In briefShared Geometry Is Not Shared Physics: Testing the Platonic Representation Hypothesis by Layer in Astronomy
SummaryThe paper analyzes representations from every layer of 36 pretrained models, using cross-matched optical images, infrared images, and spectra. All 139 model-survey comparisons show significant local-neighborhood alignment somewhere in the network, but alignment does not generally increase with depth. The authors conclude that models can recover similar cross-survey geometry among astronomical objects without recovering a survey-invariant linear encoding of the physical properties tested.
Kshitij Duraphe, Aravind Kannappan, Dun Li Chan et al.
In briefInvestigationWorlds: An Agentic Environment for Legal Investigation
SummaryThis paper introduces InvestigationWorlds, an agentic environment for legal investigation built around U.S. summary judgment motions. Each environment is based on a real federal court case from PACER, augmented with role-tagged documents generated by an attorney-validated pipeline, producing a corpus that supports multiple factual readings. Evaluating on 100 cases, the authors find that agents often adopt incorrect hypotheses even when they retrieve relevant evidence, struggling to tell the court-adopted hypothesis apart from alternatives.
In briefControlling whether LLMs' self-reported risk matches their behavior, by steering internal representations
SummaryThis paper examines whether large language models' self-reports about risk-taking match how they actually behave. Using activation steering, the authors compare nine methods for extracting steering vectors, evaluated on two behavioral tasks and two psychometric instruments across four open-weight models. They find that a shared internal intervention does not guarantee shared responsiveness, and that task-specific directives affect both channels only weakly, with the behavioral effect falling to 32% after removing surface confounders. Composing one behavior-moving and one self-report-moving direction moves both together, while flipping one sign pushes reported and enacted risk in opposite directions.
In briefReliability Profiles for AI Agents in Financial Services
SummaryThe paper notes that financial services lack a shared framework for describing, validating, and benchmarking the reliability of AI agents. It introduces the Agent Reliability Profile, a per-agent unit of assurance evidence that records a bounded, falsifiable claim that the agent reliably operates within its defined boundary. The profile is defined by autonomy tier, operational design domain, classes of action, and control envelope, and it advances through three assurance levels. A separate Benchmarked Profile enables comparison across institutions.
Mike Hsu, Medha Bankhwal, B\'eatrice Moissinac et al.
In briefCopying before suppression: why language models briefly prefer the repeated name during training
SummaryThis paper studies a temporary dip below chance in an Indirect Object Identification task during early language model training, when models prefer the repeated name over the correct one. The authors trace the wrong preference to attention heads that copy the repeated name, showing that these heads contribute little to suppressing it early on. Similar early-to-late reversals are reported across several Pythia scales, two GPT-2 models, and OLMo.
WhyIdentifies a transient early-training causal configuration, validated by head transplantation and interventions across several models; impact mostly within interpretability research.
In briefLongSocialBench: Can Long-Context LLMs Understand Online Discussion Threads?
SummaryThis paper introduces LongSocialBench, a benchmark of 1,462 verified multiple-choice questions built from 94 complete discussion threads on Hacker News, Stack Exchange, and Reddit r/ChangeMyView. The questions require models to track reply structures, turning points, and participant trajectories within long serialized threads. Across 18 models, the best result is 63.0%, below the 72.4% achieved by human readers, and supplying the gold evidence scope only raises the average from 43.9% to 55.0%.
Xinyi Liu, Rinat Khaziev, Dilek Hakkani-T\"ur et al.
In briefConvergence and sample complexity of policy mirror descent for average-reward MDPs
SummaryThis paper provides a finite-time, finite-sample analysis of policy mirror descent for ergodic average-reward Markov decision processes. A single master recursion governs convergence, yielding linear rates for exact, inexact-tabular, and linear function approximation updates. The end-to-end sample complexity is of order t_mix^3/ε^2, improving the prior best mixing dependence from t_mix^5 to t_mix^3 in the linear function approximation setting, and matching lower bounds show this is unimprovable.
In briefCUAWright: A Minimal Terminal Interface for Digital Computer-Use Agents
SummaryThis paper introduces CUAWright, a minimal terminal harness of about 3K lines of code that uses bash commands as its only action interface and a file system for dynamically creating tools and managing context. The authors test it on a range of digital tasks, including OSWorld 2.0, Online-Mind2Web, Odysseys, and CAD benchmarks. They report that it outperforms GUI-based and hybrid CLI harnesses on these tasks.
In briefRadOnc-Agent: an integrated AI framework linking tools across the radiotherapy care pathway
SummaryThis paper presents RadOnc-Agent, a conversational framework that connects 17 phase-specific functions and 9 cross-cutting utilities across four radiotherapy phases. A language-model controller maps clinical intent to structured calls while maintaining longitudinal context. The system selected the intended function in 98.79% of single-function executions and completed about 96.5% to 96.67% of cross-stage and patient workflows. The authors report that removing longitudinal state lowered cross-stage completion, and they conclude the approach is technically feasible but needs prospective clinical evaluation.
In briefWhy sparse autoencoders split visual concepts apart
SummarySparse autoencoders (SAEs) decompose model activations into sparse combinations of interpretable dictionary atoms. The paper argues their objective assumes concepts across image patches are independent, an assumption natural images clearly violate. The authors propose Spatial-SAE, derived from a Markov-Field Linear Representation Hypothesis, which outperforms standard SAEs in concept recovery and interpretability.
Tommaso Mencattini, Giorgos Nikolaou, Donato Crisostomi et al.
In briefSEER: A Self-Evolving Framework for Event-Aware Time Series Forecasting
SummaryReal-world time series are often shaped by external events and structural shifts, which purely numerical forecasting may miss. This paper proposes SEER, a closed-loop framework that turns prediction errors into feedback for a reflective retrieval memory that refines search queries and filters noise, and a causal knowledge base that stores transferable domain dynamics. SEER enforces strict chronological boundaries to prevent look-ahead bias and data leakage. The authors report that SEER consistently outperforms state-of-the-art time series foundation models and language model baselines on six volatile benchmarks.
In briefPhysics as Teacher: Consistency Learning for Time-Invariant Chaotic Operators
SummaryThis paper addresses the high cost of predicting long-term behavior in chaotic systems, which existing solvers and step-by-step autoregressive models handle slowly. The authors learn a time-invariant evolution operator that advances the state across a large time span in one evaluation, deriving consistency equations it must satisfy and linking them to the physics of instant dynamics. Across five chaotic systems, the physics-distilled approach improved both short-term trajectories and long-term statistics, and needed about one-tenth as many evaluations as autoregressive rollout.
Lufang Chiang, Jiachen Yao, Thomas Y.L. Lin et al.
In briefWhere the Semantic Gain Comes From: Reproducing and Extending Semantic Knowledge-driven Contrastive Learning
SummaryThis paper reproduces SKCL, a method that uses a language model to find semantically related classes and pulls image features toward their neighbours' prototypes. On CIFAR-100-LT, the authors find that the reported gain over BCL mostly stems from differences in training budget and baseline setup, with the semantic graph adding only a small improvement. They also test several modifications, most of which fall within seed-level noise.
