In briefJev Thinks “I Don't Know” But Doesn't Say It: A Dataset for Testing Probability Calibration
SummaryThe paper addresses the unverified claim that Jev, a System One model returning probability distributions, produces calibrated probabilities. It introduces Sys1Cal-v1, a dataset of True/False propositions with known probabilities, and evaluates Jev and the open-source baseline SemIf using total variation distance. The authors find that Jev's Choice answers appear to omit a third option, "I don't know"; recovering this term raises the median soft accuracy of Choice answers from 0.771 to 0.978.
In briefGenerative Uncertainty as a Self-supervised Signal for Semantic Similarity Learning: uses output variance from text-to-video models to pick stable video features without human labels
SummaryThe paper proposes a self-supervised method that uses generative uncertainty from text-to-video (T2V) diffusion models to learn video semantic similarity without human annotations. T2V models give consistent outputs for familiar concepts but high variance for specialized ones, and the authors use this to learn a mask over pretrained embeddings such as VideoMAE and V-JEPA. The mask keeps features that stay consistent across generations and discards those tied to generative noise. Experiments on three key tasks show the learned feature subspaces outperform the original pretrained features and baseline feature selection methods.
Enrico Pallotta, Sina Raoufi, Lars Doorenbos et al.
In briefAnswerPool: Making Multiple-Choice Benchmarks Hard Again by Pooling Answers
SummaryThe paper proposes AnswerPool, which merges the options of N questions sharing a context into one list and asks a model to assign each question its answer, without writing new items. This lowers the chance of guessing correctly by elimination. Tested on eight text, image, and video benchmarks with eighteen models, pooling made the task harder for every model, and the weakest models gained the most from elimination.
Mohamed Eltahir, Abobaker Ahmed, Nawaf Barebood et al.
In briefTMCS: Tool-Grounded Multi-Agent AI for Multi-Step Chemistry Problem Solving
SummaryThis paper addresses how large language models struggle with combinatorial chemistry problems that require constrained molecular editing, candidate validation, and revision after failures. The authors propose TMCS, a multi-agent framework that treats chemical problem solving as an interpretable workflow using external tools, few-shot trajectory memory, and structured reflection. TMCS links generation, understanding, editing, description, and optimization into a closed loop. The authors report that it consistently improves chemical reasoning across open- and closed-source base models and achieves state-of-the-art results on multiple chemical tasks.
In briefUsing LLMs to Automatically Identify Machine Learning Task Types Across Domains
SummaryThis paper examines whether large language models can infer the data domain and prediction task from dataset-level information when users supply only the target feature. The authors also release an annotated benchmark of 625 public tabular and time series datasets. On tabular data, the LLM-based approach outperforms AutoGluon (0.98 vs. 0.93 F1 macro), the best cross-domain model reaches 0.90, and smaller locally deployable models reach 0.75.
Petros Tsialis, Steffen Limmer, Tobias Rodemann et al.
In briefScalable in-context reinforcement learning via recurrent memory compression
SummaryThis paper proposes Recurrent Algorithm Distillation (RAD), a method that lets Transformers learn reinforcement learning tasks in context without explicit weight updates. RAD uses a Compression Transformer to distill long interaction histories into compact latent tokens, and an AD Transformer that generates actions from these compressed memories plus recent transitions. The authors report that across diverse environments, RAD matches the asymptotic performance of standard Algorithm Distillation while using significantly smaller context windows.
In briefReverse Sequential PAV Voting Rule: Fairness and Approximation Guarantees
SummaryThis paper studies the Reverse Sequential Proportional Approval Voting rule (RevSeqPAV) for approval-based committee elections. It analyzes the rule's proportional representation, measured by Extended Justified Representation and its approximations, and its approximation of the maximum PAV score. The authors first prove strong negative results for general, unrestricted instances, then identify settings where the rule offers meaningful fairness and optimization guarantees.
In briefA framework where AI agents automatically evolve and improve learnable optimization algorithms.
SummaryThis paper proposes a dual-agent framework that treats the design of Meta-Black-Box Optimization (MetaBBO) methods as a coding task. A task agent evolves the codebase of a target MetaBBO approach, while a hyper agent modifies both the task agent and itself to enable open-ended design. Starting from a naive template, the framework finds novel variants that outperform human-made baselines and adapts quickly across optimization domains.
In briefHow weighting schedules control what score-based generative models learn from multimodal data, and when
SummaryThis paper studies how score-based generative models learn multimodal distributions during training. The authors split the time-integrated loss into single-time contributions and analyze each at a fixed signal-to-noise ratio, finding that this ratio sets how fast mode directions and relative weights are learned. At high signal-to-noise ratio, all mode directions are learned together on one timescale while relative weights are not learned. Near the speciation time, where the ratio is of order one, every feature becomes learnable on its own timescale. The results rest on an exact high-dimensional analysis of unbalanced and hierarchical Gaussian mixtures and are checked on image and human genome haplotype generation.
WhyDecomposes the integrated loss by time and links signal-to-noise ratio to learning timescales, with exact high-dimensional analysis and validation on image and genome data.
J\'er\'emie Klinger, Rapha\"el Urfin, Giulio Biroli et al.
In briefTeacher-Student Gaps Alone Miss What Helps: Outcome-Guided On-Policy Distillation for Agents
SummaryThe paper studies on-policy distillation (OPD) for multi-turn autonomous agents, finding that teacher-student token distribution gaps do not reliably indicate when teacher guidance helps. Large gaps can be harmless, while small gaps can be critical to final outcomes. The authors propose Outcome-Guided On-Policy Distillation (OG-OPD), which weights teacher supervision using final task outcomes from paired student continuations. Across ALFWorld, ScienceWorld, and WebShop, it improves task success rates by 3.6 to 17.7 points over vanilla OPD and by up to 7.0 points over the strongest baseline.
In briefReliability engineering methods for keeping AI systems dependable in use
SummaryThis paper examines how to judge whether an AI system reliably performs its intended function over a stated period and under stated conditions. It adapts established reliability engineering methods, such as failure definitions, FMEA, accelerated testing, and field monitoring, to AI systems. A four-level diagnostic framework classifies failures as component, operational-loop, agentic-conduct, or network and governance failures. Three cases illustrate the approach, and the authors conclude that new measurements and safety guardrails are still needed as these systems self-evolve.
In briefCollaborative principle evolution via evidence transfer for parallel scientific discovery
SummaryThe paper introduces COEVOLVE, which lets parallel principle-evolution branches collaborate by sharing measurements while keeping their principle posteriors separate, aiming to speed up LLM-based scientific hypothesis exploration. Across six scientific-discovery tasks under a matched budget, it reaches a mean solution quality of 66.5% versus 57.0% for single-branch evolution, with a 1.80x mean wall-clock speedup on the GPT-5.6-Terra backbone. The authors also study when evidence sharing helps and when transfer safeguards are needed to limit negative transfer.
