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 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 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 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 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 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 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 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.
In briefEvoWorld: Out-of-Sight World Evolution in Generative Video Models
SummaryThis paper argues that generative video models render plausible visuals but lack the reasoning needed to determine how events change a world. It defines the Out-of-Sight Causal Evolution problem and proposes EvoWorld, where a vision-language model tracks world state and reasons about consequences while a segment-autoregressive renderer produces the video. The authors also release EvoWorld-Data and EvoWorld-Bench for training and evaluation, and report that EvoWorld improves event-conditioned causal evolution over recent models while preserving visual quality and consistency.
In briefTailSteer: A Method to Curb Visual Drift in Multi-turn Image Editing
SummaryThe paper observes that strong single-turn image editing does not ensure reliable multi-turn editing, since small unintended changes accumulate over repeated edits. It proposes TailSteer, a final-stage training framework that aims to reduce this drift while keeping the capabilities learned earlier by the model. Experiments on the FLUX, SenseNova, and Qwen model families report substantially less cumulative drift over 15 to 20 editing turns, with base-model performance largely preserved on benchmarks.
In briefCoupled Scene Graph Anticipation: Constrained Recall Can Mask Structural Inconsistency
SummaryThe paper proposes CoupledSGA, a model that predicts object relationships in future video frames by letting relationships sharing an entity influence each other. It adds a Structural Consistency Loss based on mutually exclusive predicates. The authors show that With-Constraint evaluation cannot capture within-group violations, and on Action Genome the loss lowers the No-Constraint contradiction rate from 45.92% to 4.07% while reducing R@20 from 42.21% to 35.18%.
Zeno Nongmaithem, Anushttha Srivastava, Anshit Sinha et al.
In briefPhone-Based VR Exergames That Turn HIIT Workouts into Games with Real-Time Pose Tracking
SummaryThis paper presents a VR exergame system that gamifies High-Intensity Interval Training with 11 predefined combo moves. A standard smartphone on a tripod captures full-body movement, which is processed in real time by a pose estimation algorithm and assessed by a backend using MCP and GPT-4o. In a within-subject study of 20 participants aged 18 to 25, VR sessions produced higher enjoyment, immersion, and likelihood to recommend than non-VR sessions. The system's evaluations agreed with a certified trainer's assessments in 87.5% of 10 sessions.
Abhiram Kothagundu, Pheelip Pritipal Raipure, Abhijit Chandrakant Patil et al.
In briefConceptARC at Scale: Testing Concept Understanding in Reasoning Models
SummaryThis paper introduces ConceptARC at Scale, a generator that creates thousands of ARC-style tasks targeting specific concepts. Using it, the authors build ConceptARC-100, a dataset of 1,600 manually reviewed tasks covering 16 concepts, ten times the size of the original benchmark. They evaluate recent reasoning models and models trained from scratch, and analyze their internal representations. The representations cluster by operations applied to objects, such as recoloring, removal, or translation, suggesting models form stable abstractions of transformation operations.
Philipp Mondorf, Adrien Deli\`ege, Solim LeGris et al.
In briefReusing output templates to simulate shell command results with small language models
SummaryThe paper introduces template gap reconstruction (TGR), a method for simulating shell command output with small language models without executing the commands. A template selector picks a reusable output template, and autoregressive models generate the varying gaps between fixed literals. Evaluated on a target environment excluded from the template bank, TGR improves output edit similarity when template coverage is high and reduces generated tokens by 26.3% to 38.3%.
In brief"Turns Matter": Improving Streaming Video Reinforcement Learning with Turn-Level Credit Assignment
SummaryThe paper addresses streaming video understanding, where models must repeatedly decide whether to respond or keep observing. It proposes TurnWise, a turn-level GRPO framework that computes group-relative advantages for each streaming decision from its remaining return. Token losses are averaged within each turn and then across turns, so response and waiting decisions receive equal weight. Experiments show the method outperforms prior RL approaches for streaming video understanding.
In briefAirlock: Limiting Fault Spread in Multi-Agent Systems via Selective Online Containment
SummaryThe paper presents Airlock, a runtime and guardrail for multi-agent systems that performs selective fault containment after an alert. Its tree-structured runtime, EdgeCitadel, narrows permissions from parent to child agents, while an append-only provenance graph records cross-agent data consumption. After an alert, Airlock traces forward to isolate tainted executions and pending actions, and backward to identify root causes. An optional optimized cut minimizes disruption to healthy work under a deadline.
