresses visual history into a textual state and reasons about it. During training, a teacher model with privileged multi-image inputs guides the student through on-policy distillation. The authors also present ECHOESBMK, a diagnostic benchmark built from real GUI agent failures, and report gains on four GUI benchmarks plus zero-shot transfer to unseen visual planning tasks.
Tengda Guo (1), Jiawang Hu (2), Yifei Tao (3) et al.
In briefNeural part-of-speech tagging for the low-resource Saraiki language
SummaryThis paper builds a manually annotated Universal POS tagging dataset for Saraiki, covering 17 part-of-speech categories. Three pretrained transformer encoders are compared under the same fine-tuning setup. XLM-RoBERTa base performs best, reaching about 88.60% test accuracy. The dataset and baseline results are offered as a benchmark for future Saraiki NLP research.
Usman Nawaz, Muhammad Junaid Iqbal, Muhammad Asaf et al.
In briefA Valid Watermark Doesn't Guarantee a Reliable Face Identity: A Watermark-Preserving Face Evasion Attack
SummaryThe paper proposes a black-box adversarial attack (PGD-L_2) that lowers face-recognition similarity while adding spatial and spectral preservation terms to limit pixel distortion and high-frequency changes that could disturb hidden watermarks. In experiments on two watermarking schemes, two face-recognition models, and two datasets, evasion rates exceed 92% at operational thresholds while watermark recovery stays high and visual quality remains above 47dB PSNR. The authors conclude that watermark-based facial certification has a structural gap: an image can stay watermark-valid and visually faithful yet become biometrically unreliable.
Jefferson David Rodriguez Chivata, Davide Ghiani, Simone Maurizio La Cava et al.
In briefTANGO: Jointly Designing RNA Topology and 3D Geometry
SummaryThis paper studies RNA inverse design, finding sequences that fold into a target 3D structure. The authors argue that prior methods treat only tertiary geometry and ignore secondary base-pairing topology, so they propose TANGO, a multi-objective multi-agent reinforcement learning framework that handles both. TANGO uses a staged curriculum and frontier-guided fine-tuning. Experiments report improvements of 18.3% in tertiary-structure and 52.7% in secondary-structure performance over the strongest baseline, with diverse and novel designs.
In briefResolution Allocation with Learnable Resizing and Spectral Filter Parameterization
SummaryThis paper examines how learnable downsampling ratios interact with spatial and spectral-domain convolutional kernels in convolutional networks. The authors find that the spectral kernel benefits more from learned resolution allocation than the spatial kernel. The pattern appears on EuroSAT, CIFAR-10 and STL-10 at two model sizes, with accuracy gains differing by +0.013 to +0.028. The difference remains positive when total retained area is matched.
In briefCSWBench: Testing whether AI switches languages appropriately in creative writing
SummaryThis paper introduces CSWBench, a benchmark for measuring whether large language models switch languages in creative writing in a way that fits the context. It covers two settings, Focused Narrative and Extended Narrative, and proposes the CSWEval framework, which evaluates both individual switching events and whole texts. Testing twelve LLMs, the authors find meaningful but uneven code-switching competence across models and narrative contexts.
In briefFlowSep-RF: Rectified Flow Matching for Separating Overlapping Radio Signals from One Channel
SummaryThis paper proposes FlowSep-RF, a rectified flow matching method that separates a signal of interest from a co-channel interferer when the two overlap in time and frequency using a single receive channel. The model combines time-domain IQ samples with frequency-domain features and recovers the separated waveform with one endpoint velocity evaluation. Tested on QPSK mixed with OFDM-BPSK, OFDM-QPSK, and RFChallenge CommSignal2 interference, it achieves the lowest average BER and MSE among the compared classical, discriminative, and generative baselines in all three settings.
In briefFUSE: Screening Client Updates to Defend Fair Federated Learning Against Byzantine Attacks
SummaryThis paper studies Byzantine robustness in q-fair federated learning, where clients with higher loss receive larger weight, a setting that had been largely unexamined. The authors formalize the problem under an update-only interface in which the server sees only fairness-scaled client updates, and show how fairness scaling changes the geometry of honest updates. They propose FUSE, an aggregator that uses a small trusted labeled dataset for coefficient recovery, scale and cone-consistency screening, and step-size control. Across four benchmarks and seven Byzantine attacks, FUSE improves worst-attack accuracy by 3.9 to 30.1 percentage points over the strongest baseline while largely preserving the influence of high-loss honest clients.
In briefAnuvani-LID: A Robust, Calibrated Model for Identifying Spoken Indian Languages
SummaryThe paper presents Anuvani LID, a dedicated 115M-parameter conformer model for spoken language identification, trained on 54,643 hours of speech covering 22 Indian languages and Indian-accented English. The authors report that some general speech models and audio-capable LLMs perform poorly on very short clips, with one-second inputs in particular. Across 37 test conditions the model reaches a mean macro-F1 of 0.914 and an expected calibration error of 0.008.
In briefHarness Policy Distillation: Aligning Agent Policies with Runtime Semantics
SummaryThis paper introduces Harness Policy Distillation (HPD), a method that makes the runtime control rules of a product-specific harness, such as which actions are executable and when the loop ends, an explicit training target for agent models. It replays reference trajectories through the harness to recover these rules as control documents, uses them to guide a teacher model, and verifies the resulting actions before jointly training a shared model. On TRAJECT-Bench with the Pi harness, the HPD-trained Qwen3-8B reaches 22.98% autonomous task success and 49.23% runtime compliance, compared with 3.09% and 19.90% for the unadapted model.
In briefWhy Long-Horizon RL Needs Muon: A Study of Asynchronous Single-Rollout Training
SummaryThis paper studies reinforcement learning for long-horizon agents that must coordinate reasoning, tool use, and environment feedback over extended interactions. The authors examine an asynchronous single-rollout recipe combining the Muon optimizer, token-averaged loss, truncated importance sampling, and auxiliary length and tool-error penalties. The method is evaluated on repository-level issue resolution across two model families, and the authors report higher training success with the full recipe than with the baselines.