In briefWhen to Correct Satellite Evapotranspiration Estimates: Stress-Testing Neural Corrections
SummaryThis study evaluates whether neural residual corrections can improve OpenET satellite evapotranspiration estimates, using 16,366 flux-tower observations from 151 stations. It tests whether selectors can predict when a correction helps and reject unsupported inputs. In one held-out station, the Gain method accepted all 32 physically invalid records, increasing mean absolute error by 21.6 mm/day. The authors describe these fault analyses as exploratory, and none of 40 preplanned temporal comparisons passed Holm correction.
In briefMemoReason: Testing How Memorized Facts Affect LLM Contextual Reasoning
SummaryThis paper examines whether large language models' strong reasoning benchmark results reflect genuine contextual reasoning or reliance on facts stored in their parameters. The authors introduce MemoReason, a benchmark that replaces real entities with fictitious ones of the same type while keeping task structure unchanged. Recent models show performance drops of up to 15.7% in the fictitious setting, but they rarely give the corresponding factual answer, suggesting parametric memory influences reasoning in ways more complex than simple recall.
Zineddine Tighidet, Andrea Mogini, Jiali Mei et al.
In briefRoGSW4RLD: A method that turns multi-camera robot world-model predictions into a single queryable 3D-over-time scene
SummaryThis paper presents RoGSW4RLD, a feed-forward framework that lifts synchronized multi-camera rollouts from robot world models into a unified, time-queryable metric 4D Gaussian field. It uses a two-stage design: the first stage fuses cross-view evidence with robot articulated geometry and kinematics, and the second refines geometry and appearance while preserving initial temporal displacements. On 256 held-out DROID episodes, it outperforms camera-wise reconstruction with calibrated merging, with gains in novel-view PSNR, depth AbsRel, and robot displacement error, and the gains extend to Cosmos 3 rollouts.
In briefHow AI Models Differ in Accepting False Facts Planted Earlier in a Chat
SummaryThe paper studies session-level contamination, where false premises inserted in earlier conversation turns are later adopted as fact by large language models. The authors define five protocols ordered by source authority and test three models over 22,500 turns, finding that the models follow structurally different policies, with adoption rising sharply under system-injected authority and instruction override. The evaluation framework and taxonomy are released for reproducible testing.
WhyProposes authority-gradient contamination protocols and compares three models' epistemic policies over 22,500 turns with human-validated judging.
Fahrell Giovanny, Geby Bayuningtyas, Sahrul Mukharom et al.
In briefPIVOT: Pivot-Aware On-Policy Self-Distillation for Training Multi-Turn Vision-Language Agents
SummaryThis paper studies reinforcement learning for multi-turn vision-language model agents. The authors find that gains from on-policy self-distillation come mostly from physically rolling back the state at the pivot step, the first unrecoverable action, which is impractical in real environments. They propose PIVOT, which internalizes pivot localization and state restoration into token-level parameter updates, removing the need for environment rollbacks during training and extra hints at test time. Across five multi-turn benchmarks, PIVOT reaches 0.90 overall accuracy on Qwen2.5-VL-3B and 0.92 on Qwen3-VL-2B.
In briefNarrowing the Horizon: Measuring Topic Shifts in AI-Generated Viewpoints
SummaryThis paper introduces a method for measuring how the prominence of topics shifts across large language model families during post-training. Using climate change discourse as a case study, it shows how homogenisation affects the representation of diverse solutions across models. The authors also test interventions and find that specialised models can help preserve a broader range of perspectives.
In briefTraining-Free Clinical Reasoning via Medical Ontologies and Symbolic-Probabilistic Knowledge Graphs
SummaryThe paper proposes CKG Reasoner, a training-free diagnostic framework that maps patient observations to explicit medical knowledge using evidence nodes, an information gate, and knowledge-weighted evidence accumulation. It was evaluated on six retrospective cohorts for dengue, malaria, and influenza, with clustering and scoring that do not use outcome labels. Positive-class F1 scores ranged from 0.634 to 0.996. The authors state the results reflect evidence separation and auditability, not prospective clinical validity or superiority over comparison methods.
In briefAbGaze: Attention-based geometric learning for end-to-end antibody design
SummaryThis paper presents AbGaze, an end-to-end antibody design framework built on attentive geometric representation learning. It encodes the distance, direction, and surface-normal orientation of antigen surfaces relative to antibody residue local frames, aggregating these interactions by interfacial context. The shared representation is used for multi-CDR co-design, complex structure prediction, and affinity optimization, and the authors report gains over prior methods on all three tasks.
In briefSimulation-based inference to recover plate reverb parameters from impulse responses
SummaryThis paper tackles Task A of the 1st DAFx Parameter Estimation Challenge, which asks for the physical parameters of a plate reverb model from an impulse response. The authors use Simulation-Based Inference, training a neural network on simulator-generated data to estimate a distribution over plate parameters given an impulse response. Inference on a new impulse response needs only a forward pass through the network, without the simulator. For each test observation, the network is fine-tuned using extra simulation rounds that sample parameters from the current estimated distribution.
In briefBeyond Saying Less: Fine-Grained Alignment for Faithful Vision-Language Models
SummaryLarge vision-language models often hallucinate objects not present in images. The paper finds that on-policy reinforcement learning can reduce hallucinations simply by making fewer valid claims. The authors propose a fine-grained framework with the DOPA dataset, which densely annotates object presence and absence, and SCAPO, which assigns credit at the subsentence level. Experiments show more informative, faithful descriptions in generation and clear gains on discriminative evaluation.
In briefDecide, Don't Generate: Competitive Dimensional ABSA with Jev's Typed Decisions
SummaryThis paper shows that aspect-based sentiment analysis can be competitive without text generation. A frozen model, Jev, answers typed questions with rubric scores, label probabilities, and yes/no judgments, which are aligned to the annotation scheme using 488 coefficients fitted on CPU. On valence-arousal regression across ten corpora in six languages it achieves the lowest aggregate error of any participating system, and on triplet and quadruplet extraction it outperforms fine-tuned Llama-3.3-70B and GPT-OSS-120B baselines.
In briefScaffold Then Internalize: Injecting Encoder Features into Diffusion Transformers
SummaryThis paper proposes REPI, a training framework that injects pretrained visual encoder representations into a diffusion transformer as a temporary scaffold, which the model then progressively internalizes. It reverses the direction of the existing REPA method, which projects diffusion features into the encoder space, and the two approaches are complementary. The authors report that REPI outperforms REPA across many backbones and that combining them yields further gains. They state that REPI + REPA with 160K steps matches vanilla SiT trained for 7M steps.
In briefPaper on how narrow multimodal fine-tuning can cause emergent misalignment in vision-language AI models
SummaryThis paper studies whether fine-tuning vision-language models on narrow multimodal tasks, such as vulnerable code, careless household-object use, and conspiratorial readings of ordinary scenes, can cause emergent misalignment. Across fifteen commercial and open-source models, the authors find that such fine-tuning induces broadly misaligned behavior that transfers to unrelated tasks. They also test mitigation strategies that partially reduce this effect.
WhyFirst extension of emergent misalignment to vision-language models, tested across fifteen models with transfer and mitigation experiments; broad impact.
