In briefA conditional independence test based on neural quantile resolution
SummaryThe paper proposes QRCIT, a conditional independence test that compares P(X|Z) and P(X|Y,Z) through a quantile deformation learned by a residual quantile-flow model. It yields a global predictive-gain statistic calibrated by conditional randomization and an inspectable deformation field over context and quantile level. The authors also introduce the Structural Dependency Benchmark (SDB) of dependence patterns motivated by real applications. QRCIT improves detection of structured alternatives on synthetic PNL and SDB benchmarks and performs competitively on Sachs, METABRIC, and CCLE/PLX4720 data.
In briefA survey of nonlinear filtering for data assimilation and new generative-AI score filters
SummaryThis survey reviews nonlinear filtering methods used in data assimilation, which combines models with noisy observations to estimate system states. It notes that Kalman, particle, and kernel-based approaches each face limits with nonlinear, non-Gaussian, or high-dimensional chaotic systems. The paper traces the mathematical foundations of the Ensemble Score Filter (EnSF), which encodes the filtering density in a score function estimated by a training-free Monte Carlo method. It also examines fourteen extensions of EnSF, twelve of which keep this core estimator.
In briefDepth or Description? One-step latent scores as alternatives to lookahead in proactive information acquisition
SummaryThis paper studies how a proactive agent should decide which questions to ask before acting. Exact lookahead evaluates combinations of questions but is expensive and relies on a fitted model, while a cheap one-step score can miss such combinations. The authors find that the need for deep search depends on how the hidden cause is described: a one-step score on the label can overlook interactions among answers, whereas a score on a suitable latent cause does not. When that latent encodes the decision, it can replace most of the deep search at a fraction of the cost.
In briefFastAFD: When Splitting Attention and Expert Compute Pays Off on NVLink Racks
SummaryThis paper presents FastAFD, an engine that separates attention from mixture-of-experts (MoE) layers on GB200 NVL72 systems and overlaps computation across microbatches. Compared with the best vLLM wide expert-parallel setting, it reaches up to 1.17x throughput on Qwen3 and 1.16x on Qwen3.5, and matches it on MiniMax-M2.5. The authors also provide an analytical model that predicts AFD throughput within 3% using only profiles from colocated vLLM.
In briefPerceptual distortion for overfitted video coding with sliced spatial matching and temporally aligned errors
SummaryThis paper proposes a perceptual distortion framework for overfitted video codecs. Spatially, it introduces a local sliced Wasserstein distortion that matches feature distributions through one-dimensional optimal transport, capturing cross-channel dependencies without enforcing exact alignment. Temporally, it uses a flow residual that constrains coding errors along motion trajectories to preserve consistency. On UVG and HEVC Class B with the COOL-CHIC codec, it reduces BD-rate by 9.2% in DISTS and 3.0% in LPIPS versus frame-wise WD, and lowers tOF by 10.2% at matched rate.
In briefDo More AI Agents Make a System More Reliable? A Study of Multi-Agent LLM Reliability
SummaryThis paper examines whether software built from several cooperating large language model agents is actually more reliable than a single model. The authors combine HIP-LLM, a hierarchical imprecise-probability method for assessing single LLMs, with a Bayesian network modeling the system's failure logic, propagating failure-probability intervals into a reliability envelope. In a pipeline extracting regulated chemical-hazard statements, majority voting over three models protected against the weakest model but did not clearly exceed the best one. The authors conclude that adding agents yields a more defensible reliability claim, but not automatically a more reliable system.
Burkay Tun\c{c}t\"urk, Koorosh Aslansefat, Yiannis Papadopoulos et al.
In briefMatriarchy Classics: A Method for Swapping Only the Gendered Words in Classic Novels
SummaryThis paper presents a computational method for gender-swapping public domain novels that programmatically guarantees only gendered words are changed. The authors demonstrate it on two novels, transforming Pride and Prejudice into a matriarchal Regency England and The Great Gatsby into a matriarchal Jazz Age. They suggest such gender-swapped novels could help investigate gender biases among readers, authors, societies, and AI systems.
In briefContext-aware automated NMR structure elucidation using chemical precursors as added input
SummaryThis paper proposes a protocol that combines AI-based spectroscopic structure elucidation with chemical context such as known reactants or a desired product. Automated retrosynthetic analysis is used to generate large training sets of molecules paired with synthetic precursors. When the context is informative, Top-15 accuracy reaches up to 95.4%, and even uncorrelated context yields accuracy comparable to the original NMR framework.
Frank Hu, Jonathan M. Tubb, Grant M. Rotskoff et al.
In briefA Method Using Proxy Variables to Measure Racial Disparities in Text-to-Image Models
SummaryThis paper presents an automated method for auditing text-to-image generators for racial bias against demographic benchmarks. It adapts racial proxy variables from U.S. antidiscrimination law, classifies generated images by race, and compares the results with public census data, while also measuring skin tone and image similarity. Experiments on three models using geographic, occupational, and surname proxies show notable variation across models and proxy categories. For example, Imagen 4 Fast overrepresents white-classified images and produces lighter, more homogeneous skin tones.
In briefTeaching Agents Behavioral Principles Instead of Rewards: A New Reinforcement Learning Approach
SummaryThe paper argues that reward-based reinforcement learning forces agents to infer designers' intent from long sequences of noisy scalars, which makes complex tasks hard to learn. It proposes prescriptive policy learning (PPL), which directly teaches agents high-level behavioral principles such as maintaining balance, accumulating energy, and landing smoothly, using key quantities and references to indicate at each step whether a principle is followed. Experiments show that PPL agents learn good, generalizable policies without any reward, outperform reward-based methods with far fewer samples, and transfer their policies to harder settings.
In briefBayesian tensor train kernel machines that infer model complexity automatically
SummaryThis paper proposes Bayesian Tensor Train Kernel Machines (BTT-KM) for supervised learning. Hierarchical ARD-type priors on the tensor train cores allow the TT ranks and feature dimensions to be inferred from data. Mean-field variational inference yields a Bayesian alternating least squares algorithm with the same computational complexity as its deterministic counterpart, along with a closed-form predictive distribution. A synthetic experiment shows that the method recovers the true TT ranks and feature dimensions from an over-parameterized initialization.
In briefScalable Bayesian Gaussian process classification for calibrated cancer histopathology predictions under distribution shift.
SummaryThis paper presents a scalable Bayesian Gaussian Process Classification framework built on frozen pathology foundation model embeddings (Phikon), combining low-rank kernel factorization, Hamiltonian Monte Carlo posterior sampling, and a pivot-based prediction scheme. It is evaluated on breast cancer metastasis detection benchmarks PCam and CAMELYON17-WILDS, the latter testing calibration under hospital-level distribution shift. The authors report improved calibration and predictive performance over stochastic variational GP and neural network baselines, with the largest gains under distribution shift.