In briefGLoD-LVSM: Compact Geometry-Based Scene Representations for Faster Novel View Synthesis
SummaryThis paper presents GLoD-LVSM, a scene representation for large view synthesis models that compresses scenes using continuous levels of detail. It learns a continuous field to initialize a flexible number of anchors and uses geometry-grounded attention to condense scene information. According to the paper, compared with the uncompressed base model, it cuts scene tokens by 33.4% and speeds rendering 1.23 times with no quality loss, rising to 85.5% and 3.00 times under a relaxed quality budget.
Weikai Lin, Nicholas Milef, Carl S Marshall et al.
In briefRepairing Speech Access After Text Updates with AFFIRM
SummaryText updates can add knowledge that is hard to reach through speech, even when the model is otherwise unchanged. The authors propose AFFIRM, a detachable operator that keeps the base model frozen and repairs speech-only access in one pass. On balanced synthetic-speech PopQA, it shows no clear advantage over SALSA or low-rank ReFT. On recorded SLUE-SQA-5 questions, the PopQA-fitted operator transfers without refitting and improves ROUGE-L by 0.0941 over its paired baseline.
In briefAdversarial Reinforcement Learning for Robust Planetary Landing
SummaryThis paper proposes training a planetary landing controller against a second agent that generates physically constrained disturbances, forming an asymmetric zero-sum game in which both are alternately optimized. Preliminary experiments in a three-degree-of-freedom landing environment indicate that these adversarial disturbances expose vulnerabilities missed by conventional randomized perturbations. The authors suggest that a more extensive training campaign could improve robustness and generalization.
Miguel Leiva-V\'elez, Nicolas Gaston-Rozado, Andrea Scorsoglio et al.
In briefContrastive Uncertainty Learning for Robust Multi-Structure Segmentation in Lower-Pelvic MRI
SummaryThis paper proposes Contrastive Uncertainty Learning (CUL), which uses Monte Carlo Dropout to estimate voxel-wise predictive uncertainty and guide representation learning at ambiguous boundaries. Evaluated on a 589-scan 3D prostate MRI segmentation task covering eight structures, it keeps average Dice comparable to the baseline (83.38% vs 83.35%). Surface distance decreases significantly in six of the eight structures, suggesting reduced risk of extreme errors.
In briefAdaptive, survey-aligned simulation of urban travel behaviour using generative agents
SummaryThe paper builds an agent-based model of daily multimodal travel in Toulouse, France, and compares fifteen decision-makers on a real household travel survey. These include simple baselines, models fitted to the survey, generative agents with minimal or tuned prompts, and a typed classifier. Generative agents with tuned prompts approach the fitted models on the modal split of an undisrupted simulated day, but not trip by trip, and they also react when unforeseen events occur.
In briefScaling multilingual mixture-of-experts: how language coverage shifts capacity and compute
SummaryThis paper studies whether mixture-of-experts models can reduce the cost of adding languages to a multilingual model. The authors propose a repetition-aware joint scaling law linking active parameters, training tokens per language, number of experts, and language coverage. It is fitted on 26,300 per-language loss observations from 64 pretrained configurations. At matched active parameters and tokens per language, 32 experts cut the loss penalty of expanding from 12 to 200 languages by 83% compared with a dense model.
Joel Dag, Nikit Srivastava, Hamada M. Zahera et al.
In briefRAMSAD: retrieval-based selection of anomaly detectors for time series
SummaryThe paper addresses the fact that no single anomaly detector performs best across all time series, and that existing model selection methods lose accuracy on out-of-distribution data. It proposes RAMSAD, which retrieves the top-k most similar series from a knowledge base and transfers detector recommendations from their performance profiles. The authors report that RAMSAD outperforms baselines in-distribution and, out-of-distribution, matches or exceeds ensembles running the full detector pool at a fraction of the cost.
Christos Panourgias, Roberto Stanzione, Adrien Petralia et al.
In briefPAWA: Panoramic action-conditioned world modeling for multiple agents
SummaryThis paper introduces PAWA, a dataset and framework for predicting how multiple players' actions change a shared Minecraft world. It pairs each player's point-of-view video with world-stabilized panoramic video and aligned actions. The authors propose a dual-tower model that jointly predicts both video types, linking them through cross-attention in a shared world frame. The reported experiments show better point-of-view prediction than prior multiplayer world models, along with high-fidelity panoramic predictions.
In briefMAEM: A replay-free, capacity-controlled mixture of experts that reduces recency bias in class-incremental graph learning
SummaryThis paper addresses catastrophic forgetting in graph neural networks when classes arrive sequentially, where predictions tend to favor recently learned classes. The authors propose MAEM, which trains and freezes a dedicated expert for each session and routes test nodes to the expert with the lowest masked autoencoder reconstruction error, avoiding task IDs and a trainable router. When the expert pool reaches a fixed capacity, experts are merged and distilled using synthesized pseudo data. The authors report that on four class-incremental graph benchmarks, MAEM reduces recency bias, outperforms strong baselines, and keeps capacity bounded without storing historical graph data.
In briefMinimal neural network width linked to immersions of real projective space
SummaryThe paper studies the smallest width of deep neural networks that can approximate every continuous function from R^n to R^m under uniform convergence on compact sets. It shows that if real projective space RP^(n-1) admits a smooth immersion into R^n, and the activation is continuous and monotone, the minimal width is at least n+min{n,m}. As a result, ReLU networks have minimal width n+m for n equal to 2, 3, 4, 7 or 8 and 1≤m≤n, extending earlier work by Hanin-Sellke and Hwang.
In briefSystematic test of reinforcement learning settings for refining chip layouts after legalization
SummaryThis paper studies whether graph neural networks combined with reinforcement learning can refine chip placement after legalization. The authors test 18 configurations of a GNN-PPO framework across three ISPD2015 benchmarks, varying reward functions, network architectures, action spaces, candidate-set sizes, and training strategies. They analyse the results to identify pitfalls and propose a design classification approach to guide future RL model design decisions.
Foteini Oikonomou, Maria Pantazi-Kypraiou, George Goudroumanis et al.
In briefAlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks
SummaryThis paper introduces AlphaDiana, a measurement system for evaluating AI agents on verifiable reasoning tasks, with modular interfaces for datasets, agents, environments, and scorers. The authors test three open-source models and three open-source harnesses on five answer-based benchmarks, with two models also tested on two interactive benchmarks. They find that harnesses do not consistently improve reasoning accuracy, and that skill guidance and retained history do not reliably help. The authors argue that behavior and output-token use must be assessed together with correctness.