In briefSurvey of dexterous robotic hands in the foundation-model era
SummaryThis survey reviews research on dexterous robotic hands from January 2024 to early November 2026, including earlier studies to show technical lineages. It organizes the literature into five connected layers: hardware, data collection and curation, cross-embodiment transfer, learning and foundation models, and evaluation and applications. The authors conclude that the value of human data depends on kinematic, visual and contact alignment rather than recording volume alone. They also argue that completing a task is not sufficient to establish physical correctness or deployment readiness.
In briefStudy of symmetry breaking at the edge of stability using central flow analysis
SummaryNeural networks have continuous symmetries that yield quantities conserved under gradient flow, such as layer balancedness in linear and ReLU networks. At the edge of stability, where iterates oscillate around a smoother trajectory described by central flow dynamics, finite-step gradient descent may not preserve these conservation laws. The paper studies how gradient flow conservation laws evolve under this central flow description, analysing a scalar two-layer model, two-layer matrix factorisation, and a two-layer ReLU network to determine which laws remain and which are broken.
In briefReasoning Over Supplied Facts in Multilingual LLMs
SummaryThis paper introduces Rel-Bloom, a multilingual diagnostic benchmark of 450 multiple-choice questions built from 30 Wikidata relations. Questions are arranged across five Bloom-inspired levels and three veracity conditions, and 43 models were tested in six languages. Accuracy drops sharply from the Apply to Analyze level in open-weight models, and Evaluate items remain difficult even for strong models. The authors conclude that greater model capability does not by itself guarantee reliable reasoning over supplied relational evidence.
Zohreh Abbasi Liasi, Batoul Hammoud, Anas Madkoor et al.
In briefEnergy-Guided Transport: A Projection-Free Method for Physics-Constrained Flow Matching
SummaryThe paper proposes Energy-Guided Transport, a training-free framework that expresses nonlinear physical constraints as a smooth potential energy landscape. Instead of interrupting ODE integration with algebraic projections, it applies a continuous dissipative force to intermediate generative states to steer them toward the physical manifold. The method is evaluated on various 1D and 2D physical systems, including the chaotic Navier-Stokes equations.
In briefBeyond the Principle-Practice Gap: Responsible AI as a Situated Ethical and Epistemic Practice
SummaryResponsible AI research has found that ethical principles are often not reflected in AI design and development, a gap usually treated as a problem of implementation, awareness, or compliance. Based on fourteen interviews with people designing and governing AI-enabled anti-trafficking systems, this paper argues that the gap arises from institutional and professional conditions that shape which issues actors recognise as ethical problems requiring action. It describes four linked moments: perception, judgement, situated action, and socio-technical materialisation. The authors conclude that responsible AI governance needs epistemic infrastructures that make the knowledge of affected communities receivable and actionable in design.
In briefALTO: A Corpus of Tech-Facilitated Gender-Based Violence in African and Levantine Languages
SummaryThis paper presents ALTO, a dataset of 3,022 examples of tech-facilitated gender-based violence in Levantine Arabic, African French, and Swahili. The examples were crowdsourced with civil society organizations in Jordan, Francophone West Africa, and Kenya, and annotated using active learning. Existing harmful content detection models show poor recall on ALTO, which improves with fine-tuning, and models often miss covert harms, with markedly higher error rates in Swahili. The annotation taxonomy, dataset, and trained models are released.
Tianwei Zhang, Ivy Njoki Nyagwencha, Zahraa Dirani et al.
In briefWhere It Breaks: Diagnosing Failures of a Vision-Language Pac-Man Agent
SummaryThis paper studies a Qwen3.5-9B-based vision-language agent that plays Pac-Man using only four movement actions, asking whether poor performance comes from reading the screen or from choosing actions over long sequences. Using separate perception checks and primitive-policy reinforcement learning on textual boards, completion on eight held-out mazes rose from 0% to 80.2% at a selected checkpoint. The authors find that accurate static screenshot reading can coexist with errors during interaction, and that further optimization can come with worse action choices.