In briefTuning open-source LLM judges with soft prompts to match human annotation
SummaryThe paper studies whether soft prompt tuning can make lightweight open-source LLM judges agree more closely with human annotation policy. The authors train soft prompts on the FairBotBench benchmark and evaluate the adapted judges on bias, toxicity, and personalization. Hybrid soft-prompt tuning, which combines textual prompts with learned continuous prompt embeddings, consistently improves agreement with human annotations over untuned baselines. For Qwen 3.5 9B, the adapted judges surpass both hosted reference judges on all three dimensions.
Tam\'as Janusko, Sven Kleinkop, Ricardo Bochnia et al.
In briefData-driven modeling of molten chloride corrosion from static test data
SummaryThis study compiles 291 literature-reported static molten chloride corrosion experiments into a dataset and builds an interpretable machine-learning framework to predict corrosion rates. A Chloride Salt Preparation Index (CSPI) is introduced to place differences in salt preparation across studies on a common scale, and it emerges as the dominant model descriptor. A blended tree-based ensemble achieves a cross-validated R² of 0.85 and an MAE of 0.41 mg cm⁻² day⁻¹, with exposure time, temperature, and alloy chemistry also identified as important, and well-calibrated prediction intervals provided.
In briefMachine learning for predicting live birth after IVF/ICSI from clinical data: review and meta-analysis
SummaryThis paper systematically reviews machine-learning models that predict live birth after IVF or ICSI using routine clinical data, covering 48 included studies. Pooled discrimination was summarised with random-effects meta-analysis of the area under the ROC curve. Models evaluated only internally pooled to 0.814, versus 0.756 for those tested on later periods or other centres, with extreme heterogeneity. Most studies were at high or unclear risk of bias, no model showed evidence of clinical impact, and the certainty of evidence is low to very low.
Abdul Raheem Shahzad, Muhammad Zaheer Khan, Umar Ajaib et al.
In briefPromptOps: Iterative Prompt Optimization for Keeping AI Judges Accurate
SummaryThe paper addresses how large language models used as automated judges can drift away from intended human evaluation standards as user interactions and score distributions change. It introduces PromptOps, a closed-loop framework that monitors and iteratively recalibrates these LLM judges, separating decisions about when to optimize, which method to use, which feedback to select, and how to update the prompt. Experiments show iterative optimization can recover judge performance after controlled distribution shifts, but the best algorithm and update timing depend on context.
Sven Kleinkop, Tam\'as Janusko, Ricardo Bochnia et al.
In briefRegulating Autonomous Mobility-on-Demand Operators Using Fixed and Adaptive Penalties via Multi-Agent Reinforcement Learning
SummaryThis paper studies how penalty-based regulation can curb excessive surge pricing by autonomous mobility-on-demand operators and improve service availability across regions. Both the operator and the regulator are modeled as learning agents using graph-based multi-agent reinforcement learning. Experiments on real Manhattan data show that fixed surge pricing penalties cut the share of travelers facing excessive prices from 22% to 1%. Dynamic region-specific penalties raise the regulator's reward by up to 15.6% over fixed penalties of equal mean intensity.
Korbinian K\"ugler, Carolin Schmidt, Florian Dandl et al.
In briefPatient-reported cancer screening barriers mined from online communities
SummaryThis paper analyzes over 56,000 posts from six cancer-related Reddit communities, extracting 576 posts from U.S. patients. It applies keyword filtering, barrier triage, open coding, and codebook synthesis to categorize patient-reported obstacles to cancer screening. The findings identify structural, informational, and psychosocial barriers that vary by cancer type and community context.
In briefConsistent yet Contextual: Learning Author-Level Style Mixtures for Authorship Attribution
SummaryThis paper addresses authorship verification and attribution, which typically model document style rather than the authors themselves. The authors introduce several ways to model an author so that the representation captures multiple consistent but distinct writing styles, and report that this yields significant performance gains. They also study when an author's writing is most idiomatic of their style across four settings: audience size, temporal distance from training, preceding context, and genre norms.
In briefA Night at the LLM Zoo: Flow Matching for Generating Pretrained LLM Weights
SummaryThe authors build a population of GPT-2 language models, varying the training data mixture across web text, books, mathematics, coding, and French. They train a flow matching model conditioned on the data mixture to generate complete pretrained model weights in a compressed weight space. The generator produces models for unseen data mixtures that perform similarly to the models in the original population.
WhyThe authors condition flow matching on data mixtures to generate full GPT-2 weights, reporting zero-shot generalization to held-out mixtures across a 300-model zoo.
In briefServer-side sparse expert assignment for federated mixture-of-experts models
SummaryThis paper proposes S³-MoE, a federated mixture-of-experts framework in which the server sparsely assigns experts to clients. It estimates each client's expert relevance from partial observations and refines assignments with an exploit-explore mechanism. The authors also give guarantees on relevance-estimation error and exact Top-K recovery. In CIFAR-100 experiments with a CNN-based MoE, each client loads only 5 experts and achieves 1.31 points higher mean final accuracy than Full-20 FedAvg, which gives every client all 20 experts.
In briefExact polynomial-time computation of proper equilibria in two-player zero-sum extensive-form games
SummaryThis paper presents a deterministic polynomial-time algorithm that computes an exact extensive-form proper equilibrium in finite two-player zero-sum games with perfect recall and rational data. Each information set is assigned a geometric permutahedron of allowable action probabilities, which is embedded in a polynomial-size sequence-form polytope via a scaled doubly stochastic representation. The authors use sign stability of a parametric linear system to obtain an exact rather than approximate equilibrium, and the resulting probabilities are rational with polynomial encoding length.
WhyThe authors report a deterministic polynomial-time exact algorithm for extensive-form proper equilibria in two-player zero-sum games, using geometric permutahedra and doubly stochastic embeddings; imp
In briefLarge-scale symbolic models of human and machine reasoning
SummaryThe paper introduces Behavior-Conditioned Symbolic Synthesis (BCSS), a method that automatically searches for interpretable, deterministic symbolic programs that reproduce observed behavior. The authors apply BCSS to model human and machine behavior across hundreds of reasoning problems. They find that both tend to oversimplify, adopt partial explanatory rules, and misinfer dependencies between problem features. Reasoning effort predicts how much machines simplify their observations.
In briefWide-View Synthesis from Multiple Cameras for Autonomous Driving
SummaryThis paper formulates wide view synthesis (WVS), which generates a geometrically consistent wide image at a chosen virtual viewpoint from several perspective cameras. The method, Wide-View Gaussian Splatting (WV-GS), is a feed-forward Gaussian splatting framework that avoids per-scene optimization. It uses Piecewise Perspective Rendering to reduce peripheral stretching and Spatial Affine Photometric Modeling to handle photometric differences across cameras. The authors also introduce WideDrive, a synthetic benchmark, and report experiments on multiple datasets.
In briefOptimistic Hedge in Games: How Much Regret It Can Avoid
SummaryThis paper studies Optimistic Hedge, a learning method that updates weights multiplicatively with one step of extrapolation, and how much regret it incurs in general-sum games. The authors show that in two-player games some learning rate gives regret of O(n log n) independent of the number of rounds. In games with three or more players, they construct games where some player's regret grows exponentially in the number of actions. Their upper bound comes from a temporal approximation framework that also gives a simpler proof of a known bound for m-player games.