In briefEvoIn: Fine-tuning agents to internalize evolved decision procedures
SummaryThis paper proposes EvoIn, a framework that evolves and validates new decision-making procedures from agent execution traces, then rewrites improved reasoning traces into self-contained form and fine-tunes the model on them. The improved decisions persist at inference time without the evolved harness. The authors report a 10.9-point pass-rate gain in-domain and gains out-of-domain, with the learned procedures generalizing to unseen tasks and yielding consistent improvements on another model family.
In briefLocality-constrained diffusion guidance for zero-shot image enhancement across domains (LocDiff)
SummaryThis paper introduces LocDiff, a locality-constrained guidance method that works as a zero-shot extension to pre-trained diffusion models for image enhancement. It keeps critical local features from the source domain during domain adaptation while leaving less important regions unconstrained. The method is tested on art-to-photo translation and on enhancing low-quality fetal ultrasound renderings. The authors report that it achieves better realism-faithfulness trade-offs than state-of-the-art methods.
Theresa Neubauer, Dimitrios Lenis, Astrid Berg et al.
In briefλ-JEPA: Spectral Anti-Collapse Regularization for Self-Supervised Learning
SummaryThis paper studies representation collapse in joint-embedding self-supervised learning, finding that anti-collapse objectives applied after a projection head may not prevent dimensional collapse in the backbone. The authors propose SACReg, a spectral regularizer motivated by λ-balance analysis, and apply it to JEPA as λ-JEPA. According to the paper, λ-JEPA outperforms LeJEPA and VISReg on ImageNet-1k classification and average linear-probe transfer across eight image datasets, and improves over LeVJEPA and V-JEPA 2 on video benchmarks.
WhyApplies a lambda-balance spectral regularizer to JEPA, with linear analysis and comparisons on ImageNet and video benchmarks against LeJEPA and V-JEPA 2.
Berker Demirel, Cl\'ementine Domin\'e, Valentino Maiorca et al.
In briefThe Argument and the Letterhead: Source–Position Coherence in AI Evaluation
SummaryThis paper asks whether AI evaluators separate an argument's quality from the identity of its source. Across preregistered studies and a supplement, 2,976 usable evaluations were collected for six fixed texts on US AI policy, Germany's debt brake and Swiss nuclear energy, each shown under several source attributions. The gap between sources changed with the argument, which the authors interpret as plausible support for source-position coherence, alongside competing accounts such as credibility, authenticity and task interpretation.
In briefTextual User Taste: Natural-language profiles of listeners for recommender systems
SummaryThe paper presents Textual User Taste, a system that builds structured natural-language taste profiles from listening behavior, interaction signals, content metadata, and optional user feedback, deployed to millions of Spotify users. It describes the production lifecycle for generating, evaluating, and maintaining these profiles at scale. Combined with behavioral embeddings, the profiles improve MRR by 0.6% for future-track prediction and NDCG@7 by 2.2% for search ranking. The authors also note limitations, including difficulties with negation and short-term temporal adaptation.
In briefMeasuring Collapse and Correction in Same-Model LLM Debate
SummaryThe paper studies multi-agent debates among identical large language models on multiple-choice questions, arguing that answer changes are not always improvements. It introduces an auditable ledger that records collapses, corrections, and signed utility across 6,925 MMLU-Pro debates, finding 253 collapses. Replay experiments show that a freeze policy preventing collapses also loses many corrections, so collapse prevention alone can recommend the wrong policy.
In briefEvaluating hierarchy-aware deep learning for recognizing Tironian notes
SummaryThis paper tests whether structural relationships among Tironian shorthand signs can help automatic recognition. The authors compare flat classifiers (ResNet18, ConvNeXt, Swin, ViT) with hierarchical coarse-to-fine and routing models, using the SNT sign set and samples from handwritten and Vergilius Turonensis manuscript domains, with and without few-shot adaptation. Hierarchical HD-CNN gives the best non-adapted Top-1 result of 45.43%, while flat classification is best after few-shot adaptation at 82.09%.
Yule Kang, Thomas Gorges, Janne van der Loop et al.
In briefMulti-Attractor GNNs: Graph networks that can reach several valid solutions
SummaryThe paper addresses recurrent and equilibrium graph neural networks, which usually converge to a single fixed point or are trained toward one target per graph, even though many problems have multiple valid solutions. The authors show that with multiple equilibria, one weight-tied message-passing network can represent set-valued, permutation-equivariant maps, since different initializations approach different valid solutions. Experiments on Ising ground states, protein graph module detection, and chemical reaction steady states show the learned updates produce multiple high-quality predictions with better average solution quality than the tested baselines.
In briefWhen Words Speak Louder than Images: Understanding Language Bias in Vision-Language Models
SummaryVision-language models often favor linguistic cues over visual evidence, leading to incorrect predictions. Building on a word completion task, the paper proposes a diagnostic framework that splits inference into four interdependent stages to trace how language bias propagates. It examines how linguistic priors and cross-modal coverage evolve across these stages. The authors reveal the interplay between these two factors as the underlying mechanism of language bias.
In briefeval-unlearn: A Unified Benchmark for Concept Unlearning in Text-to-Image Diffusion Models
SummaryConcept unlearning methods for text-to-image diffusion models are evaluated under inconsistent conditions, which makes fair comparison difficult. The paper presents eval-unlearn, an open-source Python library that integrates twelve published techniques and nine evaluation metrics in a reproducible benchmark framework. The authors also release a HuggingFace leaderboard and interactive tool, comparing all twelve techniques in a nudity concept erasure case study.
In briefSpikeCredit: A Temporal Credit Carrier Method for Reinforcement Learning with Sparse Rewards
SummaryThis paper studies how reinforcement learning can assign credit for delayed outcomes when rewards are sparse. It proposes Temporal Credit Carriers, a role that spiking neural networks can fulfill through graded membrane traces and event-driven spikes, and builds the SpikeCredit framework around this idea. On sparse-reward MuJoCo tasks, the authors report large gains in Last10 return over sparse SNN baselines, and on Swimmer the method exceeds a dense-reward baseline by 113%.
In briefWavePP: High-Throughput Pipeline-Parallel LLM Prefill with Prefix Reuse
SummaryThis paper presents WavePP, a prefill runtime built on TensorRT-LLM that overlaps request admission with pipeline execution to reduce coordination overhead when prefixes are reused across pipeline stages. It finds reusable prefixes asynchronously, protects cached state, and dynamically plans chunk sizes to keep the pipeline full. In tested settings on GLM 5.2 and MiniMax M2.7, it improved prefill throughput in 37 of 40 configurations, with gains up to 2.91x at high cache reuse and concurrency 128.
In briefRubric-Aware On-Policy Self-Distillation for Personalizing LLM Responses
SummaryThis paper proposes GRASP, a framework for personalizing large language model responses using user-specific rubrics. It turns rubric aspects into token-level supervision by aligning a rubric-free student model with a rubric-informed teacher on the student's own generated outputs. A Rubric-based Teacher Validation step keeps only cases where the teacher covers the target aspects well. The authors report state-of-the-art results on the LaMP-QA personalized question answering benchmark across multiple backbones.