In briefA study of stability and convergence in flow map training
SummaryFlow maps are a probabilistic generation method that amortizes diffusion trajectories into one-step samplers. The paper analyzes the functional gradient flow induced by the MeanFlow and Lagrangian losses and shows their convergence to the true flow map. It identifies possibly unbounded updates as a practical obstacle, and reports that adaptive normalization with a specific weighting yields bounded updates and convergence to a neighborhood of the solution.
In briefMultilingual Safety and Reasoning Evaluation via Reasoning-Guided Translation
SummaryThis paper presents a framework that uses reasoning-guided translation to build multilingual versions of benchmarks while preserving each task's structure and meaning. It separates translation strategy selection from large-scale generation and adds task-aware validation, targeted repair, and independent quality screening. The authors built an evaluation collection covering seven safety, conversational, and math reasoning resources in 11 languages. Tests on open-source LLMs show language-dependent weaknesses, including inconsistent refusals and notable reasoning degradation in lower-resource languages.
In briefGlance Forcing: Making bidirectional video diffusion models generate autoregressively
SummaryThis paper introduces Glance Forcing, a lightweight distillation framework that converts a bidirectional diffusion model into a causal few-step student for real-time autoregressive video generation. It splits the denoising trajectory into slow and fast experts, each trained with a stage-specific objective, while the teacher stays frozen. Using only one PF-ODE trajectory, the method raises the VBench Total score from 84.04 to 84.22 and reduces trajectory-collection cost from thousands of GPU-hours to about one A100 GPU-hour.
In briefMixed Poisson diffusion models for posterior sampling: generative models using counting-process noise, applied to Poisson inverse problems.
SummaryThe paper introduces Mixed Poisson Diffusion Models (MPDM), which replace Gaussian noise with noising dynamics driven by multivariate counting processes. Motivated by Poisson inverse problems, the authors derive several posterior sampling algorithms from this generative framework. Evaluations on realistic Poisson inverse problems indicate the approach is well suited to this problem class.
Thibaut Pellerin, Alain Durmus, Elinor Berger et al.
In briefContinual pretraining of a prostate MRI ViT encoder to better align with tasks
SummaryThis paper studies whether pretraining and downstream objectives are aligned in MRI foundation models. The authors continually pretrain a pretrained ViT encoder with LoRA on a prostate segmentation task and use Centered Kernel Alignment (CKA) to analyze layer-wise representational shifts. The method improves performance on 3 of 6 downstream tasks and gives comparable results on the others. For segmentation, the encoder's features resemble those of the fine-tuned encoder more closely than the original pretrained encoder does, which the authors link to the observed gains.
Bilal A. Sidiqi, Yipei Wang, Shaheer U. Saeed et al.
In briefSingle-pass scores let medical AI models classify quickly with adjustable thresholds
SummaryMedical multimodal language models answer yes-or-no clinical questions with a fixed, implicit decision threshold. The study derives a continuous score from one forward pass by renormalizing the yes and no token probabilities, so the threshold can be adjusted without retraining. Tested with MedGemma and Gemma-3 (4B and 27B) on six medical imaging datasets, normalization gave the highest AUROC on four of six datasets, and a single decision took 53-79 ms on one GPU.
Patrick Wienholt, Frederik Hauke, Christiane Kuhl et al.
In briefRobust counterfactual reasoning for Markov decision processes under statistical and causal uncertainty
SummaryThis paper studies counterfactual reasoning in Markov decision processes (MDPs), asking what would have happened had different decisions been made, to identify causes of failure and explain AI decision-making. It handles two uncertainties at once: statistical uncertainty from estimating MDP transitions from trajectory data, and causal uncertainty because several structural causal models fit the same MDP. The authors give closed-form bounds on counterfactual probabilities valid across all such models, and compute minimal robust interventions via pessimistic value iteration on the induced counterfactual interval MDP. Experiments on several case studies demonstrate robustness and practical utility.
In briefHGRP: A Heterogeneous Graph Neural Network for Predicting Routability
SummaryThis paper proposes HGRP, a heterogeneous graph neural network that predicts routability, allowing placement quality to be evaluated without performing routing. It models cells, nets, and routing grids as distinct node types and builds a graph that preserves multi-level circuit information, then uses heterogeneous message passing to learn how circuit structure relates to routing behavior. On the CircuitNet dataset, the authors report an 18.69% relative F1-score improvement over the state of the art in congestion prediction.
In briefMeasuring Conventions in Humans and AI Agents Playing Hanabi
SummaryThis paper studies how humans and AI agents develop implicit conventions in the cooperative card game Hanabi. It defines a "convention gap" as the margin by which successful plays exceed the failure probability predicted from received hints and visible cards. Humans show a larger gap than all agent levels, with their advantage concentrated in hints that point to cards other than those touched. The authors also find that convention use depends on what players know about a card and peaks in the mid-game.
WhyProposes a computable convention gap metric for Hanabi, comparing humans and agent levels, with causal interventions providing solid evidence.
In briefBehavior Cloning from Few Demonstrations for Bird's-Eye View Autonomous Driving in Simulated Cities
SummaryThis paper studies behavior cloning, a supervised learning method that imitates expert driving, for autonomous driving from a bird's-eye view in simulated urban environments. The authors add a kernel density estimator that reweights training samples by action density, improving learning of rare but critical actions such as stopping at red lights, especially when expert demonstrations are scarce. Experiments in the CARLA simulator show the weighted method achieves higher route completion than standard behavior cloning.
Eric Aislan Antonelo, Gustavo Claudio Karl Couto, Christian M\"oller et al.
In briefSimulation study of how LLMs reach consensus under competing incentives
SummaryThe paper proposes a Network Consensus game in which agents must agree on one of two colours, with different payoffs for each agent. It compares six open-source models and one closed model (ChatGPT 5.6-sol) under varying communication access and network structures. Open-source LLMs reach consensus less often than humans, and only when messages spread across the global network. ChatGPT 5.6-sol reaches consensus even with neighbour-only or no communication, while highly clustered networks reduce consensus overall.
