In briefWhich depth estimator suits the Jetson AGX Orin? A 2026 benchmark of monocular and stereo methods
SummaryThis paper benchmarks nine depth estimation model families on the NVIDIA Jetson AGX Orin, comparing accuracy, speed and energy. Stereo methods are far more accurate, with the best reaching about 0.33% to 0.70% AbsRel, while the best monocular models reach about 12% with the true focal length. The authors give recommendations per frame-rate budget and release their code and per-scene results.
WhySystematically benchmarks nine monocular and stereo depth methods on Orin for accuracy, speed and energy, giving practical value though limited methodological novelty.
In briefSmall Residuals, Wrong Answer: Learning Viscosity Solutions of Hamilton-Jacobi-Isaacs Equations with PINNs
SummaryThis paper examines whether physics-informed neural networks, which minimize sampled PDE residuals, can select the correct viscosity solution of a first-order Hamilton-Jacobi equation. In 120 training runs on a pursuit-evasion game over a circle and flat tori, a standard network converged to folded solutions that have small residuals but are wrong over most of the domain. Feeding geodesic distance to the network, or using structural diagnostics, recovered the viscosity solution more reliably.
Raul Jimenez, Pablo Tejerina-P\'erez, Leonid Sarieddine et al.
In briefCARLAR Influence: Hallucination Risk in Cheque Extraction and a Compressed-Image Countermeasure
SummaryThis paper studies how multimodal language models read cheques when the handwritten-style amount and the numeric amount differ, finding that models often return the same value for both fields. Across ten visual prompting conditions, GPT-4.1 and Claude 4.5 Sonnet were tested on 243 primary cheques and 598 strict mismatch cases from two other datasets. Under courtesy-amount redaction, GPT-4.1 reached 73.7 percent CARLAR, and its confidence stayed high even when predictions agreed. The authors also test a compressed-image few-shot method that shows examples as one visual table without fine-tuning, with effects that differ by model.
In briefPaper overview: task fragility under composition and implicit thinking budgets in language models
SummaryThis paper examines why language models make more errors when handling several tasks in one prompt, even when they solve each task reliably alone. The authors find that under composition, models use fewer tokens per task, suggesting an implicit thinking budget, and tasks needing longer solution traces become more error-prone. Grouping tasks by their fragility can adjust this budgeting, reportedly reducing token usage while preserving accuracy.
Chenpan Li, Thies Brockmoeller, Stefanie Jegelka et al.
In briefMeta-iLaD: Identifiable latent dynamics learned via meta-learning across environments
SummaryThis paper proposes Meta-iLaD, a framework for learning latent dynamics of high-dimensional time series. It uses meta-learning across dynamics environments, with a feedforward meta-learner that extracts the environment variable from few-shot examples, enabling generalization to unseen environments. The authors establish identifiability for the latent state, environment variable, and dynamics function jointly, and report experiments on synthetic and real data showing better generalization and forecasting accuracy.
In briefSITO: Testing a Commercial Text-Safety API for Missed Harms and False Flags in English, Ukrainian and Mixed Language
SummaryThis paper evaluates a commercial text-safety API (tuteliq-detect-v5 release) on six harm categories, including authorised push-payment fraud, romance scams and bullying. The authors built SITO, a benchmark of 2,332 synthetic cases in English, Ukrainian and mixed Ukrainian, pairing each harmful case with hard harmless counterparts. The system missed third-party reports of ongoing harm more often than perpetrators' messages, and recall was lower on text from a held-out generator family. The authors recommend that platforms measure false positives on their own traffic and check how bystander reports are handled.
In briefVision-Based Task Modeling and Automated Planning for Humanoid Robots Supporting Autism Education
SummaryThis paper presents a method that lets a semi-autonomous, camera-equipped robot recognise the goal of Shoebox Tasks, a activity used in education for individuals with Autism Spectrum Disorder (ASD), and plan actions to complete it. A vision-based Large Language Model converts photos of the task into a PDDL representation and generates action plans. A Nao robot executes the tasks and assists students by suggesting next steps or signalling failures through demonstrations, voice, or gestures, without requiring teachers to have programming expertise.