WhyThe authors show two-player games admit horizon-independent regret O(n log n), while three-player games have exponential lower bounds, using a temporal approximation framework.
In briefWhen Verification Is Not Agreement: Cross-agent LLM-to-Dafny formalization
SummaryThe paper studies a protocol in which one language model writes a Dafny predicate from clinical trial criteria and another writes a method checked against it. Even a same-model control agrees only 59.3% of the time when re-extracting from identical text. Across four model configurations on 96 trial documents, agreement ranges widely, and a weaker implementer can pass verification by directly invoking the specifier's predicate. The authors conclude that successful verification alone is not evidence of independent agreement.
In briefAutomated Feedback to Peer Tutors Boosts Their Questioning and Learner Reasoning, Not Speaking Balance
SummaryThis randomized controlled trial tested automated feedback on discourse with 1,030 volunteer tutors and 7,443 learners across two four-week SAT math bootcamps. Tutors received no feedback, feedback on their own talk, or that feedback plus participation feedback for their learners. Feedback to tutors increased eliciting of learner ideas, uptake of contributions, and learner reasoning by 8-15%. However, tutor talk time, averaging 91.5% in the control group, did not change in any condition, and learner-facing feedback added no further change.
In briefShaping and Monitoring Models via Internal Representations
SummaryThe paper proposes Mechanistic Shaping, a training framework that steers models through their internal representations so they avoid specified concepts in outputs and in the features used for decisions, such as harmful content or gender in résumé screening. Experiments across vision, medical imaging, and language, with models up to 32B parameters and several training methods, show targeted behaviors are substantially reduced while task performance is preserved. The authors also introduce a hierarchy of monitors and identify three ways shaped concepts can evade them, finding that shaped concepts largely remain monitorable.
WhyThe authors propose Mechanistic Shaping, a unified framework, report reduced targeted behaviors across vision, medical, and language tasks up to 32B, and characterize three circumvention mechanisms.
In briefFinite-time convergence and regret guarantees for neural reinforcement learning in POMDPs
SummaryThis paper studies reinforcement learning in partially observable Markov decision processes (POMDPs) by approximating them with a superstate MDP built from a truncated finite-memory history. Using nonlinear neural value function approximation, the authors prove a finite-time critic error bound for temporal-difference learning. Combined with policy mirror descent, they obtain finite-time sublinear regret against the optimal POMDP value. The authors claim this is the first such result based on neural function approximation for finite-memory POMDPs, and numerical experiments support the theory.
Jingjuan Huang, Zhiyao Zhang, Alvaro Velasquez et al.
In briefMulti-agent reinforcement learning that explores cooperatively under a performance constraint
SummaryThis paper tackles cooperative exploration in multi-agent reinforcement learning, where agents must learn to coordinate in a huge joint action space. The authors propose EFXPLORER, which maximizes exploration objectives subject to a constraint that protects task performance, solved via an epigraph reformulation with adaptive exploration budgets. A successor distance-based intrinsic reward is added to encourage diverse, long-horizon exploration. Experiments on SMAX, VMAS, and MPE benchmarks report outperforming strong baselines and discovering novel cooperative strategies.
In briefStochastic dynamics of generative AI, from autoregressive LLMs to interacting networks and neural SDEs
SummaryThis paper studies three types of generative AI models from a stochastic dynamical-systems perspective. It formulates large language models as high-dimensional nonlinear autoregressive models with attention-based dependencies, linking pretraining by maximum likelihood to token-by-token generation. It then analyzes networks of interacting LLMs using mean-field ordinary differential equations, covering emergent preferential attachment and persistent structural biases. Finally, it treats neural stochastic differential equations as continuous-time generative models and discusses their training via Wiener-space cubature methods based on rough-path analysis.
In briefSD-LoRA: Distilling Structural Knowledge into LoRA During Training for Few-Shot Vision-Language Adaptation
SummaryThis paper presents SD-LoRA, a method for adapting vision-language models like CLIP to tasks with very few labeled examples. During training, it uses a graph model over high-resolution image tokens and text to capture fine-grained structural relations, then transfers this knowledge into a lightweight LoRA-adapted CLIP model. The graph is discarded at inference, which needs only one CLIP-LoRA forward pass and a cache-based classifier. The authors report accuracy gains over a strong CLIP-LoRA baseline across 11 few-shot benchmarks.
WhyThe authors distill a training-only token graph teacher into LoRA and discard it at inference, reporting 1.8–3.0 point gains over CLIP-LoRA across 11 few-shot benchmarks.
Mohammed Rahman Sherif Khan Mohammad, Ardhendu Behera, Sandip Pradhan et al.
In briefSkin lesion segmentation using images generated by a text-to-image model (GENIQ)
SummaryThis paper presents GENIQ, a training-free method for segmenting skin lesions in dermoscopy images without annotated masks. It builds skin and lesion prototypes from images generated by a text-to-image diffusion model, then refines them using the most skin-like and lesion-like patches from unlabelled dermoscopy images. Test patches are scored along the refined direction with a gated threshold, achieving Dice scores of 0.85 to 0.88 on three test sets.
In briefCompetitive flow matching to widen the quality-coverage frontier
SummaryFlow matching generators often trade image quality against coverage, and this paper attributes the issue to the training objective, where one noisy image can map to many clean images but the model predicts only one velocity. The authors train with a winner-take-all scheme: each example gets K candidate velocities from random low-dimensional codes, and only the best-matching candidate receives the loss. They also introduce auxiliary conditioning scaling as a new guidance mechanism alongside traditional guidance. On ImageNet-256 with SiT and RAEv2 testbeds, the method reports substantially lower FD_r,6 at matched recall points.
In briefVariant reclassification: spotting need for reinterpretation in online health communities
SummaryThe paper presents PRISM, an agentic large-language-model pipeline that screens posts in online health communities, checks whether the discussed genetic variant was later reclassified, and drafts a disclaimer-bearing outreach message in a no-contact dry run. Across 3.9 million posts from 32 communities, reinterpretation-relevant discussion is rare (0.288%) but concentrated in specialist genetics communities and has grown fivefold since 2014. PRISM surfaces 123 candidate classification differences, five of which moved from uncertain to pathogenic or likely pathogenic.
Kamalakkannan Ravi, Mengshu Nie, Annapurna Poduri et al.
In briefClosed-Loop Agent Evaluation: Layered Measurement and Self-Calibration
SummaryThis paper proposes a closed-loop evaluation method for deployed agents, treating evaluation as performance estimation under biased raters and censored outcomes. Conclusions at each level must trace back to atomic data and be computed by versioned rules, across span, trace, conversation and engagement levels. An inner loop audits evidence, rules and metrics, while an outer loop uses per-component regression checks to drive agent improvement. The method is applied to the business-analysis agent SharpInsight and ported to public conversation data.