In briefWhat makes social agent interactions effective? Learning from verifiable outcomes in paired scenarios
SummaryThe paper introduces HICO-Bench, a benchmark of 180 everyday coordination scenarios where two agents with private information converse and then record the plan they agreed on. A program checks whether the records match and satisfy both private conditions, and an LLM judge checks whether the conversation supports the plan; passing both counts as Verified Success. Across thirteen LLMs, 36% of conversations rated highly for goal completion did not achieve Verified Success. The authors also propose Paired-Credit Policy Optimization (PaCPO), which raises Qwen3-8B's Verified Success from 20.4% to 28.3%.
In briefSWE-Stream: Scaling Executable Training Data for Software Engineering Agents via Progressive Construction
SummaryThis paper presents SWE-Stream, a system that builds executable software engineering tasks for training agents from repository development histories. It constructed 80,874 validated tasks and 118,688 successful trajectories across 10,728 repositories. Fine-tuning Qwen3-30B-A3B-Thinking-2507 on these trajectories yielded resolve rates of 69.0%, 49.3%, and 45.42% on SWE-bench Verified, Multilingual, and Pro.
In briefHEM-Med: A Layered Memory Framework for Safer Medication Recommendations Across Hospitals
SummaryThis paper presents HEM-Med, a system that recommends medications across multiple hospitals. It separates clinical experience that transfers between hospitals from prescribing habits specific to each hospital, and it accounts for safety limits such as drug-drug interactions and local formulary availability. Experiments on the real-world multi-center eICU dataset report that HEM-Med outperforms representative transfer learning, LLM-based, and medication recommendation baselines in accuracy and safety.
In briefValue-Guided Repulsion Q-Learning (VRQL) for offline RL: adds action values to particle interactions in drift-based policies.
SummaryThis paper addresses drifting policies in offline reinforcement learning, which set repulsion between generated actions using only geometric relationships and so cannot tell actions apart by expected value. The authors propose Value-Guided Repulsion Q-Learning (VRQL), which combines a geometric kernel distance with normalized Q-value scaling, giving lower-value particles stronger repulsion while keeping attraction toward behavior actions. The authors report highly competitive performance against established baselines on 50 OGBench task configurations and 6 D4RL AntMaze configurations.
In briefDimension-optimal dynamic regret for exp-concave online learners, with sharp lower bounds
SummaryThis paper studies online learning against changing comparators under exp-concave losses. It improves the curved dimension factor from d^(2/3) to d^(1/3) and provides a matching lower bound. The proposed learner adapts to the unknown path length and runs in near-linear time in T.
WhyResolves optimal dimension dependence for exp-concave dynamic regret, improving prior d^{2/3} bounds with a matching lower bound.
In briefWhen Skills Derived from Occupations Don't Help: Limits of Skill-Augmented Career Path Prediction
SummaryThis paper examines whether explicit skill representations add complementary signal to text-based job representations for career path prediction. The authors test several ways of incorporating skills, including score fusion and neural architectures, using both predicted skills and ground-truth ESCO annotations. Across two benchmark datasets, none of these settings substantially outperform strong text-only baselines, and gains with ground-truth skills are marginal and inconsistent. The authors attribute this to redundancy between text and skill representations, limited discriminative power of the ESCO taxonomy, and representational collapse in skill embeddings.
In briefIn-context neurofeedback tests whether LLMs can control their own activations
SummaryThe paper studies whether language models can deliberately control the activation-based scores that are used to detect their misbehavior. Ten open-weight models from five families were given in-context neurofeedback tasks, with some moving their own scores but no more than a comparable writer model did. Three models showed small but consistent control effects, and when asked to fool an activation monitor, no model learned to lower its score while one could raise it.
Davide Baldelli, Robyn Kozierok, Mirko Bronzi et al.
In briefTreeSpark: Training draft trees that a target model verifies in speculative decoding
SummaryTreeSpark is a training and inference framework that extends the semi-autoregressive DSpark architecture for speculative decoding. It builds parent-conditioned candidate trees from one backbone pass and replays their expansion, pruning, and packing during training, using two objectives that reuse reference-chain target distributions without extra target forward passes. Evaluated with greedy decoding on Qwen3-4B and Qwen3-8B across nine general benchmarks and six writing domains, it improves mean acceptance length over DSpark by 23.4% to 31.6%. It also raises single-request throughput by 15.1% to 25.0%.
In briefRM-OPD: On-Policy Distillation for Building Generalist Generative Reward Models
SummaryThis paper addresses two problems in training generative reward models with reinforcement learning on mixed-domain preference data: sparse outcome rewards and interference between domains. It proposes RM-OPD, a two-stage framework that first trains domain-specialist teacher models with single-domain GRPO fine-tuning, then consolidates them into one student model using on-policy distillation, domain contribution reweighting, and optional criteria context distillation. The authors report that across multiple reward model benchmarks and two base models, RM-OPD consistently outperforms vanilla Mix-RL and other strong baselines.
In briefUsing Generation-Side Representations to Improve Fine-Grained Visual Understanding in Unified Multimodal Models
SummaryThis paper studies whether generation-side visual representations in unified multimodal models can improve fine-grained visual understanding. The authors find these representations are more sensitive to subtle semantic changes and depend on correct spatial organization. They propose GAFU, a post-training framework that compresses generation-side representations into compact G tokens for task-conditioned retrieval by the understanding pathway. Experiments show GAFU outperforms SFT baselines trained on the same data, with maximum relative gains of 20.00%, 17.94%, and 8.62% on VStar-Bench, OmniSpatial, and VSR-Bench.
In briefMultiPartyTurnBench: A Benchmark for Turn-Taking Control in Multi-Speaker Voice Dialogue
SummaryThis paper introduces MultiPartyTurnBench (MPTB), a benchmark of 240 synthetic conversations with two to five speakers, testing whether speech systems can tell real turn ends and interruptions from non-events. Learned models detect turn ends with 82.5-92.6% recall, but interruption detection suffers from missed takeovers or false activations. Five full-duplex agents perform near chance on selective interruption handling, showing that multi-party floor control remains an unsolved challenge.