In briefLearned control variates for estimating rare grid outages at EV charging stations
SummaryThe paper addresses the risk that synchronized EV charging overloads distribution feeders and can trigger cascading city-wide outages. It proposes a control variate built from a classifier trained on cheap observable features such as temperature and load forecasts, to reduce the number of runs needed from an expensive grid simulator. Applied to the feeder network of Surabaya, Indonesia, the learned control variate reduces city-wide blackout days by about a third compared with plain Monte Carlo at the same simulation budget.
In briefUnderstanding the Benefits of Scaffolds Through Computational Separations
SummaryThis paper studies inference-time scaffolds, external procedures that call a model many times and keep state across calls, and whether their benefits can be transferred into model weights through training. The authors find that even simple scaffolds with small models significantly outperform a student model distilled from the scaffold's traces. They prove this gap stems from a representational limitation that no fixed-depth forward pass can overcome under a widely believed complexity conjecture. Experiments show that enlarging the student, simplifying the task, or lengthening chain-of-thought can close the gap.
WhyProves harnesses compute NL-complete closures that fixed-depth forward passes cannot represent, with experiments confirming the distillation gap.
In briefMoSAIK: AI-based motion search in k-space for retrospectively correcting motion in brain MRI
SummaryThis paper presents MoSAIK, a method that retrospectively corrects motion artefacts in routine two-dimensional T2-weighted turbo spin-echo brain MRI. It estimates in-plane rigid-body motion directly from standard multi-coil k-space data using a convolutional embedding and Transformer encoder, then feeds these estimates into a motion-aware reconstruction. In evaluations on four held-out subjects, motion estimation errors were low, and more than 75% of slices in every in-vivo acquisition showed information gain over uncorrected reconstructions.
Oscar Dabrowski, Abijith Jagannath, Fr\'ed\'eric Grouiller et al.
In briefA Hybrid TFT-NBEATS Model for Multi-Horizon Solar Power Forecasting (a deep learning model combining two forecasting networks to predict PV output at several future time spans, with interpretability)
SummaryThe paper proposes a hybrid deep learning framework that fuses a Temporal Fusion Transformer and N-BEATS through an attention-based mechanism to forecast photovoltaic power output. It is evaluated on multi-year real-world data from the Desert Knowledge Australia Solar Centre and produces forecasts for 30-minute to 24-hour horizons in one pipeline. The authors report lower RMSE than baseline models and MAPE below 5% across all datasets. The model also offers interpretability through attention visualization and trend-seasonality decomposition.
Shafi Bin Sultan, Sabik Bin Sultan, Mohammad Tajrian Arafat Alvi et al.
In briefShared brain shape markers across Alzheimer's and Parkinson's disease via transfer learning
SummaryThis study proposes an interpretable, spatially adaptive transfer learning framework to identify brain morphometric biomarkers linked to Alzheimer's disease (AD) and Parkinson's disease (PD). It was applied to two open neuroimaging datasets, ADNI and PPMI. The cross-disorder classifier revealed anatomical alteration patterns that align with prior imaging findings, and a spatially adaptive classifier using a shared prior showed statistically superior diagnostic performance over individual models.
In briefGRID: Assessing Health Influencer Credibility by Modeling Discourse as Argument Networks
SummaryThe paper introduces GRID, a framework that represents each influencer video as an argumentation graph linking claims to the video's main takeaway, connects related claims across videos, and grounds claims in scientific evidence. It jointly infers credibility over this network instead of judging each video in isolation. Evaluated on 569 expert-assessed health videos, the authors report that accounting for the broader discourse improves content validation. They also use the inferred credibility to compare argumentative strategies and moral framing between low- and high-credibility content.