In briefAuthorization Propagation in Agent2Agent Systems: A Security Test of Multi-Hop Delegation
SummaryThis paper studies security risks in delegated Agent2Agent (A2A) workflows, where the agent sending a request differs from the principal that granted the authority. The authors test a layered verification design in a synthetic four-party healthcare reimbursement workflow, rejecting 14 of 15 adversarial conditions. A separate experiment shows that a holder of the shared authorization-evidence HMAC key can alter the represented authority and produce valid evidence. The results support separating request authentication from inherited authority and call for origin-verifiable delegation and semantic attenuation.
Gabriele Donno, Kashif Mehmood, and Katina Kralevska
In brief"Gaussians Can Subtract": Neural Dissolve Fields for Hybrid Scene Rendering
SummaryThis paper studies hybrid scene rendering that combines textured meshes with 3D Gaussian primitives. The authors argue that conflicts between mesh and Gaussians are not always missing content, and can also come from geometry errors or occlusion mismatch. They propose Gaussolve, which selectively dissolves inconsistent mesh content and then refills the revealed regions with residual Gaussians. A lightweight Neural Dissolve Field learns where and from which viewing directions mesh content is dissolved, and a continuous transmittance fusion rule smoothly exposes the Gaussians. Experiments on structured and semi-structured scenes report higher visual fidelity with fewer Gaussian primitives and faster rendering.
In briefAdapting the Process of Adaptation: Darwinian, Lamarckian, and Agent-Directed Evolution in AI
SummaryThis paper develops a theoretical framework comparing Darwinian, Lamarckian, and agent-directed evolution as ways of adapting to environmental change. It finds that agents generating and evaluating their own behavioral variants can adapt faster, but misspecified evaluation targets can drive convergence toward wrong behavior even as measured performance improves. The authors argue that independent evaluation and oversight must keep pace with autonomous AI development.
In briefFollow the Policy, Fail the Benchmark: Auditing and Revising Policy-Based Decision Benchmarks
SummaryThe paper audits nine public benchmarks that test whether systems apply written rules to financial and operational cases. It finds unsupported reference answers, reference errors, and evaluators that reject acceptable answers or credit incomplete ones. Rescoring unchanged model responses shows that correcting the scorer substantially changes acceptance rates and model gaps. The authors present BAR, a versioned collection of audit evidence and revisions for reevaluation.
In briefLowRAM: Low-Rank Keys and Amortised Values for Test-Time KV Cache Compression
SummaryThe paper proposes LowRAM, an asymmetric method for compressing the key-value (KV) cache used in autoregressive decoding. Its core technique, value cache amortisation, fits a lightweight per-head predictor at test time and stores residuals only for the tokens it reconstructs worst, achieving about 4x near-lossless value compression, combined with cross-layer low-rank key compression. Across three model families it stays close to lossless and reaches up to 5x higher peak decode throughput at 4x compression. It also composes with quantisation and eviction, reaching 26.9x compression at 94.5% of baseline performance on long contexts.
Teresa Delgado, Hugo Hazard, Martin Benfeghoul et al.
In briefTraCE: A Concept Framework That Explains Video Predictions by Their Motion Sources
SummaryVideo models often rely on motion, but existing explanations rarely show which motion sources support a prediction. TraCE clusters optical-flow trajectories from training videos to discover motion concepts from people, objects, and interactions, alongside object and scene concepts. These support action recognition and question-conditioned video question answering. Across six benchmarks, TraCE stays competitive with non-interpretable baselines, with an average action recognition accuracy drop of 1.0 points and a nominal 0.4-point gain on VQA.
In briefFrom Pixels to Actions: Toward Native Vision-Language-Action Policies
SummaryThis paper introduces NEO-VLA, a vision-language-action framework without a separate visual encoder. Images are tokenized by lightweight patch convolutions and evolve jointly with language throughout the backbone, letting an action expert use representations from different depths. Under controlled comparisons, the native design outperforms encoder-based models, and layer-wise probing shows action-critical signals often peak before the final visual layer. The approach yields stronger and more robust control across manipulation tasks and distribution shifts.
In briefA framework for adding context to AI social agents while keeping their behavior stable
SummaryResearchers use large language models as agents in social simulations, but these agents often change behavior when prompts are reworded without changing their meaning. The paper introduces controlled context enrichment, which systematically adds neutral padding, generic social background, descriptive decision grounding, or normative decision grounding to prompts. It also proposes selective robustness as a validation criterion: insensitivity to meaning-preserving prompt changes combined with a directional response to decision-relevant social changes. The framework is demonstrated in a Prisoner's Dilemma-like game, where different forms of enrichment produce distinct behavior patterns that vary between the two models tested.
Mattia Greiche, Reihaneh Rabbany, Maximilian Puelma Touzel
In briefFlowReconBench: Benchmark Comparing Learned and Classical Methods for Rebuilding Flow Fields from Sparse Sensors
SummaryThis paper presents FlowReconBench, a benchmark that evaluates ten learned and five classical methods for reconstructing full flow fields from sparse sensor data. The methods are tested on a laminar cylinder wake, two-dimensional Kolmogorov turbulence, and three-dimensional isotropic turbulence, using accuracy, calibration, spectra, and computational cost as criteria. The results show that the best-performing method depends on the regime, with classical methods excelling in low-order and dense-sensing cases and data-driven methods performing better when sensor density is low or observability is poor.
In briefBEAKER: An Expert-Curated Benchmark for Embodied AI Brains in Self-Driving Chemistry Labs
SummaryThis paper introduces BEAKER, a benchmark for testing how well multimodal large language models can act as embodied decision-making systems in self-driving chemical laboratories. It contains 1,000 question-answer samples across 11 image and video task types, all manually written and reviewed by chemistry and AI experts. The authors evaluated 28 models and found they can follow coarse experimental procedures in videos but struggle with precise action grounding, fine-grained equipment and safety recognition, and ordering steps in mid-length procedures.
In briefFrontierPhysics: A Benchmark for Agents Doing End-to-End Physics Research
SummaryThis paper introduces FrontierPhysics, a benchmark of 56 tasks written by 42 physicists from their own research, spanning nine domains of theoretical, computational, and experimental physics. Agents must review literature, plan research, conduct experiments, and produce structured outputs and scientific manuscripts. Evaluation combines deterministic verifiers with expert-authored rubrics to judge correctness and publishability. Across 22 models in 26 configurations, no configuration passed more than 46 of its 168 rollouts.
WhyThe authors introduce an end-to-end benchmark of 56 physicist-authored research tasks across nine domains, reporting that no configuration passes more than 46 of its 168 rollouts.
In briefSimulation-trained humanoid robot motion driven by vision and language (HALO system)
SummaryThis paper presents HALO, a humanoid robot system that acts on visual and language inputs. It generates vision-language-motion training data across a thousand procedurally generated virtual homes, so data scales with computing rather than human effort. HALO learns skills such as sitting on chairs, navigating to named objects, passing sideways through half-open doors, crawling under tables, and pushing boxes, and combines them for long-horizon tasks. Without training on any real home, it reached a 91% success rate over 360 trials in six unmapped real-world environments.
WhyThe authors synthesize vision-language-motion data in simulation and report 91% zero-shot success over 360 trials in six real environments.