In briefOptimizing sparse Boolean rule networks with mixed integer linear programming: a method for learning compact, interpretable rule-based classifiers
SummaryThis paper studies learning accurate, simple binary classifiers that can be read directly as Boolean rules. The authors use mixed integer linear programming to train sparse disjunctive normal form networks with a fixed structure, balancing misclassification cost against sparsity through a regularized objective. An enhanced formulation reduces integer variables and redundant constraints, and experiments on synthetic and UCI datasets report better efficiency and competitive or superior accuracy against established interpretable baselines.
In briefGap Detection: Reducing Sameness in AI Answers by Searching for Uncovered but Relevant Directions
SummaryLanguage models often give similar answers to open-ended questions even when many alternatives exist. This paper proposes Gap Detection, a framework that searches a shared text embedding space for directions that are relevant to the query yet not covered by existing responses, scored by a product of demand and defensibility. Reported mean relative reductions in output collision are about 12.8% and 10.8% on curated topics and INFINITY-CHAT queries. In a synthetic-data augmentation setting, collision fell from 0.324 to 0.298 and the Vendi Score rose from 27.16 to 29.20.
In briefOneLA: A shared latent action space for controlling world models across domains
SummaryOneLA is a framework that learns continuous latent actions from videos using a shared orthogonal dictionary and a common predictive objective, without explicit action labels or cross-domain alignment losses. Lightweight adapters map each domain's native control commands into this shared space, and a world model predicts context-dependent transitions from the latent actions and current visual state. The paper reports a 24.8% reduction in action cycle-transfer rotation error over the strongest evaluated baseline, and a 20.5% improvement in HPSv3 for real-scene generation in a joint-training comparison.
In briefDataset-Level Regularization for Joint-Embedding Self-Supervised Learning
SummaryThis paper proposes out-of-batch regularization, which keeps a dataset-level state in CPU memory and uses globally built per-instance targets during minibatch training. On natural image and biomedical datasets, it approximates the full-dataset regularization gradient more faithfully than minibatch estimation at small batch sizes. Applied to SimCLR and SIGReg, it improves image classification, minority-class performance, instance retrieval, and continued pretraining, with the largest gains at small batch sizes.
In briefLLM-Based Checking of Loan Applications Against Bank Lending Policies
SummaryThis paper presents a prototype system that uses a large language model to automatically extract compliance-relevant rules from internal bank lending policy documents and convert them into machine-readable form. The extracted rules are then used to evaluate loan applications for policy exceptions, non-compliant findings, missing information, and cases needing manual review. On 11 real loan applications with known exceptions, the system detected most of them and flagged additional issues missed in manual reviews. The authors note that the approach produces false positives and cannot fully replace human judgement.
In briefIs Global News Coverage Driven by Dependence, Imitation, or Herding? A Simulation Study
SummaryThe study compares 18 national outlets in six countries with four core outlets on coverage of the Ukraine and Gaza wars and two UN climate conferences, measuring attention, framing and publication timing. It then uses language models acting as editors to test dependence, imitation and herding mechanisms in a simulation. Only one of nine tests held: outlets relying more on wire copy published sooner after the core outlets. Herding was supported and imitation was not, and the simulation could not distinguish the mechanisms because the dependence rule reduced to publishing wire copy.
In briefAttention-Based Spatio-Temporal Method for Detecting Fire and Smoke in Video
SummaryThis paper proposes a vision-based method that combines spatial and temporal features to detect smoke and fire in a scene. Infrared thermal and smoke sensors cannot be used outdoors, which motivates the vision approach. The method has three main parts, each designed to reduce the error of the previous one, and uses transformer modules to improve accuracy. The authors report proper performance and suggest it can help increase workplace safety.
Alireza Mirrashid, Mohammad Khoshbin, Ali Atghaei et al.
In briefSparse Autoencoders for Interpreting NLP Models: A Critical Survey
SummaryThis paper is a critical survey of 138 studies on sparse autoencoders (SAEs) applied to natural language processing. SAEs decompose language model activations into sparse, potentially interpretable features. The authors propose a task-grounded taxonomy, examine factors affecting feature recoverability, and assess evidence for semantic interpretability, mechanistic faithfulness, and behavioral utility. They reconcile disagreements among published results and suggest directions for more rigorous evaluation.
Salima Lamsiyah, Adelaide Danilov, Oleksandr Marchenko Breneur et al.
In briefBatch black-box optimization using flow matching to propose evaluation batches without refitting per task
SummaryThis paper addresses batch black-box optimization for expensive systems that can evaluate many candidates in parallel but allow only a few rounds. The proposed method, FMTR, learns offline a continuous, history-conditioned proposal distribution from which variable-size batches are sampled. On synthetic and real-world benchmarks, FMTR reportedly outperforms batch Bayesian optimization and in-context baselines while substantially reducing batch-selection computational cost.
Quoc-Anh Hoang Nguyen, Xinyu Zhang, Daolang Huang et al.
In briefStreamErase: Ultra-Fast Real-Time Video Object Erasing
SummaryStreamErase is a method for removing objects from videos in a streaming manner, reaching sub-second latency and up to 26 FPS on a single GPU. It first uses a text-to-video model with mask input to produce high-quality erased videos, then transfers this capability to an efficient unidirectional streaming eraser. The approach uses sliding-window causal attention with attention sinks and KV cache rolling during training to enable constant-speed generation of arbitrarily long videos. The authors also introduce LongErase, a benchmark for long-video object removal.
In briefEnergy Use of Vision-Language Models on Edge Devices
SummaryThis paper measures total-system energy, latency, and thermal behavior of the Qwen2-VL-2B-Instruct vision-language model and its text-only version on a Raspberry Pi 5 with a Hailo-10H NPU and on a Jetson Orin Nano Super. It separates each request's fixed energy from the energy that grows with output length. NPU offloading cuts energy per generated token by about two-thirds and avoids CPU thermal throttling, while at typical response lengths generating the answer costs far more than processing the image. The two platforms' energy curves cross near 40 output tokens.