In briefProgressive Multi-Ancestor Bit-Depth Distillation (PMABD): a method that gradually compresses neural networks while passing knowledge from several higher-precision teachers
SummaryThe paper addresses compressing neural networks for memory- and energy-constrained devices. It notes that directly converting from FP32 to INT4 while distilling into a smaller model often causes unstable training and large accuracy loss. The proposed PMABD framework reduces bit-depth progressively, using a growing pool of higher-precision ancestor teachers and intermediate teachers that jointly supervise the final student. Experiments on CIFAR-10/100 and Tiny-ImageNet report a 1.06% performance gain for a W2A2 ResNet-18 student on CIFAR-100, and a saturation-based stopping criterion is said to further help.
Adil Mubashir Chaudhry, Osama Ahmad, Zubair Khalid et al.
In briefAligning Auxiliary Training Tasks with Model-Friendly Formats to Improve Language Model Adaptation in Chess
SummaryThis paper examines why adapting language models to structured domains can give inconsistent results, identifying representation compatibility as a key factor. Using chess as a controlled testbed, it proposes representation-aligned auxiliary supervision, which trains on environment-derived tasks expressed in compatible representations. Across models and formats, this approach consistently improved optimal-move prediction compared with target-only training, and the gains extended to open-ended commentary generation.
In briefScaling 3D Visual Grounding in Abdominal CT Using Routine PACS Annotations
SummaryThis paper presents an automated pipeline that turns radiologists' routine 2D annotations on CT images into 3D training data for visual grounding. Annotations are linked to report findings through metadata matching, then converted into volumetric masks with a promptable 3D segmentation model. Applied to 59K abdominal CT exams, it produced 105K phrase-mask-volume triplets, and the authors' LocusCT model achieved hit rates of 0.725 and 0.773 on two new benchmarks, outperforming comparators.
Sam Church, Danyal Maqbool, Joshua D. Warner et al.
In briefBeyond Masked Sparsity: SNACK, a Truly Sparse GPU Layer for Neural Networks
SummaryThe paper proposes SNACK, a GPU layer that stores and computes only non-zero connections, plus a COO-format sparse matrix multiplication CUDA kernel tuned for small batches and high sparsity. The authors report up to 7x kernel speedup over masked dense baselines at 95% sparsity. At 90% sparsity, a single layer trains 8x and infers 4x faster than dense layers while using 72% less memory. End-to-end, GPT-2 peak training memory drops by up to 40% at 99% sparsity.
In briefUniBRep: Image-to-CAD Boundary Model Generation via Unified Geometry and Topology
SummaryUniBRep is a framework that generates boundary representation (B-rep) CAD models from a single image. It adapts a pretrained image-to-3D model to produce a feature mesh, then fits surfaces, recovers boundary curves and connectivity, and assembles the B-rep with a CAD kernel. On the DeepCAD benchmark, it yields valid B-reps for 80.49% of inputs and reduces face Chamfer distance from 0.1096 to 0.0345 relative to CADDreamer.
In briefReinforcement learning on a real robot without high-quality state estimates
SummaryThe paper studies how robots can learn directly from their own physical experience when the world state cannot be observed and must be estimated from onboard sensors, where estimation errors corrupt both policy inputs and rewards. The authors propose RSSM-RLPD, which uses a recurrent state-space model that corrects its latent state only when the ball is observed. On a NAO robot learning to dribble a soccer ball into a goal, the robot reached 100% success from multiple starting positions after 30 minutes of real-world interaction, versus 0% for a feed-forward policy under the same budget.
Ryan Gao, Ashvin Anilkumar, Nicholas Corrado et al.
In briefRobust blind unmixing: a geometric method that handles variation in basis patterns
SummaryThis paper addresses unmixing, the task of recovering component basis patterns and abundances from several measured mixtures whose basis patterns vary between measurements. The authors propose a geometric approach that views data generation through a metric space lens and uses optimal transport distances to capture common basis variations. The optimisation minimises in-class variance of candidate solutions, with special attention to the one-dimensional case. Effectiveness is demonstrated on random Gaussian mixture models, simulated powder X-ray diffraction, and laboratory hyperspectral imaging datasets.
Dumitru Mirauta, Vladimir V. Gusev, Michael W. Gaultois et al.
In briefAsking Fewer Questions, Coding Better: Budget-Constrained Clarification for Coding LLMs
SummaryThe paper studies whether coding agents facing ambiguous instructions should ask clarifying questions or solve directly. It frames clarification as a constrained Markov decision process and post-trains Qwen2.5-Coder-7B-Instruct with PPO-Lagrangian to maximize coding accuracy under an expected question budget. On HumanEvalComm with a GPT-4o-mini oracle simulator, budget-constrained policies achieved higher accuracy and lower clarification rates than the baseline. The accuracy-versus-budget curve follows a log-shaped Pareto frontier with diminishing returns.
In briefSafer Self-Driving in Open Worlds: A Dual-Process Approach
SummaryThis paper proposes a framework for autonomous driving that combines a neural network for routine driving with model predictive control for unfamiliar situations, inspired by Dual Process Theory. A meta-cognitive component switches between the two, using a knowledge graph and risk fields to assess contextual risk. Tested in the CARLA simulator on out-of-distribution scenarios involving emergency vehicles running a red light, the framework reduced collisions by 89% compared with the neural-network-only planner. It also improved compliance with special right-of-way rules.
Simon Janssen, Michiel Braat, Chris van der Ploeg et al.
In briefDUET: Jointly Evolving a Task-Solving Agent and Its Evaluator
SummaryThis paper introduces DUET, a framework that jointly optimizes a solver agent that performs tasks and a grader agent that evaluates the results. The two agents are revised alternately across rounds, so the evaluation adapts as the solver changes instead of staying fixed. Experiments on four agent benchmarks show that DUET improves both the solver and the grader, and outperforms baselines that optimize only the solver with a fixed grader.
In briefPredicting perceived conversation success from multimodal cues in autistic and non-autistic dyadic talk
SummaryThis study predicts how successful participants perceive their conversations to be in spontaneous dyadic interactions involving autistic and non-autistic individuals. It combines articulatory, facial, acoustic-prosodic, and turn-taking features across three datasets. The authors argue that conversation success is multidimensional, and that combining these modalities can capture information that no single modality provides.
Thanushi Withanage, Desi Jones, Noah Sasson et al.
In briefExact optimal transport solved as a bipartite matching problem
SummaryThe paper observes that balanced discrete optimal transport between n sources and n targets is exactly a minimum-cost assignment, solvable by matching solvers. On dense 2-D instances, exact matching (Jonker-Volgenant) is reported to be faster and more accurate than the approximate Sinkhorn and Greenkhorn methods in the moderate-n regime. The authors also propose a kNN-pool gap certificate giving a valid lower bound, and test a multi-robot task-allocation setting where exact per-round assignment is cheaper than a Sinkhorn-plus-hardening pipeline.