In briefUsing reinforcement learning to improve a general robot policy across many tasks
SummaryThe paper asks whether generalist robot policies can be improved online during deployment across many diverse tasks at once. The authors build structured evaluations and systematically compare choices in value estimation, policy extraction, and resource allocation. They report that the right recipe improves policies across diverse tasks and can also raise performance on tasks not seen during online training, with the training task distribution and value-based RL identified as critical.
WhyThe authors study online RL improvement of generalist robot policies across many tasks, reporting generalization to unseen tasks.
Michal Korniak, Marco Bagatella, Ralf R\"omer et al.
In briefMotion-Guided Contrastive Regularization for Latent World Models
SummaryThis paper addresses static collapse, a failure in JEPA-style latent world models where the encoder keeps episode-specific static background while suppressing manipulation-relevant state changes. On OGBench-Cube, randomized backgrounds substantially degrade planning performance. The proposed method uses optical-flow similarities as soft contrastive targets, combined with latent prediction and SIGReg. Reported planning success rises from 52% to 78% on OGBench-Cube with randomized backgrounds and from 14% to 54% on LIBERO-Object.
In briefOcclusion for Free: detecting pedestrian occlusion without occlusion labels by comparing two monocular distance cues.
SummaryThe paper proposes a training-free method that flags occluded pedestrians by comparing the ordering of apparent box height and monocular metric depth, which are pulled apart when an occluder is in front. It needs no focal length or occlusion annotations and reports AUC 0.909 on KITTI for heavy occlusion. It also reports AUC 0.819 on synthetic CARLA scenes, against 0.644 for distance alone.
Chibueze Praise Onyenweaku, Toyibat Adele, Sikirulahi Opeyemi Abdulkareem et al.
In briefTrust-responsive runtime authorization policy for agentic AI, with parameter sensitivity analysis
SummaryThis paper studies how an autonomous agent's runtime condition estimates are turned into operational authority, and proposes the Trust Responsive Authorization Policy (TRAP). TRAP assigns authorization levels from a readiness score combining trust, operational stability, and contextual risk, escalating only after consecutive observations above a threshold with a margin, while restricting authority immediately on a sharp drop. Compared with static, trust-only, risk-only, and symmetric policies on 100 Monte Carlo traces, TRAP restricts faster under abrupt degradation, but its restriction advantage depends on degradation being abrupt.
Ahsan Habib Shihab, Imre Felde, Ramis Akhmedov et al.
In briefProcedural Adversarial Learning: Training and testing vision models on hard cases found by a controllable scene generator
SummaryThe paper introduces procedural adversarial learning (PAL), which searches the control parameters of an interpretable procedural image generator to find difficult scenes for minimax training and adversarial evaluation. It proposes the Simple-PAT algorithm, which uses zeroth-order optimization to create adversarial training samples on demand, and releases PAL-Dojo, a set of procedural generators for surface normal estimation, image segmentation, and human pose estimation. Against a cross-model black-box adversary with 50M queries, Simple-PAT reduces worst-case error by a geometric mean of 3.15x over standard training and 2.63x over the strongest robust training baseline.
In briefAdaptive Multi Mode: routing black-box optimizers under a budget using online landscape probes
SummaryThe paper introduces AMM, which routes a fixed evaluation budget among five classical optimizers (quasi-Newton, trust-region, coordinate, multistart, evolution-strategy) using online probes of landscape features such as curvature and separability. Tested once on held-out bbob instances against up to seven competitors, AMM had the highest target share in 382 of 600 function-dimension-budget cells. The authors released it as an open-source Python package.
In briefSpecCompact: Speeding Up Context Compaction for LLM Agents with Parallel Draft-and-Verify Summaries
SummaryLLM agents compress growing interaction histories, but summarizing long histories with large models adds significant delay. SpecCompact speeds this up by having a small fine-tuned drafter generate structured summary slots in parallel, while the target model verifies each line and repairs rejected content. Across AppWorld and τ³ benchmarks with two target models, it achieves 2.04 to 5.63 times faster compaction than Hermes-style summarization. A user study within Hermes reports a median compaction speedup of 8.68 times.
In briefAutonomous AI agents drift from human preferences when formalizing proofs
SummaryThe paper introduces "simulated collaboration," a method for comparing how AI agents and humans approach proof construction. Using nine large human collaborations in Lean, covering 4.7M definitions and over 6K commits, the authors find that AI agents are strongly drawn to explicitly marked gaps. They also show aversion to densely connected proof centers and complex or surprising goals, and tend to regularize surprising proof bodies toward baseline forms.
In briefLearning to decode linear error-correcting codes with a neural network based on belief propagation
SummaryThis paper proposes a Generalized Belief Propagation Network (GBPNet) for decoding linear error-correcting codes. It keeps message-passing symmetry and the Tanner graph structure while adding cascaded variable-node updates and learned element-wise message scaling. Two training strategies, progressive training and high-SNR training, are introduced to improve decoding accuracy. Experiments on several linear codes report lower bit error rates than conventional belief propagation and the evaluated state-of-the-art neural decoders.
In briefPixelSpline: continuous video upscaling in space and time using shared pixel motion paths
SummaryThis paper presents a framework for continuous space-time video super-resolution that combines long-context encoding, continuous high-resolution feature lifting, and an endpoint-anchored module called PixelSpline. PixelSpline models each pixel's motion as one continuous trajectory shared by both endpoint frames, and uses it to warp endpoint colors into a coarse intermediate frame and align high-resolution features for residual reconstruction. Experiments across spatial and temporal upsampling settings show the method offers a favorable quality-efficiency trade-off.
In briefTurn-level best-of-N sampling to improve LLM agents on long tasks
SummaryThis paper evaluates a inference-time scaling method for LLM agents on complex, long-horizon tasks. At each turn, a generator model proposes N candidate actions and a selector model chooses one to execute in the environment. Experiments on long-horizon terminal tasks report accuracy gains of 5 to 10 percent and better consistency across repeated attempts.
In briefCounterFactual Retrieval: An Evidence-Aware Evaluation Method for Identifying Individual Animals
SummaryThe paper argues that current evaluations of individual animal identification report only retrieval accuracy, which does not reveal whether a model relies on animal features or identity-correlated shortcuts. The authors propose Counterfactual Foreground Retrieval (CFR), which evaluates the same query-gallery task using full-image, foreground-only, and shortcut-only views. They also release around 0.5M masks for 34 datasets along with code for mask generation and CFR evaluation.
Ankit K. Upadhyay, Aleksandr Algasov, Ekaterina Nepovinnykh et al.
In briefBatch-conditioned semantic anchors for robust test-time adaptation of vision-language models
SummaryThis paper proposes JPTA (Joint Prior and Transport Adaptation), a framework that improves vision-language models such as CLIP at test time by refining predictions on an unlabeled batch. It treats CLIP text prototypes as semantic references that shift toward image-side batch structure only under calibrated evidence, scaled by batch reliability. According to the paper, JPTA outperforms TransCLIP and StatA across standard benchmarks, low effective-class regimes, all-class evaluation, and online streams, with the largest gains on sparse and prior-shifted batches.