In briefFrom Events to Reasoning: Hierarchical Rule Composition over Explicit States (explicit event-graph reasoning)
SummaryThe paper presents the Event Reasoning System (ERS), which stores reasoning state as an explicit event graph. Declared rule and task conditions determine which candidate operations may execute, so candidates violating period, unit, or evidence constraints are excluded from accepted paths. Recursively composable control rules organize subgoals under a budget, and the authors report results on structured-event graphs and fourteen development tasks, including dependency chain length and inference time.
In briefBlenderWorld: A benchmark testing AI agents on realistic 3D creation from real-world references
SummaryThis paper introduces BlenderWorld, a benchmark for evaluating multimodal AI agents on realistic 3D creation tasks in Blender. It contains 223 tasks across architecture, products, natural environments, and scientific visualization, with inputs in text, image, video, and structured formats. Submitted projects are rendered from multiple viewpoints and scored against 2,651 binary rubrics plus holistic visual quality. Results across thirteen models show that current systems parse broad instructions reasonably well but struggle with fine-grained 3D spatial coherence and accurate structural reproduction.
In briefDreamHMR: Adapting Video Generation Priors for Full Body and Hand Motion Recovery
SummaryThe paper presents DreamHMR, a framework that adapts video generation pretraining to recover human body and hand motion from videos, including under occlusion. It first trains the video backbone to predict complete human renderings so it learns human geometry, then refines body and hand details using local evidence. On 3DPW, EMDB-1, and RICH, the best model reduces body MPJPE by 4.2%, 4.7%, and 6.1%. On the occlusion-heavy BEDLAM and EgoBody sequences, it reduces overall body and hand MPJPE by 7.0% and 10.0%.
In briefA new way to view neural operators: their layers as solutions to Bregman-regularized optimization problems.
SummaryThe paper views the action of neural operator layers as minimizers of Bregman regularized optimization problems over Banach function spaces. In this framework, activation operators can be interpreted as Bregman proximity operators mapping from dual to primal space. The authors introduce Bregman neural operators, which include an inverse activation operator and have the same expressivity as standard neural operators. Numerical experiments indicate that the Bregman variant of Fourier neural operators helps train deeper and more accurate models.
Abdel-Rahim Mezidi, Jordan Patracone, Saverio Salzo et al.
In briefKANLab: A Browser-Based TensorFlow Playground for Kolmogorov-Arnold Networks
SummaryKANLab is a browser-based visualization tool, forked from TensorFlow Playground, for interactively teaching the core concepts of Kolmogorov-Arnold Networks (KAN). To keep it clear for learners, it omits residual connections and learnable scaling parameters, so the displayed B-spline activation curves map directly to their trainable control points. Users can inspect and modify activation functions and initialization schemes, and watch gradient descent correct manual disruptions. It also supports exploring activation variance, interactive pruning, and retraining in a compact 2D setting. The tool is open source and available at https://kanlab.app/.
In briefMaking Context Compaction More Transparent Improves Lightweight Coding Agents' Retention of Facts and Commitments
SummaryLong coding agent sessions exceed the model's context window, so older history must be compacted, typically into a model-written summary that the agent cannot inspect for losses. This paper presents transparent-compact, a deterministic method for the open-source Pi coding agent that leaves user and assistant prose unchanged and explicitly flags where information was stripped. Benchmarks on several models show it matched or outperformed a default summary method and a similar deterministic method on commitment and fact-recall tasks, especially in avoiding confident false assertions. The advantage was smaller with a frontier model.
In briefACR-Net: Reducing Text Over-Reliance in Speech Emotion Recognition via Contrastive Decoupling
SummaryThis paper finds that speech emotion recognition models and audio LLMs fail when vocal tone conflicts with the meaning of the words, because training on consistent data makes them overly reliant on text. The authors introduce the ASPIRE benchmark of adversarial audio-text pairs across four conflict types, along with two metrics, SOP and LDD. Their proposed ACR-Net uses cross-modal attention and a contrastive decoupling loss to separate conflicting representations, and it reports higher acoustic accuracy in both adversarial and standard settings.
In briefFrom Spectral to Frobenius: When Muon's Spectral Bias Breaks
SummaryThis paper studies the implicit bias of the Muon optimizer on linearly separable classification problems. It proves that practical Muon, which uses finite Newton-Schulz iterations and a positive stabilization constant, converges to the Frobenius max-margin direction rather than the spectral one. The authors analyze how iteration count, stabilization constant, and polynomial choice affect behavior, and show that spectral bias can be recovered when Newton-Schulz iterations grow logarithmically during training.
WhyProves finite Newton-Schulz iterations and stabilization shift Muon toward Frobenius max-margin, with rigorous analysis of several asymptotic regimes.
In briefRepForcing: Representation-Space Initialization for Few-Step Autoregressive Video Generation
SummaryThis paper proposes Representation Forcing, a method for distilling pretrained video diffusion models into few-step autoregressive generators. Instead of regressing onto an intermediate autoregressive teacher's trajectories, it trains the causal initialization using normalized features from a frozen video diffusion transformer. The authors report that after distribution matching distillation, their four-step model raises VBench Total from 84.04 to 84.66 and VisionReward from 6.326 to 10.074 compared with a causal-ODE baseline.
In briefExhaustive Dataset of Canonical Formulas for Symbolic Regression Training
SummaryThe paper addresses neural symbolic regression training data, which relies on random sampling that yields redundant structures and incomplete function coverage. The authors generate an exhaustive dataset of univariate canonical skeletons using a deterministic constrained context-free grammar, paired with a sampling strategy that avoids numerical instabilities. The dataset is designed to support decomposable symbolic regression for multivariate problems. A 150-problem benchmark across three tiers is also introduced, on which current state-of-the-art methods struggle significantly.
In briefWhy Generate a Click? Plug-in decision heads that choose web agent actions without generating text
SummaryWeb agents built on vision-language models usually output text such as click(12) to choose each action, even though most of that text is just a choice among a few action types and marked elements. This paper reads these choices directly from the generator's hidden states using plug-in decision heads on a frozen generator. A confidence cascade executes the heads' action when they are confident and otherwise lets the generator decode. Reported results show task success on WebArena is maintained or slightly improved while using less than half of the generated tokens.