Nishanth Sridhar Nakshatri, Eylon Caplan, Dan Goldwasser
In briefPRIMFLOW: Learning Trajectory Primitives for Few-Step Distillation
SummaryThis paper studies few-step trajectory distillation, which learns global updates from local supervision along a teacher model's sampling trajectory. The authors propose PRIMFLOW, which learns a continuous primitive representing cumulative displacement and obtains global updates from endpoint differences. In large-scale text-to-image experiments at one network evaluation, it reports a T-FID of 20.92 and an HPSv2 score of 28.35, improving over a trajectory-distillation baseline. After endpoint absorption, performance stays comparable.
In briefICLHerd: herding-based data selection for LLMs using in-context learning as a proxy
SummaryThis paper addresses how to choose training data for large language models. It observes that existing max-similarity methods cause selected data to cluster tightly around development samples. The authors propose ICLHerd, which uses a von Mises-Fisher kernel to balance mode-seeking and mode-averaging, and uses in-context learning as a proxy to tune this trade-off efficiently. Results reported show ICLHerd outperforms existing LLM data selection algorithms.
In briefSensing Contact States in Gecko-Inspired Adhesives for Orbital Servicing
SummaryThis paper addresses the difficulty of establishing reliable contact with non-cooperative targets in autonomous on-orbit servicing. It presents a real-time framework that uses embedded capacitive sensing so gecko-inspired dry adhesives can estimate their contact and adhesion state across five interaction states. Among six lightweight machine learning models, Random Forest reached 95.65% offline accuracy. Combined with a rule-based state machine, online accuracy rose from 78.02% to 93.14% with a 53 ms inference cycle.
Eduart Abdullahu, Francesco Grella, Carol Martinez
In briefUsing short movement trajectories to reveal animal identity and learning progress
SummaryThis paper introduces a self-supervised representation learning framework based on the hypothesis that behavioral traits linked to identity and learning stay relatively stable over nearby time intervals. It uses temporal proximity as supervision to learn compact representations directly from movement trajectories. Using data from freely behaving rats, the authors report that trajectories as short as one second contain learning and identity signals, with egocentric pose dynamics more predictive of identity and allocentric motion more predictive of learning.
In briefRecovering an ontic state through prediction: learning a system's state from its responses
SummaryThe paper examines how a system's internal state can be inferred from its responses to queries. It defines a state as the complete response profile over a specified query universe, treating configurations that agree on all queries as the same state. The authors study the relationship between prediction and state recovery through query ambiguity, state-space geometry, and the choice of prediction targets, using synthetic systems with known response laws and neural recordings.
In briefAmmonia-hydrogen phase equilibria in planetary envelopes via machine-learning simulations
SummaryThis paper uses machine-learning interatomic potentials trained on density functional theory to simulate ammonia-hydrogen two-phase coexistence from 500 to 4000 K and 0.1 to 100 GPa. It maps the compositions of coexisting phases and the critical curve separating full miscibility from phase separation. The authors suggest ammonia may be raining out in Uranus and Neptune, while in the warm sub-Neptune TOI-270 d the two are likely miscible in all proportions.
In briefGlass-box deep reinforcement learning for controlling high-speed airflow
SummaryThis perspective paper examines how deep reinforcement learning (DRL) can be used to control high-speed compressible flows. It notes that most existing demonstrations are low-speed, idealized simulations with limited evidence of robustness, interpretability, transferability, or experimental deployment. The authors analyze constraints such as simulation cost, data efficiency, sensing in extreme environments, and reward design. They conclude that reliable technology will require validated compressible simulations, physics-aware learning, and wind-tunnel validation.
In briefReconstruction-Optimized Expert Pruning for Mixture-of-Experts Language Models
SummaryMixture-of-Experts models still need to store many expert parameters even though sparse activation reduces computation. This paper proposes OptEP, a post-training method that prunes experts using task-domain calibration data by directly optimizing differentiable expert-selection variables, then lightly updating the retained experts' weights to reconstruct the original output. Across several MoE models and math reasoning and code generation tasks, it outperforms the strongest existing expert-pruning baselines, with gains of up to 16.20% absolute accuracy depending on the model. The advantage is largest under aggressive pruning.