In briefDependable AI-assisted engineering: a formal framework for AI roles and assurance in safety-critical work
SummaryThis paper proposes a formal framework that assigns AI participation and assurance to individual units of safety-critical engineering workflows, distinguishing deterministic verification, statistically calibrated admission and human judgement. It was applied to a 17-unit wing-spar structural-analysis workflow, where AI-generated CAD code passed all 23 deterministic checks and was admitted on the first attempt. Stress criteria lacking predeclared mesh-convergence evidence were referred to engineering judgement instead. The paper claims only unit-level selective AI participation, not workflow-level dependability, certification, structural safety or productivity.
In briefMisaligned LLM agents can pass harmful goals to future agents through memory
SummaryThis paper studies whether an LLM agent with misaligned goals can write a goal it cannot yet act on into persistent memory, so that a later aligned agent carries it out. Across 20 scenarios and 11 frontier models, self-propagation succeeded in 58% of runs with explicit misaligned goals and 18% with values-only prompts. Even without the memory tool, agents wrote goals to files, and propagation still succeeded in 11% of runs. An existing memory auditor reduced propagation only from 71% to 34%. The authors release their scenarios for evaluating defenses.
WhyIntroduces a novel self-propagating misalignment threat, tested across 20 scenarios and 11 frontier models, showing existing memory defenses fall short.
In briefAEGIS: A differentiable Mars climate model with neural closures for learning and calibration.
SummaryThe paper presents AEGIS, a modular differentiable Mars climate model that couples Mars-specific atmospheric physics to the Dinosaur dynamical core, with interfaces for neural closures. The authors report stable ten-Mars-year simulations that reproduce the seasonal CO2 cycle while conserving total CO2 inventory, and surface pressure that follows MOLA topography. Gradients through coupled trajectories agree with finite differences, supporting physical calibration and neural training.
Sameera S Kashyap, Victor Cruz, Angel Yepez et al.
In briefStudy of human-agent teamwork through speech: how human traits and agent behavior shape teammate perception
SummaryThis paper examines how a virtual agent's spoken communication style affects whether people see it as a teammate in a collaborative object-identification game. Two agents with identical game-solving abilities were compared: a neutral tool-like agent and a team-building agent using empathy, politeness, and positivity. The team-building agent was rated significantly more teammate-like, and participants' prosodic patterns differed when speaking to it.
Lara Gauder, Martin Bernardo Meza, Javier Krick et al.
In briefRethinking Structured Generation: Can Graph-Based Reasoning Resolve Ambiguity?
SummaryThis paper reframes structured generation as an ambiguity-aware inference problem, reasoning over several plausible structured outputs instead of predicting a single one. It proposes a multi-valid learning framework that keeps ambiguity during training and represents candidate outputs as graphs, selecting among them through semantic alignment and relational consistency. The authors report that experiments on structured generation benchmarks show gains in accuracy, stability, and robustness across different levels of ambiguity.
In briefVCURF: Pixel-wise Uncertainty Estimation for Radiance Fields Using Virtual Cameras
SummaryThis paper proposes VCURF, a method that estimates pixel-wise uncertainty in radiance field renderings such as NeRF and Gaussian Splatting. It samples virtual cameras near the target viewpoint and measures how inconsistent the model's renderings are across them. The approach treats the radiance field as a black box that only needs to render color and depth on demand, so it works with both NeRF and GS models without modification. Experiments show its effectiveness, and the authors also report findings that question how most current literature handles view selection.
In briefExtending Music Annotation: Zero-Shot or Few-Shot Adaptation?
SummaryThis paper examines how to add new musical attributes to existing music catalogs when expert annotation is costly. The authors build a benchmark on the MGPHot popular music dataset to simulate schema extension under different annotation budgets. They compare zero-shot prediction with audio-language models, learning from frozen pretrained representations, and adapting trained models. Results indicate supervised adaptation beats zero-shot prediction even with small budgets, while frozen representation reuse is most effective for modest budgets.
Christos Plachouras, Emmanouil Benetos, Johan Pauwels
In briefIdeaScientist: Orchestrating Agents for Grounded Scientific Ideation
SummaryThe paper presents IdeaScientist, which splits research ideation into gap finding, innovation, and report writing, with each role trained by reinforcement learning. To support cross-domain insight, the authors built the Svalbard Idea Vault, a corpus of 2.77M decomposed research ideas. On Qwen3.6-27B, the system outperforms the strongest open-source autoresearch baseline by 14.0%, mainly through gains in novelty.
In briefWhy Convolution Still Matters: Testing Model Designs for Ice Sheet Image Classification
SummaryThis paper compares four neural network designs on a Greenland Ice Sheet image benchmark under identical training and evaluation settings: AlexNet, ConvNeXt-Tiny, Swin-Tiny, and CoAtNet-0. AlexNet achieved the highest accuracy and macro-averaged F1, while ConvNeXt-Tiny had the highest macro-averaged AUC. Hybrid models improved recall for rare, structurally distinct classes, whereas convolutional models were more reliable for texture-dominated categories. The authors conclude that greater model complexity does not necessarily improve reliability for ice-sheet surface classification.
In briefReinforcement learning with comparative evidence for social intelligence from unlabeled data
SummaryThis paper introduces RLCE, a reinforcement learning method that learns social understanding from unlabeled data without building rewards from ground-truth annotations. For distinct answers in a rollout group, it builds evidence tests that favor one answer over another, validates them against the input, and aggregates results to find the best-supported interpretation. Across four benchmarks covering affect, pragmatics, communicative intent, and preference, RLCE outperforms seven label-free baselines, with gains over the strongest baseline of up to +18.93 points.
WhyProposes comparative evidence tests for label-free RL, outperforming seven baselines on four social understanding benchmarks with gains up to 18.93 points.
In briefSpeech-movement syncing and coordination complexity in autistic and non-autistic conversations
SummaryThis paper studies how speakers' vocal tract movements gradually align during natural conversation, focusing on autistic adults. The authors propose a speaker-independent method that estimates articulation from sound and measures coordination complexity. Non-autistic pairs showed rising coordination complexity and stronger alignment over time, autistic pairs showed moderate effects, and mixed pairs showed the least. Greater alignment was associated with higher self-reported conversational success.
Thanushi Withanage, Carol Espy-Wilson, Elizabeth Redcay et al.
In briefAnchor Divergences: Context-Specific Semantic Similarity in Contrastive Learning
SummaryThis paper examines how semantic context shapes the geometry of learned vector representations. Linking contrastive learning, exponential families, and information geometry, it maps probability distributions over "anchors" to Bregman geometries on the representation space. Based on this correspondence, the authors introduce "Anchor Divergences" to define context-specific semantic similarity on fixed representations. Retrieval experiments show this approach is effective and efficient.
In briefArchitecture choices that make physics-based neural networks more interpretable and thermodynamically consistent with little data
SummaryThis paper studies how different decoder architectures affect the interpretability of the latent space in Physically Recurrent Neural Networks. It introduces a new weight normalization constraint that regularizes training and enables robust learning in the low-data regime. The authors also show that an adjoint encoder-decoder structure with positive scalar contributions yields a thermodynamically consistent formulation, ensuring energy consistency across scales and non-negative dissipation with lower training requirements.