Mohammed Rahman Sherif Khan Mohammad, Ardhendu Behera, Sandip Pradhan et al.
In briefTrade-offs between hierarchical representations and quantization in text embeddings
SummaryThis empirical benchmark study examines how two existing techniques, hierarchical text embeddings with variable-length representations and quantization that lowers numerical precision, interact. The authors evaluate 38 tasks from the Massive Text Embedding Benchmark using four transformer-based embedding models. Int8 quantization is statistically indistinguishable from float32 for pair classification and re-ranking, while binary quantization is significantly worse across all task categories. Retrieval degrades much faster than other categories.
Kai Ponel, Marco Polignano, Erasmo Purificato et al.
In briefCompressionBench: Can coding agents compress large language models?
SummaryThe paper introduces CompressionBench, which frames LLM compression as a terminal task where agents receive a checkpoint, storage limit, inference environment, and time budget, then build compressed weights and executable inference code. Across six agents and six model-size settings, 29 of 36 submissions passed independent verification. The best per-task designs average a 3.4% perplexity increase over BF16 and 77.33% downstream accuracy, versus 79.56% for BF16.
In briefLongBanana: An Expert-Verified Benchmark for Multi-Reference Image Synthesis
SummaryThis paper introduces LongBanana, an expert-designed and expert-verified benchmark for long-context image synthesis from multiple reference images. It contains 90 tasks and 4,670 atomic evaluation items, scored on content fidelity, relational binding, and global constraint satisfaction using an entity-relation graph. The authors also present LongBanana-Gen, an agent-harness framework, and report that a ReAct-Plan-Exec-Reflect workflow yields gains that depend on the harness.
In briefVisual Token Routing: a method for keeping only the image tokens that matter
SummaryThis paper introduces Visual Token Routing (VTR), a training framework for computer-use agents that decides which visual tokens from screenshots are passed to the language model. A lightweight spatio-temporal router, trained jointly with the vision-language model, selects image regions from the current and previous screenshots. The authors report that after 2,000 training steps, VTR matches full-resolution performance on OSWorld, CUA-World, and video game action prediction while using 2.5 to 5.8 times fewer visual tokens.
In briefTrainspotting: Learning 3D Object Reconstruction from Scattered Online Photos
SummaryCurrent feedforward 3D reconstruction models perform poorly on object photo collections taken in the wild, since they are mainly trained on single-session captures. This paper introduces Trainspotting, a dataset of 12,068 objects and 311,886 photos from Wikimedia Commons, with recovered camera poses, depth, and object masks. Objects are grouped into fixed-site, moving, and shared-design regimes, and the models fail most severely on the latter two. Fine-tuning VGGT-Ω on this dataset improves pose, depth, and geometry on held-out objects, with similar gains on the NAVI benchmark.
In briefPersona Perspectives: How Adopting Viewpoints Reshapes LLM Internal Representations
SummaryThis paper examines whether prompting a large language model to adopt a political persona, such as liberal or conservative, changes the geometry of its internal representations. The authors compare representations of over 9,000 words across 33 semantic axes under two political personas and two neutral baselines. They find that persona conditioning systematically shifts how words project onto semantic axes, and classifiers trained without a persona yield politically slanted truthfulness and morality ratings when applied to persona-conditioned models.
In briefHow to Handle Conflicting Criteria for Judging Whether AI Explanations Are Faithful
SummaryThis paper examines why existing criteria for evaluating whether algorithmic explanations faithfully reflect a model's behavior can reach conflicting verdicts on the same explanation. The authors introduce a faithfulness functional that locates these disagreements in what an explanation claims, which comparisons support that claim, and how agreement across comparisons is aggregated, and they reconstruct existing criteria within it. A case study of adverse action notices shows how a specific setting can justify some evaluation choices while leaving others open.
Grant Shanklin, Ji\v{r}\'i N\v{e}me\v{c}ek, Asia Biega
In briefWhere to Connect: DGP, a method that lets multimodal LLMs dynamically choose visual layers via gated pathways
SummaryThe paper proposes DGP (Dynamic Gated Pathways), which creates all-to-all connections between every ViT layer and every LLM layer, with each LLM layer dynamically gating its preference over visual layers based on its own semantic state. Built on LLaVA-1.5-7B and 13B, DGP improves over the 7B baseline on several benchmarks and surpasses prior multi-layer visual fusion methods on most. Layer-wise routing analyses show that different LLM layers select different visual granularities on demand.
In briefPredicting grain growth in polycrystalline materials under changing temperatures with neural networks
SummaryGrain growth during heat treatment affects the microstructure and mechanical properties of polycrystalline materials. This work extends a previous forecasting framework, which worked only under constant temperature, to non-isothermal conditions by combining a convolutional network for grain size distributions, a multilayer perceptron for thermal inputs, and a recurrent network for temporal prediction. The model is trained and evaluated on data from validated front-tracking simulations across diverse non-isothermal thermal paths. It reproduces grain growth dynamics under complex thermal histories with high accuracy.
In briefServerBench: Can coding agents design specialized LLM serving systems?
SummaryThis paper introduces ServerBench, a benchmark testing whether coding agents can design complete, specialized large language model serving systems. Agents experiment toward a shared goal of latency-qualified token value per dollar of GPU cost, with stateful replay preserving session dependencies and independent verification checking submitted systems. Across 11 coding-agent models and three deployments, Claude Sonnet 5.5 reaches 2.97 on Gemma E4B, 4.87 times the vLLM reference.
In briefSame Layout, Different Count: How Number Words Behave Above Ten in Text-to-Image Models
SummaryThis paper studies whether number words control object counts in five open text-to-image models when requests exceed about ten objects. Holding the object and starting noise fixed, the authors changed one condition at a time and had people count 3,088 images. Models follow small numbers, but their adherence weakens past a limit, and the number word does not reliably steer the count in a consistent direction.
In briefBefore You Continue: Checking if AI Chat Assistants Remember Time-Based Safety Limits
SummaryThis paper introduces TASC, a benchmark that tests whether large language model assistants apply time-dependent constraints when answering conversational requests. It includes 3,418 inputs built from 128 templates, plus 33 scenarios comparing elapsed-time calculation with applying that information. Across 13 models, assistants considered time-dependent constraints in only 30.5% of cases where the proposed action violated a rule, and in 62.1% of cases they failed to apply constraints even when they calculated the elapsed time correctly.
In briefContext-aware image generation that adapts its order to reason better
SummaryThis paper proposes a context-aware autoregressive image generation framework that dynamically chooses the next token location based on global context and current local information, instead of using a fixed or random order. Experiments show improved structural robustness across masked autoregressive and scale-wise autoregressive models. The authors also report that the approach helps models better understand complex semantic instructions and produce images that follow them more faithfully.