In briefWhat Foundation Models for Market Microstructure Learn: A Controlled Study with Synthetic Order Books
SummaryThe paper trains GPT-style transformers on order flow from simulators with known latent parameters, so Bayes-optimal predictions and generating states serve as references. Next-token pretraining on order flow is found to approximate amortized Bayesian inference over market regimes, with generalization driven by regime diversity rather than token count. On real NASDAQ data, zero-shot transfer depends on how well the simulator prior covers real order flow.
WhyNovel controlled design comparing autoregressive models to Bayes-optimal and oracle references with known ground truth; well-supported experiments with clear value for the subfield.
In briefHow Language Exposure Affects Multilingual Pre-training on Limited Data
SummaryThis paper studies how training data is split across English, Dutch, and Chinese, and in what order it is presented, when data is limited. It compares balanced exposure with exposure that favors one language, and mixed training with curriculum training. Mixed training outperforms curriculum training in all 15 matched configurations, and some language-dominant mixed setups score higher than the balanced one. Dutch-primary configurations achieve the best overall performance.
In briefTask-State Governance: A Virtual Machine for Recoverable GUI Agent Execution
SummaryThis paper proposes task-state governance, an interaction paradigm that represents a task as an editable contract linking desired outcomes, facts to preserve, and observed application state. The authors implement it as TaskVM, which turns live state into editable task controls, runs user edits as real interface actions, verifies outcomes through independent observation, and labels each action as reversible, compensatable, or irreversible. Across an automated benchmark, a user study, and a live-model exploration, the paradigm improved contract compliance and users' recovery control and trust while execution outcomes stayed the same.
In briefEvalBuilder: Can LLMs Build Cheaper, Reliable Benchmarks?
SummaryThis paper asks whether LLM agents can build lower-cost evaluation benchmarks that reach the same conclusions as expensive reference benchmarks. It formalizes automatic benchmark construction as a meta-evaluation problem and introduces EvalBuilder, which measures fidelity, efficiency, and robustness of generated benchmarks. The framework also includes an iterative environment where agents refine their benchmarks within a fixed time budget.
In briefRTAformer: A transformer that tracks changing relationships between objects in videos
SummaryThis paper addresses video scene graph generation, which captures objects and how their relationships change over time in videos. It identifies a failure mode called relational semantic inertia, where models keep old predictions after a relation has changed. The proposed RTAformer models entity persistence and relation changes separately, using differences between adjacent-frame relation queries and cross-frame instance alignment during training. Experiments on Action Genome report higher recall and mean recall, with larger gains on relation-transition frames than on non-transition frames.
In briefWhen Pairwise Preferences Cycle: A Reliability-Anchored Preference Walk for Video GRPO
SummaryThe paper studies pairwise preference feedback used to train video generation models with reinforcement learning. Auditing 6,976 on-policy comparison graphs, it finds cyclic preferences are common and persist during training, while average-win-rate reward aggregation can yield unreliable advantage estimates. It proposes a graph-aware method, the Reliability-Anchored Preference Walk, which uses reliable relational evidence from cyclic preference graphs to strengthen win-rate signals. Under matched conditions, it needs no extra evaluator queries and reportedly outperforms Pairwise-GRPO in controlled comparisons, transfer settings, and blind human evaluations.
In briefAFCS: Atomic-Fact-Conditioned Synthesis for Factual Knowledge Pretraining
SummaryAFCS is a synthetic data pipeline that first extracts source-supported atomic facts from documents, then uses each retained fact to drive text generation for language-model pretraining. In continued pretraining of OLMo-3-7B and a 32B base model, it outperforms WRAP and Nemotron-CC on SimpleQA, SimpleQA Verified, PopQA, and HotpotQA. Ablations show fact-level synthesis beats chunk-, paragraph-, and entity-level synthesis.
In briefText interface design to improve LLM reasoning on audio-visual question answering
SummaryThis paper studies how automatic prompt optimization (APO) applies to audio-visual question answering, depending on the textual interface used. It compares three designs: a single omni-modal LLM for direct inference, a captioner-reasoner system, and an agentic architecture where a primary agent coordinates subagents. Experiments with Qwen3-Omni and Gemini show that direct optimization gains depend on the model, while the other two designs are more interpretable but harder to optimize. The authors conclude that textual interface design is an important factor for effective prompt optimization in omni-modal reasoning.
In briefVisualWorldBench: A Fine-Grained Benchmark for Testing Visual World Knowledge in Multimodal AI Models
SummaryThis paper introduces VisualWorldBench, a benchmark that tests how well multimodal large language models know visual entities, built on a taxonomy of 21,291 leaf-node classes across 8 visual domains. It uses four tasks: closed-set recognition, contrastive selection, precise localization, and open-world naming, with most questions drawn from taxonomically similar categories. The strongest closed-source model reached 80.4% average accuracy. The authors also release VisualWorldBench-50K, a training resource, and report that fine-tuning on it improves benchmark scores by over 21 points with transfer to other fine-grained benchmarks.
In briefLarge-dimensional analysis of least-squares federated learning via generalized total variation minimization
SummaryThis paper studies federated learning over decentralized networks, where collaboration suppresses local finite-sample noise but introduces bias between heterogeneous nodes. Using random matrix theory, the authors analyze the global solution of generalized total variation minimization (GTVMin) with least-squares models in large dimensions. They derive exact analytic expressions for asymptotic performance under i.i.d. and non-i.i.d. Gaussian mixture data on a fixed network. Experiments on synthetic and real-world data confirm the theoretical results.
In briefFeasibility of a Patient-Specific Digital Twin for Shunt Catheter Path Planning
SummaryThis study built a patient-specific digital twin that reconstructs skull, scalp, ear and ventricles from preoperative CT and searches around Keen's point for the best catheter trajectory, returned as ear-based ruler measurements and two depths. In 11 adults, freehand insertion following the plan achieved Hayhurst Grade 1 placement in 10 cases, versus 22 of 213 historically, with no intraoperative complications. The gain came from choosing entry and depth per patient, and whether it reduces revision remains to be tested in the next stage.
Kritsanavis Chongsrid, Vich Yindeedej, Supaporn Konmun et al.