In briefAgent-driven generative priors for synthesizing and segmenting pathology images
SummaryThis paper proposes a three-stage framework to address the scarcity of pixel-level annotations in pathology images. A pathology agent system screens image quality to build TCGA-200K, a dataset of 217,030 patches from TCGA, and a generative model trained on it is adapted to semantic segmentation with a lightweight head. The authors also introduce PathSeg-380K, with expert annotations for about 380,000 cells across 12 cell types. Experiments on multiple pathology datasets report state-of-the-art segmentation performance.
In briefPINR-ESI: A Physics-Informed Self-Supervised Neural Method for Mapping Brain Activity from EEG
SummaryThis paper proposes a method for extended-source EEG imaging that requires no paired source labels or simulation-based supervised pretraining. It uses a lead field matrix to constrain an implicit neural representation so that predicted and measured EEG agree, modeling cortical current density as a continuous field. Simulations showed the clearest advantage in identifying multiple and larger active cortical regions. Evaluations on real EEG from epilepsy patients and intracranial stimulation experiments supported the framework's feasibility for source imaging.
In briefAdds a Detector-Agnostic Subspace Signal to Improve Near-OOD Detection
SummaryThis paper proposes PLUS, a method for near-out-of-distribution (near-OOD) detection that builds a low-dimensional per-class subspace from a sparse autoencoder's most class-active neurons. Inputs are scored by their whitened squared Mahalanobis distance within the subspace of their predicted class. This signal is fused additively with existing detection scores at a single fixed global weight. Across 18 detectors and 126 detector–backbone combinations, it consistently improves near-OOD AUROC while far-OOD performance is broadly maintained or improved.
In briefLifeVal: A Benchmark for Testing AI on Personalized Consumer Tasks
SummaryThis paper introduces LifeVal, a benchmark for evaluating how well AI models handle personalized agentic consumer tasks across 148 categories, such as managing finances, moving apartments, searching for a job, and finding childcare. The authors estimate the tasks' value through a survey of over 3,800 consumers, cross-referenced with the American Time Use Survey and the Multinational Time Use Study. Task creators who met an expertise threshold built the tasks from real ChatGPT conversations, memories, reference files, and personal context, and each task passed at least three rounds of human quality control.
Jenny Yang, Elizabeth Proehl, Anjan Bharadwaj et al.
In briefSekai: Incremental Event Memory for Reasoning Video Segmentation That Tracks Changing Targets
SummaryThis paper introduces Generalized Reasoning Video Segmentation (GRVS), a task where the objects matching a query can change as new events occur in a video. The authors build Isekai, a benchmark of 60 videos and 197 queries, and propose Sekai, an online framework with an event memory reasoner and a memory-grounded decoder. Sekai reaches 50.33 J&F, improving on the strongest baseline by 3.2 points.
In briefSecureGARD: Ontology-Guided Security Framework for Tool-Using AI Agents
SummaryThis paper presents SecureGARD, a framework that governs large language model agents which call external tools, aiming to block indirect prompt-injection attacks. It maps tool calls to a shared semantic representation and uses an executable ontology to decide whether each action is ready, waiting, or forbidden, with waiting states enabling bounded replanning. The ontology is refined offline from verified training runs and frozen after held-out validation. The authors report a 0% attack success rate on AgentDojo, ASB, and AgentDyn, along with improved task utility under attack.
In briefDo unlearning algorithms learn the right features?
SummaryThis paper examines how machine unlearning, which removes the influence of deleted training data from a model, is evaluated. The authors argue that comparing only post-deletion outputs can hide differences in the features encoded by an unlearned model versus one retrained on the remaining data. They prove that models agreeing in output can be broken apart by a single gradient step on their prediction heads. They also introduce a feature recovery measure that separates missing from extraneous features, finding that many methods performing well on standard metrics fail to recover the retrained features.
Polina Dolgova, Amartya Sanyal, Sebastian U. Stich
In briefSelf-Directed Apple Tasting: Online Learning with Partial Feedback
SummaryThis paper studies the classical apple-tasting setting, a partial-feedback model where the learner sees the true label only when it predicts 1. It considers a self-directed variant in which the learner adaptively chooses the order of points to predict on. In the realizable case, the authors establish a tetrachotomy for the expected minimax number of mistakes, with rates depending on the VC dimension, the ∅-centered star number, and the self-directed dimension. In the agnostic case, the minimax regret is either $\tilde \Theta(\sqrt{n})$ or $\Theta(n)$, depending on the VC dimension.