In briefPredicting LLM answer activations from context before the first answer token
SummaryThe paper studies how a single context can produce many different language model answers under stochastic decoding, which normally requires costly rollouts. The authors train context–answer metamodels that map context activations to point or distribution forecasts of mean-pooled answer activations. The four metamodel families forecast these activations with high fidelity, improve up to 500,000 training prompts, and generalize zero-shot to out-of-domain data. Forecasts retain behavior-predictive signal and transfer from Qwen2.5 to Qwen3.5 and Gemma 4.
In briefLooM: Loop-Invariant Memory for Efficient Looped Transformer Language Models
SummaryThis paper proposes LooM (Loop-invariant Memory), a method for looped Transformers that addresses memory growth with loop count, coarse compute adjustment, and difficulty changing loops after training. In a causal encoder-decoder, only the decoder layers after the first are looped, reading a global memory written once per token, and key-value entries are published only on the first visit so the cache does not grow. One model trained on randomly sampled budgets supports every budget from 3 to 21 layer executions per token without a router. With enough loops, LooM outperforms a single-pass baseline trained on the same data at the same cache size.
In briefRobust stream classification with time-neutralising decision trees
SummaryData streams are potentially endless sources of data whose distribution changes over time (concept drift), forcing deployed predictive models to be repeatedly retrained. The paper introduces Time Neutralising Trees (TNT), a decision tree that filters out old training samples no longer relevant to the current stream context and re-includes them if they become relevant again. Evaluated on 24 real and semi-real classification streams from the USP DS repository, TNT reportedly achieves best-in-class performance across a wide range of concept drift types.
In briefAlignment Myopia in Generative AI: a paper arguing current AI alignment ignores long-term effects on people and should shift toward long-horizon alignment.
SummaryThis paper argues that current generative AI alignment relies on short-horizon feedback that prioritizes immediate conversational satisfaction. It contends this overlooks long-term effects of persistent human-AI interaction, such as cognitive atrophy, social erosion, and diminished human agency. The authors map tensions between near-term utility and long-term costs across four facets of human well-being and advocate alignment strategies that actively foster long-term well-being beyond simple preference optimization.
Remi Denton, Aida Davani, Cynthia L Bennett et al.
In briefDeveloping AWS Cloud FPGA Designs with High-Level Vitis HLS
SummaryThis paper presents a toolchain for developing machine learning models on AWS EC2 F2 cloud FPGAs from high-level code, connecting Vitis HLS to the HDK and AFI workflow. It introduces a four-level methodology for localizing errors, including FPGA-specific circuit errors. The authors validate the approach by implementing the embedding layer and the first of four encoder layers of a TinyBERT address Named Entity Recognition model on FPGA hardware. Weights are streamed from host DRAM, and performance is weaker than CPU and GPU implementations.
In briefSystolic Linear Attention: A TPU-Tailored Gated Delta Attention Design
SummaryThis paper redesigns gated delta linear attention layers, which were originally tuned for GPUs, to better fit Google TPU hardware. The proposed Systolic Linear Attention (SLA) widens keys to 256, groups value heads and decay gates, and computes chunk solves in one fused kernel. The authors report that on TPU v7x, SLA layers run 1.6x to 2.5x faster than the open-source Tokamax KDA kernel, and train 11% to 24% faster than KDA in large MoE pre-training at similar perplexity.
In briefTraining and calibration data for open models on ARC-AGI-2 puzzles
SummaryThe paper studies which training and calibration data help three open models solve ARC-AGI-2 tasks under offline, four-GPU, 12-hour limits. For the NVARC Qwen3-4B model, continued fine-tuning on 1,920 real task records raised the semi-private score from 28.96 to 33.06 (mean of five runs), outperforming checkpoints trained on synthetic tasks. For Qwen3.8-27B, combining fine-tuning on its own correct reasoning traces with an improved inference procedure raised the score from 26.81 to 44.03, with most of the gain attributed to the inference procedure.
In briefToroFit: Fitting Shapes Made of Joined Tori to High-Dimensional Data
SummaryThis paper presents methods for detecting and parameterizing Euclidean point cloud data that lies near a wedge sum of tori, a shape formed by joining several toroidal components. The approach uses persistent cohomology and cup products to identify the distinct toroidal parts, then optimizes low-energy multivariate trigonometric polynomials to best fit the data on each part. The resulting problem reduces to generalized ridge regression, which can be solved efficiently. The authors demonstrate the framework on synthetic and real datasets, including video interpolation, attractor recovery in quasiperiodic dynamical systems, and parameterizing gene-space cycle data for parasites and humans.
In briefOpen-HumanScale: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models
SummaryThis paper presents Open-HumanScale, a fully open recipe for scaling egocentric human video pretraining to 50,000 hours, along with the dense hand-motion and language-annotated corpus Open-HumanScale-50K. The authors report an empirical log-linear relationship between pretraining data hours and downstream robot success rates. They build HumanWAM-Vanilla, pretrained only on human video, and HumanWAM-Goal, pretrained with joint human-robot data, and report results on RoboTwin 2.0 and RoboDojo benchmarks. Data, pipelines, checkpoints, and evaluation suites are open-sourced.
In briefSAUCE: A Fact-Based Method for Checking Conversation Summaries
SummaryThis paper introduces SAUCE, a reference-free, fact-based pipeline for evaluating cross-lingual summaries of conversational speech. It measures summary accuracy and fact coverage by checking entailment between the conversation and the summary text. Unlike black-box LLM evaluators or dense embedding metrics, SAUCE maps scores to discrete, verifiable facts, so exact hallucinations or omissions can be identified. Preliminary investigations show that its results align strongly with human judgment.
In briefEvolving interpretable spatial descriptors to uncover structure-property links
SummaryThis paper presents a framework for building spatial descriptors that predict local spectra and hysteresis loops from images. The authors benchmark regression models, use SHAP maps to examine spatial attribution, and study how patch size and compression affect accuracy and descriptor dimensionality. A genetic algorithm is then used to optimize spatial partitions and the choice of mean, minimum, and maximum pooling, with Cartesian and polar constructions compared.
In briefLLMs Don't Know What They Know: A Benchmark for Testing Whether Models Can Judge Their Own Knowledge Dates in Financial Backtests
SummaryThe paper introduces FinDendro, a benchmark of 2,226 dated U.S. facts from 1999 to 2026, including macroeconomic series, month-end stock prices and Federal Reserve decisions, to test whether language models accurately judge what they know. For most frontier models, self-reports understate actual knowledge; for example, GPT-5 answers 76% of controlled facts correctly but claims to know only 38%. The authors conclude that self-report, instructions, or stated knowledge cutoffs cannot certify that a backtest within a model's training window is free of leaked outcomes.
In briefMeasuring Risk, Not Harm: Detecting Signs of Psychological Harm Risk in AI Chats
SummaryThis paper proposes monitoring psychological harm from large language models by detecting observable risk-factor behaviors in AI responses rather than measuring harm directly. The authors define eight such behaviors and build a synthetic multi-label corpus of 47,062 AI responses. They compare embedding-based classifiers, direct LLM judging, and a classifier-judge cascade, with the best cascade reaching an F1 of 0.79 on held-out human-labeled data.