In briefImprint Reader: Turning Model Weight Updates into Plain-Language Descriptions and Behavioral Edits
SummaryThis paper introduces Imprint Reader, a model trained with SMaRT to describe frozen weight updates in natural language, so a model can explain what it has learned. On held-out updates, judge-based Pass@100 is 2% for knowledge and 16% for behavior, showing feasibility while indicating reliability across updates remains the next step. The Reader also serves as a differentiable proxy that guides intervention through MetaEdit, with reported gains in harmful-prompt refusal and BFCL Overall score.
In briefTrade-offs between latency and accuracy in spiking neural networks
SummaryThis paper studies the latency of spiking neural networks (SNNs) for speech command recognition, arguing that overlapping computation across layers can make them faster than comparable quantized networks. The authors propose Pipeline Delay Search to balance per-layer accuracy gains against added latency, and combine it with quantization-aware training and threshold tuning in a framework called Falcon. On GSCV2 and SSC, Falcon reports accuracies of 96.31 and 83.02 at modeled latencies of 119.64 and 124.00 microseconds.
In briefOn-policy or off-policy learning? A systematic study of how distillation behaves.
SummaryThis paper studies how the rollout policy affects strong-to-weak model distillation, varying rollout policy, token-level KL direction, and learning rate in a controlled setup across Llama3 and Qwen2.5 on scientific, medical, and arithmetic reasoning tasks. It finds that token-level KL direction mainly shapes task performance and output coverage, while learning rate governs forgetting and update sparsity. Forward KL is robust to the rollout policy, whereas reverse KL is more sensitive and favours student-generated rollouts. On-policy data improves generalisation to harder Countdown variants, but this advantage does not reliably persist after subsequent RLVR.
Julianna Piskorz, Antonin Berthon, Mihaela van der Schaar
In briefAIM-ZO: Activation-guided subspace maintenance for memory-efficient LLM fine-tuning without backpropagation
SummaryThis paper proposes AIM-ZO, a zeroth-order method for fine-tuning large language models that estimates updates using only forward evaluations, avoiding backpropagation and activation storage. It uses forward activations as local directional information and continuously integrates them into a broad, evolving subspace, while activating only a small set of shared and sampled directions per step. Across 5 LLMs and 11 downstream tasks under matched forward-evaluation budgets, its six-task average exceeds the strongest fully evaluated ZO baseline by 1.26 points on OPT-2.7B and MeZO by 2.85 points on OPT-30B.
WhyAdds activation-informed subspace maintenance to zeroth-order fine-tuning, a substantive improvement over prior low-dimensional subspace methods, with broad evaluation across five LLMs and eleven task
In briefToward Reliable AI Data Scientists: Data Agents with Workflow Harnesses
SummaryThis survey examines data agents, AI systems that automate data science work, through a workflow-harness lens. It organizes the literature into five stages (perception, planning, execution, verification, and repair) and analyzes the technical approaches within each. The paper also identifies four open reliability problems and summarizes task families, application settings, and evaluation benchmarks.
In briefSCBO: Batching and ordering survey questions for more efficient LLM-based social surveys
SummaryThis paper proposes SCBO, a training-free framework that lets large language models predict several survey questions in one prompt instead of one at a time. It extracts semantic representations of survey items, groups related questions into batches with a shared reference bank, and orders target questions from easy to hard. Experiments on four large-scale survey datasets and four LLMs report substantially lower token consumption and inference time, with generally improved accuracy over a non-batched baseline.
In briefSpatial Grafting: Adding Metric 3D Geometry to Robot Action Policies
SummaryPretrained robot policies such as vision-language-action and world-action models do not explicitly encode metric geometry relative to the robot. The paper proposes Spatial Grafting, a lightweight module that binds frozen 3D reconstruction features to robot-relative geometry and injects them into a flow-matching action expert via cross-attention, leaving the host's perception pathway unchanged. On the RoboTwin 2.0 dual-arm benchmark, the graft improves every tested host, and grafted π0.5 reaches 94.0% (clean) and 92.4% (randomized), above the published WAM4D baseline.
Dingsheng Liu, Yangzheng Wu, Mahboubeh Asadi et al.
In briefAnswerMap: Faithful Spatial Explanations for Vision-Language Models
SummaryThe paper introduces AnswerMap, a training-free, task-agnostic, black-box method that explains which image regions a vision-language model uses to answer a query. The image is split into row and column bands, each judged for relevance by the frozen model, and their yes-probabilities are combined into a spatial map. In tests across four models and three query types, the map matched the model's own pointing (AUC 0.85 versus 0.38 for attention), and deleting its region flipped 53% of correct answers, compared with 19% for attention.
Mohamed Eltahir, Fardows Adam, Duaa M. Tahir et al.
In briefBregman Consensus: an iterative method for agents to agree on shared parameters
SummaryThis paper studies how a group of agents can reach consensus on a set of parameters. The agents use the same Bregman-type divergence to measure disagreement among their estimates but trust one another to different degrees. Each agent revises its estimate toward the weighted barycenter of all individual estimates, giving more weight to more trusted agents. The authors show that these revisions form an iterative algorithm that converges to a unique consensus estimate.
In briefDrawingsDreamer: A Unified Model for Generating Multi-View Engineering Drawings
SummaryThis paper presents DrawingsDreamer, a unified framework driven by a large language model that generates multi-view vector engineering drawings. It treats drawing generation as a pure sequence modeling task, so no raster image encoder is needed, and uses a streamlined representation with hierarchical postfix tokenization. A progressive task-aware curriculum lets the model move from local structural repair to full drawing generation. According to the authors, the model performs well on geometric fidelity and syntactic accuracy across diverse conditional and unconditional tasks.
In briefSystematic map of CT/LDCT lung-cancer AI studies: tasks, evidence chains, and gaps
SummaryThis paper systematically reviews artificial-intelligence studies using CT or low-dose CT for lung cancer, covering 293 studies from 2016 to 2026. Most address detection (230), with 8 on future risk prediction and 55 in other categories. External validation, calibration, and decision-curve analysis were less common than clinical variables, 3D CT, and radiomics. Only 31 studies (10.6%) completed the full six-tier substantive evidence chain, with the largest losses at the reasoning/explanation stage.
In briefLong-Horizon Scaling: How Model Capabilities Shape Returns to Computation
SummaryThis paper studies how existing capabilities of long-horizon agents affect the gains from extended interaction and task feedback, using the AutoLab and EdgeBench benchmarks. It finds that early performance and later growth relate to different capabilities and models these trends with category-specific logistic power laws. The authors also derive a policy for deciding whether to continue a run, which in replay experiments saved about one-third of full-run time with relative score losses of 2.4% on AutoLab and 3.3% on EdgeBench.
In briefEP-Mem: Privacy-Aware Memory for LLM Agents That Respects Social Relationships
SummaryThis paper proposes EP-Mem, a memory architecture that lets users control what information an AI agent may share across different social roles. The authors also build EP-Bench, a long-term multi-party benchmark for policy-based relational disclosure. Reported experiments show 94.0% privacy classification accuracy, disclosure-permission accuracy rising from 22% to 68%, and a 75.6% reduction in privacy leakage.