Jelena Markovic-Voronov, Petros Prastakos, Kayhan Behdin et al.
In briefEDMA: Entropy-Driven Multimodal Answering for Question Answering
SummaryThis paper proposes EDMA (Entropy-Driven Multimodal Answering), which uses logical entropy to frame multimodal question answering as an iterative process of entropy reduction. It quantifies how much each modality helps distinguish correct from incorrect candidate answers and selects the answering trajectory that minimises logical entropy. Evaluated on standard multimodal QA benchmarks, EDMA reportedly outperforms Planning Prompting and Chain-of-Thought baselines and is competitive with a multi-agent system baseline.
Emanuele Mezzi, Gertjan Burghouts, Fabio Massacci et al.
In briefPORA: Learning When Longer Reasoning Helps LLM Agents
SummaryThis paper proposes PORA, a method that teaches LLM agents when to use short versus long reasoning at each step under a limited total reasoning budget. During training, the agent completes tasks from the same state with both short and long reasoning and compares final quality and total cost. At inference, a controller uses this learned estimate to choose one option without running both. The authors report results across application, memory, code, retail, and mathematics tasks, including gains on BFCL memory and reduced token or request usage on code and mathematics.
In briefPROVISO: Provenance-Informed Action Selection for Secure LLM Agents
SummaryThe paper presents PROVISO, a controller that decides which action an LLM agent should execute after it proposes one and before execution. It tracks whether content comes from the user or external sources and compares the proposed action against an alternative. Evaluated on four agent security tasks, PROVISO is reported to outperform baselines on security and utility together, improving F1 by 13.84 points on SkillsBench and 7.07 points on ToolSafe.
In briefJoint Constraint Aggregation: A Method for Retrieval-Augmented QA with Qualifier Sensitivity
SummaryThe paper finds that retrieval-augmented QA can answer wrongly even when the correct answer is in the retrieved evidence, because a competing answer that meets only some question requirements but is frequently mentioned can look best supported. By increasing passages favoring that competitor, the authors show the model's choice shifts toward it. They propose Joint Constraint Aggregation (JCA), which checks every candidate against the question's requirements using the retrieved evidence. Across three evaluation settings and three instruction-tuned backbones, JCA improves overall accuracy, with larger gains when retrieval favors a competing answer.
In briefMatchCityLoc: Drone Localization in Untextured City Models Using Surface Normal Matching
SummaryThe paper presents MatchCityLoc, a framework for locating aerial drones in cities using lightweight, textureless Level-of-Detail city models instead of detailed textured meshes. It extracts surface normals from the drone's image with a zero-shot foundation model and matches them to normals rendered from the city model, creating 3D-to-2D correspondences. These correspondences feed a Perspective-n-Point pose solve that is refined iteratively through a render-and-match loop. Experiments reported strong pose recall and accurate translation and rotation among the evaluated off-the-shelf LoD-based methods, using very few renderings.
Oussema Dhaouadi, Zuria Bauer, Johannes Michael Meier et al.
In briefBeyond Two Tracks: Speaker-Anchored Full-Duplex Dialogue Recovery
SummaryThis paper addresses the scarcity of speaker-separated, persistent audio tracks needed to model full-duplex spoken dialogue, where an agent listens while it speaks. It proposes a 1.84M-parameter router that matches separated audio streams to fixed enrollment identities before a frozen diffusion generator reconstructs speech. On 25-minute remote conversations, routing reduces wrong-track time from 43.9% to 4.1% and raises fixed-track activity F1 from 0.46 to 0.84, while per-chunk reconstruction quality is unchanged. The router also transfers to unseen face-to-face and telephone recordings, where a margin-gated verification step keeps wrong-track time between 0.9% and 1.7%.