Nil-Jana Akpinar, Eugenia Kim, Natalie Isak et al.
In briefEfficient Agent Evaluation Using Cached Execution Traces
SummaryThis paper proposes estimating an AI agent system's full benchmark score from only a few task executions. Each execution trace is summarized by an LLM and embedded with a text encoder, preserving correctness- and task-related signals. On SWE-bench Verified and Terminal-Bench 2.0, mean absolute error drops from 0.13 to 0.08 and from 0.14 to 0.12 when only one random task is executed.
In briefConDR: Conservative Reachability via Neural Hamilton-Jacobi Subsolutions
SummaryThis paper addresses safety reachability analysis for dynamical systems, where DeepReach can misclassify truly unsafe states as safe. The authors propose Conservative DeepReach (ConDR), which learns a subsolution of the Hamilton-Jacobi-Isaacs variational inequality so that the learned unsafe set contains the true one, and prove its value lies within ε·τ of the true value function. Compared with DeepReach and MADR on Air3D, Air6D, a 9D multi-vehicle system, and a 13D quadrotor, ConDR reports the lowest misclassification and closed-loop collision rates.
Kendra Givens, Taylor T. Johnson, Bryan C. Ward et al.
In briefMessage-passing graph persistence: topological descriptors that incorporate GNN message passing
SummaryThe paper introduces Message-Passing Graph Persistence (MGP), a family of topological descriptors that integrates message passing into graph filtrations. At each filtration step, node embeddings are updated on the current subgraph and linked to persistence events, and they also determine representatives of connected components. The authors prove MGP is invariant under graph isomorphism and strictly more expressive than the persistent homology descriptors it extends. Experiments on an expressivity benchmark and several graph classification and regression datasets show MGP outperforms other PH-based descriptors on most of them.
Jorge Luiz Franco, Arthur Queiroz Moura, Moacir Ponti et al.
In briefVA-TDM: A unified framework improving few-step diffusion models via velocity adversarial distillation
SummaryThis paper introduces Velocity Adversarial Distillation (VAD), which uses a learned critic to compare velocity fields of a student and teacher model and turns their disagreement into training feedback. Building on it, VA-TDM provides supervision for partial denoising along the student's trajectory. The method achieves state-of-the-art four-step results in text-to-image and text-to-video generation across several backbones and scales to a 33B MiniMax-H3 model.
In briefTurning integer grids into spatial intuition: perception, not planning, drives ARC-AGI-3 performance
SummaryThe paper presents SARBLOH, an open-weight agent for ARC-AGI-3, a benchmark of uninstructed visual games. Its zero-parameter perception engine converts each frame into an object inventory, a geometric briefing, and an annotated canvas, and the agent writes code to express hypotheses. The authors report that across eight presentations the rubric score rose from 3/28 to 25/28 while token use fell by 89.4%.
In briefTwo-Ledger Model: Per-Transaction Cost Accounting for Multi-Agent AI
SummaryEnterprises run business transactions through multi-agent AI systems across clouds, LLMs and systems of record, but finance only sees AI costs aggregated at settlement level. The paper proposes the Two-Ledger Model, which joins FOCUS-conformant billing data with invocation-level usage records built on OpenTelemetry GenAI telemetry. A layered correlation method assigns each record to its business transaction, and simulations show cost coverage above 99% even when trace propagation drops to zero.
In briefWhether Two Paper Acceptance Predictors Agree on Which Paragraph Edits Raise Scores
SummaryThe paper compares a fast classifier with two released text-only paper acceptance predictors on 198 ICLR papers and their paragraph rewrites. The classifier agrees with the predictors on ranking original papers, but shows no detectable agreement on which rewrites raise scores. A full-precision re-check gives the same result, with only a tiny average gain over random choice. The authors recommend checking revision selectors against the intended evaluator on the revisions they rank.
In briefFedFiOT: Fisher-weighted optimal transport for merging models in one-shot and few-round federated learning
SummaryThe paper addresses problems in one-shot and few-round federated learning, where client models drift apart and neurons become misaligned. It proposes FedFiOT, which uses Fisher information to weigh parameter importance and guide neuron alignment and aggregation. The method jointly optimizes neuron correspondences and global parameters by alternating minimization of a shared Fisher-weighted objective. The authors report consistent improvements over baselines across multiple datasets in both settings.
Quan Minh Nguyen, Hoang M. Ngo, Bhakti Chowkwale et al.
In briefSemantic Residual Vector Quantization for Sign Language Generation (paper)
SummaryThis paper proposes Semantic Residual Vector Quantization (SRVQ), a framework that generates sign language motion from written sentences in two stages. It first encodes the sign content at a semantic level, then refines pose, handshape, trajectory, and body coordination in later residual levels. On CSL-Daily and Phoenix-2014T, the authors report R@1 scores of 82.3 and 70.3, respectively, which they say exceed the previous state of the art by 30 and 46 points.
In briefBig Models, Small Gains: Frozen Forecasters as Classifiers on MONSTER
SummaryThis paper benchmarks two time series forecasting models, Reverso and Chronos-Bolt, as frozen feature extractors for classification on six large MONSTER datasets, using 1%, 10%, and 100% of labels. Pretraining beats random-weight versions on five of six datasets, but the frozen representations do not reliably outperform the strongest non-pretrained baseline. The authors conclude that strong task-specific baselines and random-weight controls are important when evaluating forecasting foundation models for classification.
In briefA benchmark testing whether AI models can spot fine details in drone images for city management.
SummaryThis paper introduces UMBench, a fine-grained urban remote sensing benchmark built from UAV images for tasks such as traffic monitoring, urban construction, environmental quality, and disaster warning. Evaluations show that current multimodal large language models often rely on scene-level priors and overlook fine-grained visual evidence. The authors propose a training-free, evidence-first chain-of-thought method that directs models to inspect such evidence. Experiments report that this method improves the performance of both open-source and commercial models on UMBench.
In briefRoute First, See Later: separating routing policy from action spaces so vision agents reason scientifically
SummaryThe paper introduces ZETA, an architecture that separates routing policy from action execution in vision agents for scientific image analysis, so routing errors and execution failures can be evaluated independently. On a zero-shot benchmark of 51 scientific-intent queries, explicit routing semantics substantially improved routing accuracy, especially at semantically adjacent action boundaries, though gains varied across models. The authors report that for Qwen 3.8, segmentation routing accuracy rose from 6.2% to 75.0% and detection routing accuracy from 60.0% to 100%.
In briefContinual Classification Beyond the Kernel Regime: Using L2 Regularization's Sparsity Bias in Diagonal Linear Networks
SummaryThis paper studies continual logistic regression with linear predictors in a high-dimensional statistical setting. It derives the optimal L2 regularization strength for continual linear classification and characterizes its generalization performance. Using a minimal two-layer diagonal linear network, it shows that L2 regularization promotes sparse feature learning in single-task settings but loses this sparsity bias in continual learning, and proposes a weighted L2 variant that restores it. The authors report that, for sufficiently sparse teachers, this variant converges substantially faster over moderate task horizons.