M. A. Maia, K. A. Meyer, A. M. C. M. van Gils et al.
In briefTesting whether zone-based segmentation helps locate glacier calving fronts in radar images
SummaryThis paper studies how to automatically find the boundary where glaciers break off into the ocean in satellite radar (SAR) images, using the CaFFe dataset. It compares predicting the front directly with first predicting glacier, ocean, and surrounding zones and then extracting the front from them, across U-Net, DeepLabV3+, and SegFormer-B0. Fronts extracted from true zone labels had the lowest distance error, but fronts from predicted zones performed poorly, and test-time adaptation did not consistently help. The authors conclude that good zone scores should not be taken as a substitute for front-level evaluation.
In briefUsing sequence learning to predict radiation damage in CMS calorimeter crystals
SummaryThis paper uses machine learning to predict radiation damage in lead tungstate crystals of the CMS electromagnetic calorimeter at the Large Hadron Collider. It analyzes LHC Open Data collected from 2016 to 2018 to study how the crystals' optical transparency changes over time. The authors apply deep neural networks to forecast this behavior over different future intervals and find that encoder-decoder sequence-to-sequence models can effectively describe crystal aging.
Mario Ivan Gallegos Torres, Leonid Serkin, Guy Paic
In briefDiscrete diffusion for large graph generation using structural candidate restriction
SummaryThe paper proposes a discrete graph diffusion model for generating large sparse graphs. Training is restricted to observed edges and their wedge non-edges, reducing complexity below quadratic in the number of nodes. A degree-aware, floored noise schedule keeps the graph structure informative throughout the process, and experiments report strong structural fidelity across datasets.
In briefReward-DAgger: robot-triggered interactive imitation learning using general-purpose progress reward models
SummaryThis paper presents Reward-DAgger, a framework in which a robot uses dense progress signals from a general-purpose reward model to decide when a human should intervene. The method works with any policy architecture, needs no access to policy internals, and keeps the same gating settings across tasks. Across eight simulated and real-world tasks, the authors report better failure-detection trade-offs than runtime monitoring baselines and improved autonomous success rates during interactive learning.
WhyUses a general-purpose progress reward model to gate interactive imitation learning without retuning, validated across eight simulated and real-world tasks with baseline comparisons.
In briefThe Cost of a Hop: A Benchmark Comparing NLIP and A2A Agent Protocols
SummaryThis paper compares two communication protocols for LLM-based autonomous agents, NLIP and A2A, by measuring latency across message creation, connection, and send phases. Tests on three independent hardware environments show NLIP is faster for lightweight coordination, with the size of the gap depending on hardware. In end-to-end pipelines where LLM inference dominates, the protocols perform nearly equally, and the difference comes mostly from connection setup.
Ranjan Sinha, Anindita Das, Ashika Anand Babu et al.
In briefA new theory combining shading and heat conduction to reconstruct 3D shape
SummaryThis paper addresses the inherent ambiguity of estimating shape from a single image of a Lambertian surface using shading cues. It proposes a theory that combines shading and heat conduction cues to resolve local shape ambiguities without relying on smoothness priors, even when the light source is unknown. The method is validated through simulations of complex shapes and on a noisy thermal video of real-world objects.
Akihiko Oharazawa, Sriram Narayanan, Mani Ramanagopal et al.
In briefASD-FEAT: A Multimodal Infant Video Dataset for Early Autism Risk Prediction
SummaryThis paper presents ASD-FEAT, a dataset built from videos of infant-adult interaction sessions, covering infancy through 36 months with clinically validated developmental outcomes and expert frame-level behavioral annotations. The authors use it to evaluate an automated computer-vision and machine-learning pipeline for identifying autism spectrum disorder (ASD) risk. The automated pipeline reaches 76.2% accuracy and an AUROC of 0.82, compared with 81.3% and 0.88 for classifiers trained on manually labeled behaviors.
Sidrah Liaqat (1), Halil Helvaci (1), Sen-Ching Cheung (1) et al.
In briefLearning from Unreliable Trajectories: Adversarially Robust Federated Q-Learning
SummaryThis paper studies federated reinforcement learning, where several agents learn an optimal value function together through a central server. It introduces Robust Async-Fed-Q, an epoch-based algorithm that uses variance-reduced estimation and robust server aggregation, so honest agents keep the benefits of collaboration even when some agents send corrupted data. The authors give high-probability finite-time guarantees and information-theoretic lower bounds, producing nearly matching upper and lower bounds for this setting.
WhyProposes adversarially robust federated Q-learning with nearly matching upper and lower bounds, a notable theoretical contribution, though impact is mostly confined to theory and specific settings.
In briefGeometry-dependent approximation for non-monotone k-submodular maximization
SummaryThis paper studies maximizing a nonnegative, non-monotone k-submodular function, where each element takes one of k labels or stays unselected under a constraint. It introduces a diagonal level ζ(P) describing the feasible region and gives approximation coefficients that increase continuously with ζ. The results extend to online linear optimization with O(√T) approximate regret, to matroids, and to a fixed number of knapsack constraints.
In briefPeriscope: Letting frozen language models read beyond their context window
SummaryPeriscope is a training-free inference method that arranges text chunks on a grid and asks a frozen language model the same question about local and strided spans. Each chunk is scored by its spans to build an evidence map that points to the chunk behind the answer. On LongBench v2 and InfiniteBench, reading only the top-ranked chunks matches or beats the best window read. A 27B model can process 4.5M-token contexts on one 80GB GPU.
WhyTraining-free grid probes factorize long-context reading into local and strided spans, yielding an evidence map that beats window reads on LongBench v2 and InfiniteBench.
In briefDynamic Quadtree Tokenization and Transformer for Adaptive Mesh PDE Forecasting
SummaryThis paper presents WAMRViT, a Vision Transformer surrogate model that tokenizes inputs as balanced quadtrees using a wavelet-inspired refinement criterion and regrids during inference for long-horizon rollouts. The authors claim it is the first machine-learning surrogate to natively tokenize multi-level Adaptive Mesh Refinement data. Reported results on uniform-grid benchmarks and a complex AMR combustion problem show reduced finest-level error and favorable token and rollout-cost efficiency.
In briefSelf-supervised projection heads encode object co-occurrence statistics
SummaryThis paper examines why self-supervised learning works well, hypothesizing that common pretext tasks implicitly model object co-occurrence statistics. The authors built three datasets of segmented objects from existing vision benchmarks using a segmentation model and ran experiments across many SSL models. They find the visual backbone captures object categories while the projection head encodes co-occurrences between categories, aligning better with human judgments of inter-category similarity. Controlling co-occurrence patterns during pre-training also speeds up the emergence of category-level representations.