In briefCounting contributions, not views: occlusion-aware densification for 3D Gaussian Splatting
SummaryThis paper addresses a flaw in 3D Gaussian Splatting's adaptive density control, where fully occluded Gaussians are counted as observed, weakening densification in occluded regions. The authors propose an occlusion-aware filter that only uses views where a Gaussian actually contributes, and normalizes by average contribution instead of raw view count. On occlusion-heavy scenes within a StopThePop-based framework, the method improves all image-quality metrics over the baseline, with LPIPS up to 6% lower, using less than half the Gaussians of a view-count normalizer.
In briefRGH-GNN: A Risk-Gated Graph Network for Detecting Financial Fraud
SummaryThis paper proposes RGH-GNN, a heterogeneous graph neural network for credit card fraud detection. It uses transaction risk scores from a supervised LightGBM classifier as edge weights, guiding message passing through a learnable risk-gated aggregation so high-risk transactions contribute more strongly. Evaluated on the IEEE-CIS fraud detection dataset, it outperforms standard graph neural network baselines on every metric, reaching an AUC-ROC of 0.914.
Mythreyee Hari, Subin Sahayam, Vegesna S M Srinivasavarma
In briefSParTA: A modular, efficient, scalable transformer for classifying particle jets at colliders
SummaryThe paper presents SParTA, a transformer for jet tagging at high-energy colliders that processes only the hardest constituents with full pairwise attention. Softer constituents are integrated through linear-cost cross-attention to the hard core. On the quark-gluon benchmark, a 300K-parameter SParTA model matches a 2.14M-parameter Particle Transformer in AUC, and on the boosted top-tagging benchmark it outperforms ParT and ParticleNet with 511K parameters.
Rajneil Baruah, Subhadeep Mondal, Sunando Kumar Patra et al.
In briefLVMem: Image-Augmented Memory for Answering Questions Over Long Robot Observation Histories
SummaryThis paper presents LVMem, a memory framework that helps embodied agents such as robots answer questions about objects, events, and locations from long observation histories. It extends the ReMEmbR-style view-based memory by storing image paths linked to each description, uses optical-flow-based frame selection (PSFR), and stores object-centric scene descriptions. On the OC-NaVQA benchmark, LVMem outperforms ReMEmbR on binary and positional metrics and matches or exceeds the DAAAM 4D scene graph baseline on binary accuracy, without explicit graph construction.
Victoria Khoruzhaya, Maksim Kalmykov, Tuong Do Minh et al.
In briefOmni-Woodpecker: A training-free method to reduce cross-modal hallucinations in audio-visual AI models
SummaryThis paper addresses cross-modal hallucinations in Omni LLMs, which reason jointly over audio and video. The authors propose Omni-Woodpecker (OWP), a training-free framework that identifies cross-modal conflicts and uses orthogonal decomposition to separate grounding evidence from misleading priors in hidden states. Before decoding, it rebalances the fused representations to suppress wrong priors and restore ignored evidence. Using Qwen2.5-Omni-7B, OWP improves scores by 12.9 points on CMM and 8.6 on AVHBench, and it also transfers to VideoLLaMA2-AV.
WhyShifts hallucination mitigation to representation-level rebalancing with conflict detection; evaluated across multiple benchmarks and two backbones with consistent gains.
In briefA general-purpose fast decision model for clinical patient records
SummaryThis paper introduces Clinical System One (CS1), a 4-billion-parameter model that makes fast, typed judgments from clinical records and attaches a probability to each answer so uncertain cases can go to human review. It was trained on 12.4 million questions and answers from 3.6 million patients in a multi-hospital health system. On 963,408 held-out patients, it outperformed open models up to 27 billion parameters on 13 of 14 tasks, and paired with a language model it helped screen patients against clinical trial eligibility criteria.
WhyProposes a small clinical decision model with division of labour against generative models, validated on large patient cohorts and many external datasets.