In briefBeyond Selection: Token Parameterization for Extreme Visual Token Compression
SummaryThis paper addresses visual-token compression for vision-language models under extreme budgets, where pruning can harm visual grounding and learned resamplers add cost. It frames compression through token parameterization, defining compressibility and learnability as two coupled objectives. The authors propose Braco, a lightweight four-step coder, which reports a favorable accuracy-efficiency trade-off at 23x to 64x compression and remains competitive at 144x.
In briefToken-Disentangled Latent Test-Time Scaling for Vision-Language Reasoning: Refining Hidden States by Token Role
SummaryThis paper proposes an inference-time method that optimizes a short hidden-state prefix of an initially generated reasoning trajectory. Visual feedback is routed to image-sensitive tokens and reasoning feedback to high-entropy tokens, while remaining tokens are held near their anchors by a regularizer. On Qwen2.5-VL-7B and InternVL3.5-8B, the authors report macro-accuracy gains over CoT of +2.57 and +1.51 across perception and reasoning benchmarks. Code is publicly available.
In briefGenerative AI data augmentation for oral lesion classification, with the PhotoMOCI dataset and benchmark
SummaryThis paper introduces PhotoMOCI, a new photographic dataset for oral cancer diagnostic tasks, and benchmarks how different data augmentation methods affect image classifiers. It compares traditional augmentation with generative approaches such as GANs and diffusion models. The authors also propose the Synthetic Image Filter (SIF), which uses two auxiliary models to keep synthetic images that look realistic and represent the target class. SIF-filtered augmentation outperformed traditional augmentation in all reported cases, improving ResNet50 and ViT accuracy on PhotoMOCI by 1.73% and 2.35%.
Marco Parola, Mario G.C.A. Cimino, Sabrina Senatore
In briefSignFLIP: A unified model that translates and generates sign language through staged alignment.
SummaryThis paper proposes SignFLIP, a unified framework centered on a large language model that handles both sign language translation and generation. It uses a symmetric architecture and a stage-wise training strategy built on large-scale data, first pre-aligning the shared sign-text representation and then refining it for translation. Experiments on multiple benchmarks report competitive results against task-specific models on both tasks, along with strong transfer to sign language recognition.
In briefFine-tuning LLMs by updating only the layers that matter most for a task
SummaryThis paper studies how large language models reason over cross-linguistic text and proposes that their layers divide work into conceptualization, reasoning, and textualization. It uses sensitivity analysis to identify the most critical functional stage for a given task. The authors introduce Layer-Informed Fine-Tuning (LIFT), which updates only these critical layers, and report that it speeds up training and improves performance.
In briefPretraining Temporal Heterogeneous Graphs for Relational Deep Learning
SummaryThis paper studies how to handle time in relational deep learning, which models database rows and foreign-key links as heterogeneous graphs. It combines Multi-scale Time Encoding for record age with Rotary Time Encoding for signed intervals between records, and uses three self-supervised objectives in a two-stage pretraining scheme. On five RelBench datasets across 11 tasks, the best staged schedules improve over supervised training with the same encodings by 3.02% and 1.06% on two backbones, and over controls without pretraining or either encoding by 3.24% and 2.37%.
In briefCADENCE: A Confidence-Adaptive Dual-Expert Network for Fast and Accurate Time Series Classification
SummaryThe paper presents CADENCE, a CPU-native dual-expert architecture for time series classification that pairs a convolutional linear expert with a distributional interval expert, and uses a validation meta-router to choose between them or blend them by confidence. Across 109 equal-length UCR Archive datasets with 30 resamples each, it achieves a grand mean accuracy of 0.8864, ranking second behind HIVE-COTE 2.0 (0.8895), with no statistically significant difference. It averages about 17.53 seconds per dataset on a dual-core CPU.
In briefAttributing multivariate forecast uncertainty to input features via entropy-Shapley games
SummaryThis paper addresses how to attribute the predictive uncertainty of multivariate probabilistic models to input features while accounting for dependence between output components. It introduces a hierarchy of three entropy-based Shapley games, ranging from per-component marginal entropy to fully joint entropy, which isolates a cross-component term capturing how features shift dependence. The authors give a chain-rule decomposition and characterize the cross-component term via conditional total correlation, with closed-form and sample-based estimators. The framework is demonstrated on differences in learned joint structure across models, from distributional regression to a zero-shot time series foundation model...
Niklas Koenen, Claudia Battistin, Jeriek Van den Abeele et al.
In briefASCT: Attentive Search over Counterfactual Trees for Agentic RL Credit Assignment
SummaryThis paper proposes ASCT, a framework that uses training-time multi-step search to give local credit to individual decisions within an agentic workflow, rather than relying only on the final outcome. On HotpotQA agentic retrieval-augmented generation, three instantiations (Uniform, UCT, and cost-aware AgentUCT) all improve mean held-out utility over trajectory-return PPO and workflow-adapted VinePPO. Across three seeds, ASCT-AgentUCT reaches 0.6187 utility versus 0.5939 for VinePPO.
In briefUncertainty quantification for personalising heart models using ultrafast ultrasound
SummaryThe paper uses ultrafast ultrasound shear wave elastography to track myocardial stiffness over the cardiac cycle and uses these measurements to personalise a cardiac model's parameters. It applies simulation-based inference with neural posterior estimation to obtain posterior distributions over stiffness scale, contraction rate, and relaxation rate. Of six healthy volunteers, four passed prior-support diagnostics and were analysed, with curve-level RMSE dropping from 12.61 ± 5.55 kPa to 1.14 ± 0.38 kPa for posterior predictive medians compared with prior predictive medians.
In briefAdversarial Consistency-Guided Representation Learning for Multi-view Clustering: a method that reduces view-identifying information in shared representations to improve cross-view clustering consistency.
SummaryThis paper proposes ACGRL, a framework for multi-view clustering. The authors note that shared representations learned for cross-view consistency may still retain information that identifies each view, which can compromise consistent cross-view clustering structures. ACGRL uses a gradient-reversal view discriminator to reduce view identifiability and obtain invariant reference representations, which are then frozen to separate view-specific information from common information. Experiments on four benchmark datasets reportedly show better clustering performance than representative multi-view clustering methods.
In briefUnderstanding On-Policy Distillation via Sparse Crosscoders
SummaryThis paper uses sparse crosscoders to study what on-policy distillation (OPD) changes inside a student language model's internal representations. The authors propose a swap readout that measures how training changes the student's use of each shared feature. They find that OPD does not create new features or transfer the teacher's own features, but reweights existing shared ones, and that SFT warm-up before OPD works mainly the same way.
In briefRevealing model identity splits cooperative LLM agent groups into factions
SummaryThe paper studies multi-agent LLM systems in which each agent is told the model family of its peers. When agents know these labels, they form factions and prefer interacting with agents sharing their label, even though the task does not call for this. Across two cooperative games and a reasoning benchmark, success fell from 96% to 81%, while labeled groups used about 30% more rounds and 55% more tokens. Withholding the identity labels removes the effect and is presented as a simple mitigation.