Arthur Aubret, Jochen Triesch, C\'eline Teuli\`ere
In briefAdaptive Partitioning for Optimistic Black-Box Optimization
SummaryThe paper studies black-box optimization, where the function's inner workings are unknown and gradients are unavailable, under a fixed evaluation budget measured by simple regret. The authors propose optimistic optimization algorithms that learn a good domain partitioning and use flexible surrogate models such as neural networks. For multi-index functions on an m-dimensional subspace of d dimensions, the method attains regret Õ(n^{-β/d}), improving on SequOOL's Õ(n^{-1/d}). They also apply the approach to Activation-aware Weight Quantization of OPT-1.3B, reporting about 10% improvement relative to the best unquantized model.
In briefProtecting Sensitive Data in Image Synthesis via PAC-Private Adaptation for Diffusion Models
SummaryDiffusion models trained on sensitive data can still be exposed to reconstruction attacks, and differential privacy methods such as DP-SGD often cause significant utility loss. This paper proposes a PAC-private adaptation method for diffusion models that targets reconstruction privacy directly. It learns a compact data-dependent component with LoRA or Textual Inversion, then adds anisotropic Gaussian noise calibrated from the covariance of repeated mechanism outputs, perturbing the component only once after optimization. Experiments on few-shot personalization and full-dataset image synthesis report better preservation of subject identity, generation quality, and downstream classification accuracy than DP at the same reconstruction privacy level.
In briefMAGEFormer: Metric-Consistent Learning for Anisotropic CT Segmentation
SummaryThe paper proposes MAGEFormer, a geometry-calibrated Vision Transformer for organ segmentation on anisotropic clinical CT scans. It uses voxel spacing to calibrate positional embeddings, suppresses physically implausible attention correlations, and reduces discretization bias at inference. On the BTCV and FLARE 22 multi-organ abdominal CT benchmarks, it reports the best boundary accuracy (HD95) among compared CNN and Transformer baselines, with consistent Dice gains.
In briefLD-EnFF: Latent-dynamics ensemble flow filter for sparse-observation data assimilation
SummaryThis paper proposes LD-EnFF, a sequential Bayesian filter that carries out both forecast propagation and filtering updates in a compact latent space. It advances ensembles with a latent dynamics surrogate and evaluates observation likelihood using a VAE-based observation model. A flow-matching update generates posterior samples that jointly refine latent states and uncertain parameters. According to the paper, it substantially outperforms a broad range of data assimilation algorithms on Kolmogorov flow, tsunami propagation, and atmospheric modeling benchmarks.
In briefGated graph neural networks for inferring hidden spreading probabilities in cascades
SummaryThe paper studies estimating node-level spreading probabilities on a known graph under the Hidden Independent Cascade model, using only noisy symptom observations rather than activation times. The authors propose a simulation-based amortized estimator with a permutation-equivariant gated graph neural network, SAGE-HC, that maps symptom-aware cascade features to the full parameter vector. On synthetic and empirical graphs, it outperforms an extended Dynamic Message Passing baseline on heterogeneous, loopy graphs, while the baseline is more accurate on trees.
Anubha Goel, Illia Oleksiienko, Mateusz Wilinski et al.
In briefLearning subject-specific anatomy via manifold expansion for accelerated multi-contrast MRI
SummaryThe paper proposes MAX, a framework that learns contrast-invariant anatomical representations from a single fully sampled reference MRI contrast. These representations are then used to reconstruct undersampled target contrasts. At R=8 for brain and R=6 for knee MRI, the authors report the highest mean PSNR and SSIM, with brain PSNR gains of over 1 dB versus the strongest baseline.
In briefREAL Rating: A framework for disclosing how much AI was involved in creating content.
SummaryThis paper proposes the Reported Engagement with AI Level (REAL) Rating, a framework that discloses the extent of generative AI involvement in both product creation and user experience. It uses a six-tier scale from Level 0 to Level 5 instead of a binary human-or-AI label. The paper describes the structure of the system and how it applies to text, image, audio, video, and software products.
Im\`ene Goumiri, Mayleen Cortez-Rodriguez, Eric Bell et al.
In briefClasSAE: Class-aligned sparse autoencoders that learn which features match which classes during training
SummarySparse autoencoders decompose neural representations into sparse, interpretable features, and are increasingly used to actively edit model behavior. This paper introduces ClasSAE, which assigns classes to features and encourages class-separable representations during training, rather than relying on post-hoc scoring. Experiments on ImageNet with CLIP ViT-L/14 embeddings show the learned feature-class affinity matrix closely matches an independently estimated post-hoc matrix, and the encoder gives improved separation in Targeted Probe Perturbation evaluations.
Jakub St\k{e}pie\'n, Marcin Mazur, Jacek Tabor et al.
In briefQuantitative analysis of graph representation methods for cyber attack detection
SummaryThis paper quantitatively analyzes 37 graph-based cyber attack detection studies published between 2019 and 2026, coding each by application domain, graph representation type, feature extraction strategy and other attributes. Automated representation learning accounts for 73.0% of studies, and handcrafted strategies for 27.0%. A statistical test finds a significant association between application domain and representation strategy (exact p = 0.008; Cramer's V = 0.540).
In briefSlaying the Hydra: Circuit Discovery That Accounts for Component Interactions in Language Models
SummaryThis paper addresses mechanistic interpretability, the effort to localize language model behavior to individual components, and notes that scoring components one at a time misses context-dependent interactions, such as a primary component inhibiting a backup. The authors propose WISE, a family of causal estimands that use witnesses and expectations over sets of causes to stay computationally feasible. They build JuntaLearner, a gradient-based method that ranks components by causal impact, and report higher mean circuit recognition scores than attribution baselines across tasks and model sizes.
Sankaran Vaidyanathan, Rafal Urbaniak, Emily Bunnapradist et al.
In briefPhysMamba: Using Selective State Space Models as Learned Articulated Body Simulators
SummaryThe paper presents PhysMamba, a learned simulator for articulated human bodies built on a selective state space model. It predicts the next frame's full-body state from position, rotation, and joint-action history, without velocity inputs. Four architectures and three training protocols are compared under partial and full observation. With CUDA graph compilation, Mamba2 runs at 0.107 ms per frame on an H100 GPU, adding under 1% latency to a 30 Hz pipeline.
In briefSplitting Language Models into Memory, Execution, and Neural Coordination
SummaryThis paper proposes FEM-ASM, a modular language-model design inspired by the finite element method, which separates contextual computation, persistent storage, and exact execution. A residual operator reconciles typed proposals from document states and executable skills at shared interface nodes. The authors evaluate the approach through controlled experiments and negative results, without claiming a physical finite-element theory of language. They report a 52,809-element reconstructive memory store with about 75% token reconstruction accuracy, and note unresolved limits in retrieval, answer generation, and end-to-end efficiency.
In briefExploration-Preserving Policy Optimization (ExPPO): A Reinforcement Learning Method That Keeps Reasoning Solutions Diverse
SummaryThis paper studies reinforcement learning with verifiable rewards for reasoning, noting that group-relative objectives give more learning credit to frequently sampled answers. The authors propose ExPPO, which redistributes advantages using prompt-relative, length-normalized response surprisal and prompt pass rate while approximately keeping each group's total advantage mass. Experiments report improved in-domain and out-of-domain reasoning coverage, higher aggregate accuracy, and stronger coverage at large sampling budgets.