In briefAgentTrustCI: Build-Time Tests for Java LLM Agent Policy, Grounding, and Trajectories
SummaryThis paper presents AgentTrustCI, a Java and JUnit 5 framework that checks LLM agents at build time for policy compliance, grounding of answers in tool output, and adherence to required workflows. It applies these checks to recorded traces and was evaluated on 55 synthetic cases, where the full pipeline correctly classified all injected faults and passing controls and missed no critical cases. The authors note that these results reflect coverage of the authored regression suite rather than production accuracy.
In briefCardioDrift: fully automatic heart motion tracking on a 3D atlas mesh with slice-shift correction
SummaryThis paper presents CardioDrift, a fully automatic, atlas-based probabilistic method for tracking subject-specific cardiac motion from cine cardiac MRI. It generalises Coherent Point Drift to jointly register atlas surfaces of several cardiac regions to segmentation contours across the cardiac cycle, with slice-shift correction built into the same model. Evaluated on 1301 UK Biobank subjects and the M&M dataset, it outperforms the strongest baseline on all distance and overlap metrics on both datasets.
Reza Akbari Movahed, Abuzar Rezaee, Shatha Alamri et al.
In briefFinding Where Search Agents Make Mistakes: Graph-Based Step-Level Checking
SummaryThis paper studies search-augmented reasoning systems that can reach correct answers despite errors in intermediate steps. The authors introduce GraphRail, which represents requirements, evidence, and reasoning as structured claims and compares them without a generative judge or gold answer. Across three multi-hop QA benchmarks and three reasoning frameworks, GraphRail achieves higher error recall than generative verifiers and remains stable under stochastic extraction.
In briefTiming analysis to identify dispatchers in distributed model extraction
SummaryThe paper examines how API distillation campaigns spread requests across many load-balanced accounts to evade content classifiers. It shows that request arrival times and stream completion times reveal the underlying dispatching strategy, such as round-robin counters or key-shuffling pools. Using a 20 percent seed set, the timing analysis ranks the remaining accounts of a campaign within 20 minutes with an AUC of 0.92 to 1.00.
In briefBeyond Accuracy: What AI Assurance Must Establish in Public Services
SummaryThis paper examines how public bodies can decide whether AI-generated evidence, such as classifications and summaries, is reliable enough to use. The authors tested eight GPT-4o configurations classifying 1,200 UK Prime Minister's Questions into 21 policy topics against human labels and repeated runs. Accuracy was similar across settings, but 24.3% of labels changed between runs and agreement between configurations did not ensure correctness. The authors also analysed 37 UK and EU policy documents on testing and oversight requirements.
In briefDetecting hidden robot action errors using calibrated relational world models
SummaryRobot actions can meet shape and range checks yet still have wrong meaning, such as swapped axes, inverted signs, or reversed gripper polarity. The paper uses learned state-action-effect relations as offline contracts, calibrating alarms on clean held-out transitions and ranking likely faulty fields with a fixed inverse vocabulary. On ten LIBERO Spatial tasks with 432 source episodes, at a 2.5% clean false-positive rate, the exact conditional of a four-component joint Gaussian mixture reached 65.7% recall, versus 46.1% for ridge residuals. These are retrospective diagnostics, not deployment calibration guarantees.
In briefDynamical State-Space Diffusion for Spatiotemporal Prediction
SummaryThis paper proposes Dynamical State-Space Diffusion (DSSD) for spatiotemporal prediction. Unlike standard diffusion models that add independent noise to each frame, DSSD learns the temporal structure of the noise process itself, building a temporal kernel from damped oscillatory modes of a stable state-space model. The denoiser architecture stays unchanged. On scientific-field and robot-video datasets, DSSD outperforms diffusion baselines that share its denoiser.
In briefReproducibility and robustness of large language models in political and social science research
SummaryThis chapter examines the reproducibility problem in research using large language models, whose probabilistic outputs vary across models and versions. It proposes a distribution-based standard for replicability, reproducibility, and robustness, along with a standardized prompt script and error-structure analysis. Using three datasets, the authors show that the choice of measurement can reverse the sign of downstream regression coefficients. They also present a stacked ensemble of open-weight models that improves precision in the correction step while retaining estimate validity.