Xavier Del Giudice, Alessio Palma, Matteo Migliarini et al.
In briefNucShare: a lightweight framework sharing one decoder for nuclei segmentation
SummaryNucShare proposes a single U-Net-style decoder shared across three nuclei segmentation tasks, each with only a lightweight head, using 37% of the parameters of three independent decoders. A teacher model uses a UNI2-h encoder (690.5M parameters), while the student NucShare-S uses a ConvNeXt V2-Tiny encoder (34.7M parameters) and learns via output-level distillation. On PanNuke, the student keeps 99.1% of the teacher's mPQ (0.519), above CFR-SAM-H's 0.511, and runs 27.5 times faster per tile than the teacher.
In briefFrom Normative Frameworks to Alignment Data: Building and Testing Training Data for AI Values
SummaryThis paper describes how seven experts turned Islamic ethical, theological, and jurisprudential principles into Arabic-English training data for language models. The data includes about 2.8K supervised fine-tuning examples and 5.4K preference pairs. In blind evaluation, a model trained on the curated fine-tuning data was preferred over the baseline, while adding preference data produced a smaller, non-significant difference. Standard benchmarks showed no broad drop in general capabilities.
Husrev Taha Sencar, Rezart Beka, Danish Naeem et al.
In briefReCAT: Structured Memory Helps Robots Remember, Count, and Track Time
SummaryThe paper presents ReCAT, a language-conditioned robot policy with structured recurrent memory that lets robots use information no longer visible to their sensors, such as earlier visual cues, event counts, and elapsed time. It reports 95.3% average success on LIBERO and 62.4% on RMBench. On three real-robot tasks, the best variant averages 66.7% success, compared with 8.3% for the strongest short-history baseline.
Pankhuri Vanjani, Mostafa Hatab, Can Mizrakli et al.
In briefAlignment Games: A Framework for Fixing Concept Mismatches in Human-AI Collaboration
SummaryThis paper introduces Alignment Games, a framework for making task-relevant differences in how people and AI understand a concept visible and repairable during interaction. The authors describe task-specific conceptual frames using attributes, values, relations, constraints, and priorities, drawing on situated conceptualization theory. They define alignment moves that act on the situation, the interpretive reasoning, or the resulting frame. Examples from educational content generation, creative coding, and argumentative writing show how these moves form repair sequences, and the paper derives design principles for runtime conceptual alignment.
In briefGAC-PINN: A Geometry-Adaptive, Constraint-Enhanced Physics-Informed Neural Network
SummaryThis paper proposes GAC-PINN, a physics-informed neural network for problems with steep gradients, sharp interfaces, or strong spatio-temporal coupling. It combines gradient-driven adaptive grid mapping, adaptive-bandwidth hard constraints, Gaussian Fourier feature mapping, and an operator-aware router. On the viscous Burgers, sharp-peaked 2D Poisson, and Allen-Cahn benchmarks, it reports relative L2 errors that consistently outperform the baselines.
In briefCarveMix-RC: Handling rare resection-cavity imbalance in brain metastasis segmentation with lesion-aware synthetic augmentation
SummaryThis paper addresses segmentation of post-treatment brain metastases, focusing on the rare resection cavity (RC) subregion. The authors build an nnU-Net-based framework with RC-weighted loss, anatomically consistent cavity augmentation, a residual encoder, and lesion-aware post-processing. On the BraTS-MET 2026 Task 1 online validation benchmark, the best ensemble reports an RC lesion-wise Dice of 0.575 and NSD of 0.474.
Md Shibly Sadique, Md Fayaz Bin Hossen, Michael L. Evans et al.
In briefConRAG: lightweight method for finding links between two known entities across documents
SummaryThis paper addresses how to find the intermediate bridge entities and evidence chains connecting two known entities across a document corpus, and to explain the connection using that evidence. The authors propose ConRAG, which builds a lightweight entity-document graph from entity co-occurrence and LLM-based entity filtering, then retrieves and semantically ranks paths between the two endpoints. On MuSiQue and 2WikiMultiHopQA, it improves bridge entity and reasoning chain recovery over strong RAG baselines while reducing graph-indexing token cost substantially.
Kilian B\"anziger, Sonia Laguna, Markus Kreft et al.
In briefG³-LoRA: Grouping reward-weighted video training data by gradient compatibility
SummaryThis paper studies post-training text-to-video models on reward-weighted data whose categories may conflict in their training updates. The proposed G³-LoRA method measures category-level gradients, removes the shared global direction, clusters categories by residual gradient compatibility, trains group-specific LoRA adapters, and merges them into one adapter using distillation. On Wan2.1-T2V-1.3B-Diffusers, the merged adapter improves VBench2.0 evaluation over the base model and several baselines, though camera motion and some local-quality dimensions remain challenging.
Jia Song (1), Wenhow Li (1), Lichen Bai (1) et al.
In briefBeneath the Tokens: GPU Performance Study of Multi-Token Prediction in LLM Inference
SummaryThis paper compares two-token multi-token prediction (MTP) with standard autoregressive decoding on a single NVIDIA A10G GPU, using 360 requests across plain-text, reasoning, and tool-calling workloads. MTP raised output throughput by 1.91 to 2.19 times and cut time to first output by 10.0 to 14.2 percent. It also reduced repeated CUDA Graph executions per generated token by 56.4 to 78.1 percent, despite a longer execution path.
In brief5W1H+Which: A Context-Valid Indexing Method with Progressive Ontology Binding
SummaryThe paper proposes 5W1H+Which, a design that separates source-grounded content extraction from progressive ontology binding. The 5W1H questions organize independently searchable content units, while Which links them to versioned ontology elements, mappings, applicability scopes, and validation states. The authors specify falsifiable hypotheses and baselines, but no experiments have been conducted and no performance gains are claimed.
Yaxiao Liu (PwC China AI Center), Pengbo Liu (PwC China AI Center), Yiwen Liu (PwC China AI Center) et al.
In briefA research agent platform that enforces research discipline through mechanisms (paper)
SummaryThis paper describes AfS (Agent for Science), a platform for long-horizon scientific projects that run for tens of hours across many agent runs. The authors argue that research discipline should be enforced mechanically so non-compliant states cannot occur, rather than relying on the model's behavior. The system covers node contracts, write-path gates, a two-tier memory, and the runtime, illustrated with three failure traces and two closed campaigns. No benchmark results are reported.
In briefSubgroup rank-1 lattice points for fast high-dimensional black-box integral estimation
SummaryThis paper proposes subgroup rank-1 lattice point sets for estimating integrals of high-dimensional black-box functions, such as kernel mean embeddings and the softmax kernel in self-attention. Restricting the Korobov generator to a fixed multiplicative order allows feature-map computations to be done exactly by FFT in O(n log m) time and O(n) memory. The authors prove a convergence rate for the squared worst-case error and report better results than random features and scrambled Sobol' and Halton points across synthetic and real-data settings.