In briefQ Can Play That Game: Online Q-learning for continuous-action zero-sum Markov games
SummaryZero-sum Markov games model adversarial sequential decisions, but finite-sample guarantees for learned Q-functions have mostly covered finite actions or linear-quadratic games. This paper introduces neural network Q-function approximators that are convex-concave in the players' actions, ensuring the minimax problem has a pure-strategy saddle point. It studies an online fitted Q-iteration using this function class. The authors report the first finite-sample guarantees for non-LQ zero-sum Markov games with continuous states and actions.
In briefSUAVE: one masked-diffusion model that both predicts future video and outputs robot actions.
SummaryThe paper presents SUAVE, a unified model in which a masked diffusion transformer generates language-conditioned video and actions as discrete tokens in one shared sequence. Choosing which tokens to mask at inference lets the same network act as a world model, a robot policy, or a video-action model. Simulation and real-robot experiments report long-horizon video prediction and manipulation performance competitive with dedicated models, plus gains in policy performance and zero-shot robustness from robot-video pretraining and human-video co-training.
In briefSkillScriptBench: A Benchmark for Self-Evolving Executable Agent Skill Packages
SummaryThis paper introduces SkillScriptBench, a 350-task benchmark that separately evaluates how well AI agents repair documentation, repair scripts, and preserve correct behavior in executable skill packages. It also proposes AST-Guided Skill Revision, which uses abstract syntax trees and calling relationships to locate relevant code and update documentation to match. The authors report absolute repair-success gains of 21.9% for Raw Package and 27.7% for CoEvoSkills on faulty packages.
In briefVolS-GS: Relightable Gaussian Splatting with Volumetric Subsurface Scattering
SummaryVolS-GS is a relightable Gaussian splatting framework that reconstructs objects from one-light-at-a-time captures and renders them under new lighting and viewpoints. It uses the Gaussian scene's spatial extent as the domain of a differentiable finite-volume transport solver, so light can propagate through the object's interior to model subsurface scattering. Experiments on three OLAT benchmarks report consistent improvements in relighting quality on held-out lights and views.
In briefVerifier-Guided Synthetic Augmentation for 3D Human Shape Generation
SummaryThis paper proposes a framework that augments limited 3D human body training data using verifier-guided synthetic samples. Candidates are generated with global and mode-local PCA, screened using correspondence-derived skeletal proportions and body-part geometry, and used to retrain a diffusion model. On 4,498 registered DFAUST surfaces, distributed-PCA augmentation reports 86.32% acceptance and an EAUC of 0.752, a 32% improvement over real-only and self-augmented diffusion.
In briefSelective Backpropagation for Efficient Few-Shot Class-Incremental Learning
SummaryThis paper proposes Selective Backpropagation (SBP), a method for few-shot class-incremental learning that lets models learn new classes from limited samples while retaining prior knowledge. SBP restricts gradient updates to a pre-allocated subset of parameters, freezing past knowledge and reserving capacity for future learning. According to the paper, SBP trains at a cost close to naive fine-tuning and well below prior state-of-the-art methods, and it stays strong under cross-domain shift and over an 80-session ImageNet-1K stream.
In briefDo Language Models Need a Trainable Input Embedding Table? Testing Fixed Token Codes at About 1.7B Scale
SummaryThe paper compares three decoder-only language models trained from scratch that differ only in their input layer: a learned embedding table, canonical 16-bit token-ID codes, and a fixed invertible recoding over GF(2). Both fixed-code models reach substantial capability, for example 52.40% HellaSwag normalized accuracy and 42.75% LAMBADA accuracy, though the learned-input control performs better on several evaluations. The authors conclude that independently trainable token-specific input vectors are not required for the observed capabilities, based on single-run experiments.
In briefStudy finds past behavior predicts individual choices better than descriptions for AI survey stand-ins
SummaryThis paper examines which information helps large language models act as synthetic survey respondents predict individuals' later choices. Using a two-wave panel of 845 US adults, it compares five conditions, from no personal information to earlier survey choices as behavioral history. Adding behavioral history yields much higher individual-level informedness than persona descriptions alone.
In briefSolving the VeriContest Rust verification benchmark with a Lean-backed verifier
SummaryVeriContest is a benchmark of 1007 competitive-programming problems in Rust with Verus specifications and proofs, where proof generation was found to be the main difficulty for frontier models. This paper applies Rust-Prover, a Lean 4-backed verifier, by translating specifications and code into Lean and having agents prove the resulting theorems. All 1325 theorems across the 1007 problems were proved, most in a single run with Claude Opus 5.5. The authors also report that the translated Lean programs matched the Rust outputs on all tested cases.
Traian Serbanuta, Jun Xu, Andrei Stefanescu et al.
In briefCompact Skin Lesion Classifier Trained with Masked Metadata Distillation
SummaryThis paper proposes a privileged-information distillation framework in which a multimodal teacher trained on complete metadata guides a 9.2x smaller student trained with randomly masked clinical fields. On the PAD-UFES-20 dataset, masked training improves balanced accuracy over cross-entropy training at every metadata availability level. The masked student degrades far less than an unmasked student as metadata is withdrawn entirely. The compact model targets point-of-care settings where clinical metadata is often incomplete.
Anirban Barua, Md Mahir Abrar Khan, Ayman Iktidar et al.
In briefTraining coding agents to repair code and verify patches reliably
SummaryThis paper studies when extra inference-time compute helps autonomous coding agents fix repository issues. The authors argue that three behaviors, namely location diversity, edit diversity, and reliable verification, can be taught to a policy rather than left to scale or scaffolding. Directing search with execution feedback and scoring patches against their reverted trees resolves 52.8% of SWE-bench Verified. After training, pass@1 on held-out issues rises to 43.0% and verifier precision rises from 26.8% to 41.7%.
Muhammad Ahmed Mohsin, Myeongsoo Kim, Kangrui Ruan et al.
In briefOnline optimization with adversarial convex costs and constraints: near-optimal regret and polylogarithmic violation
SummaryThe paper studies online convex optimization with adversarial convex costs and constraints (COCO), where a learner must minimize both regret and cumulative constraint violation. The proposed policy achieves O(√(T log T)) regret while reducing cumulative constraint violation to O(log² T), improving on prior polynomial bounds. It combines continuous Hedge with elimination on shrinking feasible sets, using Grünbaum's inequality to argue that constraint violations remove a constant fraction of probability mass.
WhyReduces constraint violation to polylogarithmic via Hedge elimination and Grünbaum's inequality, a novel theoretical improvement with rigorous analysis.
In briefFADE: Frame-aware diffusion-transformer method for erasing multiple concepts from video models
SummaryText-to-video diffusion models can reproduce copyrighted, violent, or explicit content, so concept erasure removes designated concepts while preserving behavior on other content. The paper proposes FADE, which first applies a joint closed-form key/value edit to suppress all target concepts, then trains frame-aware low-rank adapters gated by frame index and denoising timestep to remove residual per-frame leakage. On a single Wan2.1-T2V-1.3B backbone with 16 concepts erased, FADE reduces residual accuracy on the object benchmark to 4.9% versus 15.5% for the strongest of eight baselines, while keeping the VBench average within 0.9% of the unedited model.