In briefPredicting human motion by progressively compensating for lost prior knowledge
SummaryThis paper addresses human motion prediction, which forecasts future human poses from observed past motion sequences. The authors propose Continual Prior Compensation (CPC), a framework that splits predictions for different time horizons into subtasks trained in multiple stages, with a learnable Prior Compensation Factor measuring prior knowledge loss. An extended version, CPC++, adds a Fine-Grained Prior Compensation Factor for each subtask. Experiments on three benchmark datasets with several state-of-the-art backbone models show the frameworks are effective and flexible.
In briefUncertainty-guided efficient communication for drones sharing sensor data
SummaryThis paper proposes UCP, a framework that makes communication efficient for UAV teams sharing perception data. It uses observational uncertainty to find sensor-limited regions and discriminative uncertainty to find ambiguous regions near the decision boundary, generating request and supply maps. Received messages are fused by U-Fusion, an uncertainty-guided cross-attention module. Experiments on three benchmarks cover 3D object detection and trajectory prediction.
In briefTestBED: An Open-Source Toolkit for Bayesian Experimental Design
SummaryThis paper presents TestBED, a general-purpose open-source toolkit for Bayesian experimental design (BED). It breaks the BED process into modular, reusable components covering tasks, objectives, training, policy architectures, and evaluation, along with scalable implementations of core primitives such as expected information gain estimators. The authors use TestBED to run a large-scale benchmark of contemporary BED methods, including controlled studies of design policy choices such as time embeddings.
In briefTrick-or-Treat: Adversarial Testing of Medical AI Using Synthetic Images
SummaryThe paper introduces Trick-or-Treat, a benchmark that pairs real chest radiographs with synthetic clinical photographs to test whether models reason correctly across image types. In mismatched pairs, three frontier models hallucinated clinically significant photograph findings 45-57.5% of the time, compared with 25-37.5% when the photographs were shown alone. The models tended to ignore the mismatch and invent findings that fit the radiograph instead.
In briefFrom Certified Pairs to a Certified Hub: Checking Information Exchange Standard Crosswalks with a Reasoner
SummaryThis paper addresses how digital twins in a national ecosystem are linked to a common foundation through mappings, and defines valid mappings as coherent and conservative when merged into one ontology. The authors test crosswalks from the Information Exchange Standard (IES) to thirteen ontologies, first pair by pair and then across whole hubs, using matching, human review, SKOS certification and the HermiT reasoner. Pairwise certification and review catch different errors, and together they make eleven of twelve overlapping pairs coherent and conservative. However, merging three building crosswalks made five Brick classes unsatisfiable, showing that pairwise certification does not compose and that the hub is the unit to certify.
In briefReal-time interactive world simulation with Cosmos for training physical AI
SummaryThis paper presents a shared framework that adapts Cosmos world foundation models for real-time interactive simulation. It streams visual observations in response to actions and scene controls, letting users and embodied agents interact with evolving virtual environments. The approach is demonstrated on camera navigation, robot manipulation, controllable video transfer, and multi-view autonomous vehicle simulation, with code and model checkpoints released.
NVIDIA: Anis Ahmad, Jon Allen, Malik Aqeel Anwar et al.
In briefLearning to track cells in 3D microscopy videos using sparse labels
SummaryThis paper presents a pipeline for tracking cells in 3D time-lapse microscopy, where only some cells are annotated, to reconstruct cell lineages. It detects cells from sparse annotations, calibrates candidate links using image features, and jointly selects lineage graph events. On 199 videos from 2 zebrafish embryos, it achieves an overall score of 0.9762 in 5-fold cross-validation.
In briefPrivate mean estimation without known covariance cannot avoid dimension dependence
SummaryThis paper studies the limits of differentially private Gaussian mean estimation when the covariance matrix is unknown. When the covariance is known, optimal sample complexity depends only on its spectrum rather than the ambient dimension. The authors prove that private estimators cannot achieve similar guarantees when the covariance is unknown. They give lower bounds showing error gaps of order √(d/n) when the sample size is below the effective dimension, and of order √d/n^{3/2} above it, even for diagonal covariances.
WhyThe authors prove dimension-dependent lower bounds for private Gaussian mean estimation with unknown covariance, contrasting the known-covariance case.