WhyCombines subgroup rank-1 lattices with FFT to reduce feature-map integration cost, and proves an exact convergence threshold for worst-case error.
In briefProtoSeam: Improving Classifier Training with Latent Gaussian Mixture Models
SummaryThis paper proposes a lifted training method for supervised classification that inserts one learnable prototype per class at a single interface inside the network. Training combines a consensus penalty pulling features toward their class prototype with a classification loss on samples drawn around the prototypes, and the prototypes are discarded at inference. Experiments on CIFAR-10, CIFAR-100, and TinyImageNet with ResNet and vision transformer backbones are reported, along with theoretical justification.
In briefQAM: Merging Saved Training Checkpoints Accurately via Sequential Consistency
SummaryThis paper studies how accurately saved training checkpoints can reconstruct the endpoint of a sequential training run with different update strengths. The authors prove a lower bound showing that gradient-descent checkpoint histories alone cannot achieve better than o(h^3) endpoint error uniformly over smooth, strongly convex losses. They propose Quadratic-Accurate Merging (QAM), which attains a matching O(h^3) error bound. Experiments on two public Adam checkpoint trajectories and 15 tasks show mixed results for short windows and advantages over WSM for longer windows.
In briefEdgeCraft: Automated Model Building for Edge IoT Devices
SummaryThis paper presents EdgeCraft, an LLM-driven system that turns high-level intent into machine learning models deployable on edge IoT devices. It uses a constraint-aware synthesis tree to explore candidates guided by measured SLO gaps, and a multi-fidelity verifier to reduce the cost of target-device verification. Across 50 public tasks, it exceeds the task-specific reference in best-observed quality on 40 tasks and performs competitively on a self-collected SEN dataset.
In briefLearning to Re-Draft: A Variational Stackelberg Game for Discrete Diffusion
SummaryThis paper proposes Variational Stackelberg Discrete Diffusion (VSDD), a framework for learning which token corruptions best train a discrete diffusion model to revise its earlier outputs. A leader learns a corruption process parameterized by the denoiser's token embeddings and is rewarded by how much the denoiser improves after training on those corruptions, while a follower optimizes a variational denoising objective. The authors evaluate VSDD on molecular, text, and playlist generation, reporting substantially better molecular validity than uniform and masked diffusion, lower text perplexity than uniform diffusion, and competitiveness with masked diffusion.
Dmitrii Moor, Federico Tomasi, Paul N. Bennett et al.
In briefA simulation framework linking mass gatherings to citywide epidemic spread
SummaryThis paper presents a multiscale simulation model that links short-term close contacts at mass gatherings such as concerts, sports matches, and festivals to citywide commuting networks. Using Madrid, Spain as a case study, the authors compare several types of gatherings under different levels of disease transmissibility. They report that mass gatherings consistently increased outbreak size, sped up progression, advanced arrival times across districts, and synchronized spatial spread. The initial number of seeded infections at the event explained much of this acceleration, while post-event transmission conditions gave complementary information about invasion timing.
Jaeyoung Kwak, Michael H. Lees, Chin Chun Ooi et al.
In briefSmall transformers track Bayesian evidence for hidden common causes
SummaryThis paper examines whether small, tractable transformers can learn a Bayesian style of reasoning about latent common causes that generalizes across contexts. It separates the causal mechanisms inside the model from the causal structure of the true data-generating process, and moves toward natural language prediction. The authors test whether accumulated Bayesian evidence can be represented in ways that transfer to novel test cases.
In briefFONDANT: A Planner for Strong and Best-Effort Plans Using Antichains
SummaryThe paper presents FONDANT, a planner for fully observable nondeterministic planning that handles both strong planning and best-effort planning. It represents sets of states, such as winning regions, by their minimal elements and returns uniform policies along with a certificate for losing states. On a benchmark drawn from leading strong planners and a best-effort planner, it matches or exceeds their coverage and is faster on larger instances, though slower on small and medium ones.
In briefPEARL: An Elastic, Adaptive System for Faster Agentic Reinforcement Learning
SummaryThis paper addresses the high cost of multi-turn rollout in agentic reinforcement learning. It presents PEARL, an asynchronous system that coordinates external GPU elasticity, temporary reuse of idle training GPUs, and adaptive prefill-decode configuration. PEARL uses runtime profiles to predict completion time and switches configurations with minimal worker and role changes. The authors report 2.17 to 2.79 times the throughput of fixed-resource ROLL, and up to about 26.9% and 36.3% gains over RLBoost+ for Qwen3-8B and Qwen3-30B-A3B.
In briefTrajectory-based method to reduce forgetting in incremental learning
SummaryThis paper addresses catastrophic forgetting in continual learning, where models lose performance on earlier tasks after training on new data. It proposes TMLN, which treats continual learning as an optimal control problem over a curved loss landscape, using a diagonal empirical Fisher Information Matrix to define a local Riemannian manifold. The preconditioner is modulated by the normalized historical trajectory of parameter values, rather than relying on additive penalties. According to the paper, experiments on class- and domain-incremental benchmarks show a significant reduction in the loss barrier between consecutive tasks.
Isabelle Aguilar, Zayn Andre Zainal, Luis Fernando Herbozo Contreras et al.
In briefCoordinated lane-level speed limits and ramp metering for congested weaving sections using hybrid MPC and multi-agent reinforcement learning
SummaryThis paper addresses recurrent congestion and collisions on successive weaving segments of urban expressways. It proposes a hybrid controller that combines model predictive control, using a lane-level traffic flow model (L-METANET) for long-horizon optimization, with multi-agent reinforcement learning that produces short-horizon residual actions. Experiments on an 18-km expressway in Changchun, China, report faster convergence and better results on multiple metrics than MPC- and MARL-based baselines, and stronger generalization than pure MARL under fluctuating demand.
In briefA Wizard-of-Oz study of nonverbal behavior in Chinese-German intercultural interaction
SummaryThis paper presents a Wizard-of-Oz study design and real-time system for collecting multimodal behavioral data from native Chinese speakers reacting to social norm violations by German learners of Mandarin. The system uses a photorealistic avatar, a wizard control console, and synchronized multimodal logging. A layered annotation framework grounded in psychological theory supports training future cultural interpretation and behavior generation models. Four interaction scenarios developed with experts provide a foundation for a culturally adapted Chinese language learning agent.
Siddhant Jain, Anna Lea Reinwarth, Dimitra Tsovaltzi et al.
In briefFrom Migration to Calibration: Aligning AI Agents with Deployment Requirements
SummaryThe paper argues that after an agent is deployed, migrated, or scaled, it should be calibrated to the destination's requirements rather than imitate the source system. It proposes three standard families (capability and operational reachability, technical and professional precision, and user context) and four instruments that supply tunable reward signals. The authors describe their pilot evidence as partially reconciled and state that confirmatory validation remains outstanding.
Yaxiao Liu (PwC China AI Center), Pengbo Liu (PwC China AI Center), Yiwen Liu (PwC China AI Center) et al.