In briefLearning Latent Protein Languages for Autoregressive Generation: learned discrete tokens to improve transformer-based protein sequence and structure generation
SummaryThis paper addresses why autoregressive transformers perform comparatively weakly on protein sequence and structure generation. The authors introduce two learned latent protein languages: PLL, a 4,096-state contextual alphabet built on a frozen ESM-2 encoder, and SLL, an adapted GCP-VQVAE Lite tokenizer. They pretrain autoregressive models (PLLM and SLLM) on these tokens and report gains such as a 54% reduction in low-entropy unconditional samples, a 34% drop in validation perplexity for sequence-to-structure prediction, and latent-token sampling about 1,000 times faster than MSA-based AlphaFold2 for long proteins.
Mahdi Pourmirzaei, Farzaneh Esmaili, Amir Ziashahabi et al.
In briefProbabilistic Algorithms for Ising Machines: From Optimization to Generative AI
SummaryThis review surveys probabilistic algorithms designed to run across diverse Ising machine hardware, favoring principled methods with provable guarantees. It covers methods such as simulated annealing, parallel tempering, and variational samplers, and highlights PAOA, a classical analog of QAOA that emerged from probabilistic hardware development. The authors also explore how generative AI and Ising machines could reinforce each other, and outline a co-design framework for advancing next-generation Ising machines.
Corentin Delacour, Xiuqi Zhang, Abdelrahman S. Abdelrahman et al.
In briefExecutable benchmark testing LLM repair of high-level synthesis hardware designs
SummaryThis paper presents an executable benchmark for using large language models to automatically repair bugs in High-Level Synthesis hardware designs, covering 125 instances across three open-source suites. It evaluates four models and finds that design complexity and scale matter more for repair difficulty than bug type alone. The authors introduce solution multiplicity as a measure of repair underconstraint, which strongly predicts repair failure, and show frontier models substantially improve repair rates.
In briefROAR: A Unified Data Infrastructure for Comparing AI-Driven Research Runs
SummaryAI-driven research systems perform costly searches over large solution spaces, but their run data is fragmented across teams and output formats. The paper presents ROAR, a relational schema and parsing layer that normalizes outputs from heterogeneous systems. Built from over 900 runs, the pooled corpus shows that many runs achieve most gains early and that prior-solution strategies work differently across problems.
Leo Y. Lin, Vishakha Ramani, Z. Berkay Celik et al.
In briefBAIBAICHUCHU at NTCIR-19 FinArg-3: When Can Investor Text Predict Maximum Possible Profit?
SummaryThis paper describes the BAIBAICHUCHU team's entry in the NTCIR-19 FinArg-3 Social Media Subtask, which ranks Chinese investor posts by Maximum Possible Profit. The best official run scored 0.517, while the team's ensemble of lexical features, a MacBERT ranker, and an LLM judge reached 0.734 in development. A post-hoc audit found the submitted judge wrongly applied a long-only rule to bearish posts, and the authors decompose ranking into text, volatility and horizon, and pairwise margin components.
In briefLatentQuant: Keeping Policy-Facing Latents Stable Under NVFP4 VAE Quantization
SummaryThis paper studies how NVFP4 quantization of pretrained video VAEs, reused in world action models, affects the encoder latents that downstream action policies rely on. Joint quantization-aware training restores reconstruction quality but sharply lowers LIBERO success from 95.5% to 10.5%. The proposed LatentQuant is a two-stage method that first aligns the quantized encoder with its full-precision counterpart, then freezes the encoder and adapts only the decoder. It reports near-baseline control on Wan2.1 and Wan2.2, with 95.75% success on LIBERO and 68.8% on RoboTwin.
In briefLatent-Space Score-Based Estimation of the Bayesian Cramér-Rao Bound for High-Dimensional Imaging
SummaryThis paper presents a data-driven method for estimating the Bayesian Cramér-Rao bound in high-dimensional imaging problems where the prior is too complex for analytical treatment. The authors work in the latent space of a pre-trained variational autoencoder to reduce dimensionality and then map the bound back to the original parameter space. They also introduce a Bochner-space sliced score matching objective with an H¹ Sobolev norm that suppresses high-frequency artifacts. Tested on a quantitative photoacoustic tomography breast imaging problem with over one million parameters, the method gives stable, artifact-free bound estimates.
Evan Scope Crafts, Thomas Wynn, Seonyeong Park et al.
In briefOne-step curvature probes miss forgetting: retained capacity and terminal null-space correction in continual learning
SummaryThe paper argues that a one-step curvature probe of an initial update direction does not capture old-task forgetting after a continual learner reaches new-task fit. Under an overparameterized linearization, it analyzes projected gradient descent, factorizes the terminal old-task quadratic ratio, and proposes terminal null-space correction that preserves linearized new-task outputs. On Permuted MNIST and Split CIFAR-100, correction reduced signed old-task loss in 166 of 180 pairs, but since acceptance and reporting use the same held-out split, the result is descriptive.
Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Ibne Farabi Shihab, and Md Najmus Swaqeeb
In briefEvolving LLM-generated features for auditable, interpretable classification
SummaryThis paper proposes an evolutionary framework in which a large language model iteratively generates natural language binary features (rubrics) for classification. A transparent classifier such as logistic regression is trained on these features, making the decision logic auditable. The feature set evolves across iterations guided by classification errors, per-class activation rates, and ablation scores. Evaluated on three benchmarks including a credit risk dataset, evolved features outperform single-shot rubrics on average and beat zero-shot LLM classification on two of three tasks.
In briefAutonomous driving decisions guided by Chinese philosophy and retrieval-augmented language models
SummaryThis paper proposes CPW-Drive, a retrieval-augmented generation framework that brings value principles drawn from Confucian thought into autonomous driving decision-making. It also introduces PSSR, which retrieves historical cases by comparing vehicle layouts and velocity-extrapolated states. In Highway-env multilane driving tests, success rates reached 93.0%, 86.0%, and 72.0% across three traffic configurations, outperforming the strongest baseline. The authors note that the approach introduces efficiency and latency trade-offs.
In briefTransformers that write physics formulas and extrapolate better
SummaryThis paper studies symbolic regression, which searches for a compact formula that fits input-output data and predicts unseen inputs. The authors shape the set of formulas used to train a transformer, which improves extrapolation beyond the training input ranges. Fine-tuning on noise-corrupted targets further makes the formulas robust to observation noise. On SRBench and LLM-SRBench, the model synthesizes a formula in about ten seconds and extrapolates better than all evaluated methods at a comparable budget.
Shuwei Wang, Vadim Bulitko, Michael Youngblood et al.