In briefGDBA-Net: A Gated Dual-Branch SAM2 Adaptation for Cross-Domain Nuclei Segmentation
SummaryThis paper presents GDBA-Net, a parameter-efficient framework for segmenting cell nuclei in pathology images. It adapts a frozen SAM2 encoder using a Gated Dual-Branch Adapter that combines pretrained semantic features with frequency-derived structural priors through adaptive channel gating. With only 4.3 million trainable parameters, it reports state-of-the-art in-dataset results and better cross-dataset generalization across six benchmarks. The authors attribute these gains to the semantic-frequency design and gating rather than added model capacity.
Biplab Poudel, Betserai Albert Magarire, Anish Kumar Shah et al.
In briefRethinking Prediction Horizons: Separating Model Queries from State Updates in Few-Step MeanFlow
SummaryThis paper studies an inference-time control for few-step MeanFlow image generation, separating the time interval the model is queried over from the step size used for state updates. With the model and grid held fixed, varying the prediction horizon affects FID differently depending on query position and the incoming states. A frozen four-query rule reduces mean CIFAR-10 FID without retraining or extra queries, but the authors report that it is not universally preferred over alternatives.
In briefTrainable deep, exactly solvable small: how a polynomial activation's loss surface guides network design, initialization and training
SummaryThe paper proposes a network built around a polynomial activation with an unnormalized identity path. As a deep stack it trains to one hundred layers on MNIST without normalization machinery. As a small model its square loss is an exactly solvable polynomial system. The authors argue the small-scale analysis prescribes how to design, initialize and train the large stack.
In briefWhich Knowledge Should LLMs Trust? A Steering Direction That Generalizes Across Knowledge Conflicts
SummaryThe paper studies how large language models choose between their internal parametric knowledge and external contextual knowledge when the two conflict. The authors find this choice is linearly represented and can be controlled along a single direction in the model's residual stream. The direction generalizes across factual-conflict settings and transfers to retrieval-augmented generation, memory, and rule-following tasks, where steering improves use of prompt-provided information.
In briefFoveaVLA: Replacing Wrist Cameras with Learned Gaze Actions in Robot Policies
SummaryThis paper presents FoveaVLA, a vision-language-action model that replaces the wrist camera with learned foveation actions. At each step, it predicts the interaction region to observe next and feeds crops from original-resolution images alongside full external views to the policy. Built on π0.5 in single and dual configurations and tested on four real-world task suites, FoveaVLA achieves 70.8% and 77.1% success, compared with 52.1% for the wrist-camera π0.5 baseline.
In briefBeyond Correctness: Training Reasoning Models to Use Extra Test-Time Compute Better
SummaryThis paper asks whether reasoning models can be post-trained to use test-time compute more effectively, and which training examples help most. It defines compute-responsive examples as problems that switch from failure to success when more reasoning is allowed, and proposes Compute-Responsive Data Selection (CRDS) to prioritize them. On Qwen3-8B, training on these examples recovers more held-out low-compute failures than training on already-solved examples, and the benefit holds across several inference budgets.
WhyProposes a new principle selecting training examples by marginal utility of extra reasoning, validated across models, data sources, and compute budgets.
In briefGBTSVR: Granular-Ball-Based Twin Support Vector Regression
SummaryThis paper proposes granular ball twin support vector regression (GBTSVR) and its quantile extension GB-TSVQR, which replace individual samples with compact granular balls as basic learning units to reduce sensitivity to noise and outliers. The models build nonparallel lower- and upper-bound regression functions over these balls, and GB-TSVQR adds quantile regression to capture heterogeneous trends and asymmetric errors. According to the authors, both models outperform state-of-the-art regression methods on synthetic, time-series, and outlier-contaminated datasets.
In briefWhat Is a Clarification? Estimating Model-Dependent Input Ambiguity
SummaryThe paper distinguishes world ambiguity, meaning all interpretations an input admits, from model ambiguity, meaning those a responding model can access. It formalizes natural-language clarifications through uniqueness, faithfulness, and answer equivalence. The authors introduce CircleClar, a black-box method that estimates a model's available answers by checking consistency along an answer–clarification–answer cycle. It recovers these answer sets well on open-domain QA and executable Text-to-SQL.
In briefFACES: A fine-grained gated method for aligning meaning in multimodal emotional support chat
SummaryThis paper presents FACES, a framework for Multimodal Emotional Support Conversation (MESC) that combines multi-source rewards into semantic and functional alignment signals. It uses a non-compensatory gated fusion mechanism to reduce reward offset across dimensions and structured Chain-of-Thought anchors for token-level credit assignment. Under a controlled response-length protocol, FACES raised its average MLLM-based score from 3.61 to 4.04, won 80.22% of pairwise comparisons against SOTA, and earned a 75.3% human preference rate. Results on three external datasets indicate transferability.
In briefStaged physics-informed neural network for estimating power converter component parameters
SummaryThe paper proposes a staged physics-informed neural network (PINN) that identifies the inductance, capacitance and parasitic resistances of a DC buck converter from time-domain data. Gradients are released in a sensitivity-derived order of L, C, rL and rC. The five-stage schedule recovers L and C within 5% and the series ESR within 10%, and a three-stage variant reaches the same accuracy 11% faster.
Samuela Rokocakau, Panagiotis Kakosimos, Giulia Tresca et al.
In briefSWE-Nexus: Automatically building multi-part coding tasks from test runs
SummaryThis paper presents SWE-Nexus, an automated pipeline that builds repository-level implementation tasks from test execution rather than from historical code changes. It traces the production function bodies each selected test exercises, keeps only scopes where masking those bodies causes a test failure, and deterministically composes compatible scopes into larger tasks. The benchmark contains 600 tasks from 105 repositories across six languages, and the best of 12 models under three scaffolds resolves 37.8% of tasks in one attempt.
In briefWeighted Adjoint Matching: a time-weighted way to finetune flow matching models
SummaryThis paper proposes Weighted Adjoint Matching, a method for finetuning pretrained diffusion or flow matching models using a differentiable reward. It extends Adjoint Matching by adding a time-dependent weight to the control cost, which is high early in generation to preserve coarse-feature diversity and low later to allow refinement of details. The authors prove the method is equivalent to reward maximization under time-varying KL regularization on generative paths. Experiments on Gaussian mixtures and MNIST show a better reward-mode balance tradeoff than Adjoint Matching.