In briefTimeline-Bench: Testing AI agents on realistic video editing, from raw footage to finished cut
SummaryTimeline-Bench is a benchmark of 56 real video-editing tasks in which AI agents must turn raw production footage into a finished video. Each task is judged by tests covering delivery format, content, the brief's requirements, and a quality test calibrated on blind judgments by 43 video editors. Among 16 agents, the best resolved 15 of 56 tasks (26.8%), and the average was 14.0%. The authors release the tasks, verifier and per-run results.
In briefGame-theoretic trajectory planning that stays robust for air traffic
SummaryThe paper addresses autonomous trajectory planning for Advanced Air Mobility aircraft that must avoid collisions without centralized coordination. It proposes a strategically robust game in which each agent guards against a fictitious adversary that perturbs other agents' controls within a bounded budget at each timestep. The authors show the game remains an exact dynamic potential game under reasonable assumptions and has a quasi-closed-form solution for the inner adversarial problem with linear dynamics. Experiments with up to eight agents show more robust trajectories in high-risk cases, with nearly unchanged low-risk trajectories and only a modest runtime increase.
Victor L. Qin, Nicolas Lanzetti, Saverio Bolognani et al.
In briefCacheRepair: learning to fix missing cross-chunk context when reusing cached text in RAG
SummaryRetrieval-augmented generation over several documents can be sped up by precomputing each chunk's key-value cache and concatenating them, but this skips attention across chunks and lowers answer quality. CacheRepair is a lightweight network trained to predict the difference between independently computed caches and caches from jointly processed chunks, then adds that residual to each token's cache. Across three target LLMs and four datasets, it sits on the measured answer-quality-latency Pareto frontier in eleven of twelve combinations. Its largest repairers cut median time to first token by 1.69 to 4.61 times versus full prefill and raise mean F1 by 2.1 to 26.1 points over direct cache reuse.
Genglin Wang, Wangsong Yin, Yeerzhati Abudunuer et al.
In briefFlexiWorld: Planning with Flexible Action Chunks Across Time Scales
SummaryFlexiWorld is a JEPA-based latent world model for long-horizon, goal-directed control. It trains on mixed-length goal spans and variable-length action chunks, jointly learning a causal action encoder and an autoregressive actor, with Student Forcing to reduce exposure bias. Planning uses the ARCEM method, and the authors report 89.29% mean success across four benchmarks versus 83.98% for the strongest baseline.
In briefVideoPhysEdit: Physics-Based Counterfactual Video Editing via Rigid-Body Scene Reconstruction
SummaryThis paper introduces physical counterfactual video editing (PCVE), which generates a video showing how a physical edit to a source video would change subsequent motion and interactions. The authors propose VideoPhysEdit, a training-free pipeline that reconstructs a rigid-body physical scene reproducing the observed motion, applies edits as simulated interventions, and uses the resulting trajectories to guide video generation. They also release the PCVE-RigidBench benchmark and the Physical Edit Score metric, reporting higher physical edit accuracy than compared open-source and commercial methods with competitive visual fidelity.
In briefCTP-FL: Federated learning that computes gradients along a shared predicted path
SummaryThis paper proposes CTP-FL, a federated learning method in which all clients evaluate gradients along the same predicted path, built from the current global model and the previous aggregated direction. Clients upload the average of their gradients, and the server makes a single global update. Per-round computation and communication match full-participation FedAvg-M. For smooth non-convex objectives, the paper establishes an average-stationarity convergence bound and notes a trade-off: a longer path gives more forward-looking information but larger displacement bias.
In briefExplaining hyperbolic neural networks with geometry-aware relevance propagation
SummaryThis paper studies how geometric operations in hyperbolic neural networks affect feature attributions, showing that equivalent geometric realizations can yield different explanations even when local relevance is conserved. The authors propose LRP-radial-all, which treats radial geometric scaling as modulation and assigns relevance entirely to the signal branch, satisfying relevance conservation, invariance to equivalent factorizations, and zero-curvature consistency. Experiments on hyperbolic MNIST, sEEG, and CIFAR-10 classifiers report competitive attribution fidelity with runtime comparable to Gradient×Input and much lower than Integrated Gradients.
In briefA multimodal autonomic sensing system for objectively assessing patient responses to dental pulp stimulation
SummaryThis study addresses the subjective, hard-to-communicate nature of pain during dental pulp testing. Forty-nine patients underwent cold pulp testing, and the framework combined ECG-derived skin nerve activity, R-R intervals, and electrodermal activity with baseline signals, anxiety scores, and sex. It reached 80.2% balanced accuracy for no-response versus response classification and 60.0% for three-class classification. The findings support the feasibility of non-invasive, objective multimodal autonomic sensing for assessing responses to pulp stimulation.
In briefRestoring Handheld Ultrasound Images via Synthetic Paired Data and Degradation-Aware Enhancement
SummaryLow-cost handheld ultrasound images suffer compound degradation that can mislead diagnosis. The paper proposes a two-stage framework: a cycle-consistent style-transfer model first turns unaligned real hospital images into pixel-aligned low-quality data, then a Dual Degradation-Guided LoRA method fine-tunes an enhancement model on these pairs. On the USenhance2023 dataset, the FID score improves by 16.7% over the strongest baseline.
Yu-Kai Wang, Chun-Xin Tan, Manh-Hung Nguyen et al.
In briefHow Tool Use Changes When Language Models Refuse Harmful Requests
SummaryThis paper examines why large language models are less likely to refuse harmful requests when they operate through external tools than in ordinary conversation. Across several open-weight models, the authors find that harmfulness information is encoded in both interaction modes, but tool-mediated inputs stay permissive until harmfulness reaches a substantially higher effective threshold. Tool-mediated refusal is also more fragile, suggesting that conventional safety evaluations may not fully transfer to LLM agents.
WhyShows tool mediation weakens refusal, using representation geometry, neuron analysis, and intervention experiments across several open-weight models, highlighting agent safety implications.
Abel Rodr\'iguez, Giuseppe Garofalo, Lieven Desmet et al.
In briefDuplexCadence: Exact State and Execution from a Speech Model's Declared Timelines
SummaryThis paper addresses real-time serving of full-duplex speech models, which must turn each second of input into a second of speech before the next arrives. The authors find that the token-to-audio synthesis stage causes overruns through orchestration slack and wastes memory by over-provisioning state. They propose DuplexCadence, which declares the model's native clocks to the runtime and derives demand-sized state allocation and padding-free exact-shape graph replay. On four released models across three decoder architectures, outputs are bit-for-bit identical, speed reaches 2.85 times the stock runtime, and peak memory drops by 38.8%.
In briefSymbolicArena: A Unified Framework for Distilling and Dynamically Evaluating Symbolic Regression Benchmarks
SummaryThis paper introduces SymbolicArena, an infrastructure for evaluating symbolic regression, the task of finding concise mathematical formulas from data. It standardizes 664 tasks and distills them into Core50, a 50-task benchmark that reduces evaluation workload by 92.5% while staying consistent with full-set results. Evaluation shows current methods fit numbers well but recover correct symbolic equations poorly.