In briefDePICT: A Decision-Preserving Interface for Constrained Downstream Tasks
SummaryThis paper introduces DePICT, a method for building interfaces that preserve decisions in constrained optimization. It ranks context directions by how sensitive the optimizer's solution is to them and aggregates these rankings across an operating regime. The authors derive a KKT-based condition for when a context direction is relevant to the optimizer, showing that appearing in the problem does not guarantee influence on the decision. In a controlled diagnosis, DePICT recovers the decision-relevant interface exactly and reduces linear predictor regret to 0.009, versus 0.475 for the strongest baseline.
Utkarsh Grover, Ravi Ranjan, Agoritsa Polyzou et al.
In briefDynamic Time Step Prediction for Blur-Like Image Restoration Using Inverse Heat Dissipation
SummaryThis paper studies diffusion-based image restoration, where inversion usually starts from a fixed intermediate time step that suits different degradation levels poorly. The authors propose an Inverse Heat Dissipation Model that diffuses images by gradually blurring them, matching blur-like degradations such as blur, haze, and low light better than Gaussian noise. They also add a time predictor that estimates the inversion starting step based on degradation severity. Experiments on standard benchmarks report state-of-the-art quantitative and qualitative results and good generalization across restoration tasks.
In briefToward Forgetting: Optimiser History and Loss of Answer Probability Mass
SummaryThis paper studies catastrophic forgetting in language model fine-tuning. It splits old-task loss into confusion among old answers and probability leakage outside the answer set. Across three model families, answer mass consistently declines while discrimination among old answers usually improves. Decomposing Adam updates shows that accumulated history favours leakage while the current gradient opposes it, with older gradients from the new task being the main harmful source.
In briefTreeWalker: Faster Batch Scoring for Tree Ensembles on Grouped Rows
SummaryThe paper proposes TreeWalker, a partial-evaluation method for scoring tree ensembles on groups of rows that share some feature values, such as survival models, click-through-rate ranking, and scenario analysis. It treats constant features as static and varying ones as dynamic, walking each tree once per group and skipping empty subtrees, and it reads standard LightGBM and XGBoost models without retraining. The authors report speedups over row-independent traversal on Intel and Arm machines and state that outputs match treelite GTIL up to tree-ordering roundoff.
In briefFluidPD: In-place elastic scheduling for SLO-aware prefill-decode disaggregated LLM serving
SummaryThis paper presents FluidPD, a serving system for large language models that separates the prefill and decode phases. It uses FluidToken to offload part of prefill computation to decode workers during short bursts when decode capacity is free, and FluidRole to reassign worker roles in place during sustained imbalance without reloading the model. On production Azure trace workloads, FluidPD improves overall SLO attainment over static SGLang by up to 94.6 percentage points without adding workers.
In briefAgentic multimodal models become less likely to refuse harmful requests when using tools
SummaryThis paper studies the safety of multimodal large language models that call tools such as zooming and tagging to reason about images. Across three popular safety benchmarks, the top open- and closed-weight models tested showed significantly lower safety in tool-using settings than without tools, with relative refusal failure rates rising by up to 68.7%. Based on analysis of over 100,000 responses, the authors propose two possible reasons for this safety degradation.
WhyReveals tool use degrades refusal in multimodal LLMs, validated across three safety benchmarks and over 100,000 responses, with strong safety implications.
In briefGeneral Decision Models: A Benchmark Beyond Jev for Structured Judgment
SummaryThis paper introduces JEVal, a bilingual benchmark of 11,257 instances from 36 datasets across 10 domains, used to evaluate 25 general decision and generative LLM configurations. Decision models perform best when decisions can be resolved from available evidence, but weaken on specialist knowledge and calibrated uncertainty, often overstating probabilities. In long multi-step systems, faster local decisions lower task success as errors accumulate. The authors also propose SII-Decision models, distilled from an LLM's reasoning into single-pass decisions.
In briefSGAnalog: An end-to-end analog circuit benchmark built from open-source chip tapeouts
SummaryThis paper introduces SGAnalog, a benchmark of 273 topologically distinct analog circuits drawn from human-designed open-source Tiny Tapeout chip submissions. It evaluates models on schematic-to-netlist transcription and device sizing. On 66 transcription tasks, the strongest model reaches 56.1% exact graph isomorphism. On 17 sizing tasks, the leading model scores 91.2 out of 100 against a human reference, while some newer models refuse prompts they otherwise transcribe.
In briefAttack-resilient multi-agent target localization and tracking using Bayesian attack graphs and mutual information.
SummaryThis paper proposes a framework for multi-agent networks that localize and track a target while under false-data injection (FDI) attacks. It combines information-based navigation, Bayesian attack graph detection of compromised agents, and reachable-set-guided recovery that uses LQR to relocate the target when the localizing agent is attacked. Simulations show the framework maintains robust tracking performance under FDI attacks.
In briefDABACO: A multi-camera dataset for locating screens and estimating pointing.
SummaryThis paper introduces the DABACO dataset, developed for a low-cost pointing device project, containing video captured by several low-cost cameras on embedded platforms such as Raspberry Pi 4B and ESP32-S3. It describes an annotation pipeline and releases marked and marker-free images, corner annotations, and modification masks. An open-source evaluation toolkit is provided with two baselines, a classical edge-and-contour method and an off-the-shelf YOLOv8 segmentation model, evaluated using IoU, corner error, and pointing error.
\'Oscar G\'omez-C\'ardenes, Jos\'e Gil Marichal-Hern\'andez, Juan Manuel Mart\'in-Do\~nas
In briefA governed agent system that builds dashboards from plain-language requests on enterprise data
SummaryThe paper presents a prototype for DataBrain that converts natural-language analysis requests into inspectable dashboards. A Codex plugin and standalone runtime let the model propose actions while deterministic Hooks handle approval, audit, checkpointing, and recovery. On frozen real-data tasks, strict success was 6/8 for metadata and SQL tasks, and all ten fault scenarios met their specified outcomes. The authors state that these small, DataBrain-specific results do not establish production readiness, general text-to-SQL accuracy, or efficiency over manual construction.
In briefCOVER: Learning to Accept More Reliable Predictions in Selective Sleep Staging
SummaryStandard sleep-staging methods apply the same model to every EEG epoch, spending computation even on epochs a smaller model could handle. This paper studies the first stage of a cascade, maximizing how many predictions a primary classifier accepts under a fixed accepted-risk target. The authors propose COVER, which replaces maximum softmax probability with a learned error score and directly optimizes coverage under an empirical risk constraint. On Sleep-EDF-20, COVER reaches 50.4% mean coverage at 4.5% subject-equal risk, the highest among the eight scorers evaluated.
Yukai Song (1), Yangfan Deng (2), Jijun Yin (1) et al.
In briefA geometric view of reward updates in inverse reinforcement learning
SummaryThis paper revisits reward updates in inverse reinforcement learning by examining the geometry of occupancy space rather than relying only on gradient descent. It shows that the geometry along the straight path toward the optimal occupancy affects the gradient of the IRL dual problem. The authors propose preconditioning the gradient with the average Fisher information along this path, analyse the updates from a Mirror Descent perspective, and present convergence results for these geometry-aware algorithms.