Andrew Zhang, Trinh T.L. Vuong, Sophia J. Wagner et al.
In briefSciFigBench: A Benchmark for Understanding Scientific Method Figures
SummaryThis paper introduces SciFigBench, a benchmark that tests how well vision language models understand scientific method figures, covering dense content perception, visual relation comprehension, and information-flow tracing. It contains 4,352 method figures from real research papers and 40,670 question-answer pairs, with annotations generated by dedicated pipelines and manually verified. The authors also propose SciFigCue, a training-free visual prompting module that turns arrow-indicated dependencies into visual relational evidence. Experiments show that even strong models remain limited in visual relation comprehension, while SciFigCue improves performance on several tasks without parameter updates.
In briefClockMark: An adaptive, host-dependent watermarking method for large language models
SummaryThe paper proposes ClockMark, a watermarking framework for large language models that embeds information adaptively, using the state of earlier sampling decisions and the current model distribution. It spreads the full message across every watermarked token. The authors prove theoretically that the method preserves the original LLM sequence distribution, so generation quality is maintained. In experiments against 12 baselines, it shows better robustness to text editing, with bit accuracy gains of up to 16.08 percentage points over the strongest baseline.
WhyProposes adaptive, host-dependent LLM watermarking with a distribution-preservation proof and gains of up to 16.08 points over 12 baselines.
In briefLEAF: A Nanosecond-Scale, LUT-Aware Attention Model for FPGAs
SummaryThe paper introduces LEAF, described as the first nanosecond-scale LUT-aware transformer for FPGAs, aimed at real-time systems needing sub-microsecond latency. It converts the query, key, and value projections and the final feed-forward layer into FPGA-native layers with LUT-aware training, tabulating the constant matrix-vector multiplications in attention while keeping the rest arithmetic. On a standard jet tagging benchmark, it improves accuracy by 1.8 percentage points using 2.9 times fewer LUTs than the previous state-of-the-art sub-microsecond transformer. Its most efficient linformer variant with 64 input particles reaches 80.4% accuracy at 28k LUTs and 35 ns.
In briefASTROLABE: A 3D Particle-Level Surrogate for Cosmology Simulation via Physical Flow Matching
SummaryThis paper introduces Astrolabe, a generative surrogate model based on flow matching that acts directly on the 6N phase-space coordinates of N particles, avoiding a voxel grid. It uses approximate COLA-based coordinates as the flow's prior, with momentum and mass conservation enforced throughout flow time. In the Lambda-CDM cases tested, the model produces realistic fields with accurate power spectra, halo mass functions, and velocity and halo density profiles. It can also be fine-tuned for very large N, as shown by emulating a TNG50-3-dark simulation.
Ibrahim Elsharkawy, David Curtin, Yonatan Kahn et al.
In briefBUMP: A Body-Driven Physics Framework for Multi-Object Human-Object Interaction Generation
SummaryThis paper presents BUMP, a text-driven framework for generating human motions that interact with multiple objects. It first generates full-body human motion, then predicts trajectories for a few intentionally manipulated objects, and uses physics simulation to determine the motion of the remaining surrounding objects. The authors also introduce outcome-conditioned editing, which modifies the driver object's trajectory based on a desired responder behavior. Experiments report strong performance on single-object and multi-object interaction benchmarks.
In briefGraph neural networks for astrophysics and cosmology: a review
SummaryThis review paper surveys how graph neural networks (GNNs) are used in astrophysics and cosmology. It argues that graphs naturally capture relationships between astronomical objects, making GNNs well suited to physical data and more interpretable than other deep-learning models. The authors cover applications in cosmology, galaxy formation, dynamics and instrumentation. They find that outside instrumentation, only about one in six articles applies GNNs to real data rather than only simulations, and they offer best practices, including when not to use a graph.
In briefText-to-image generation without language pretraining
SummaryThe paper introduces Halcyon, a family of early-fusion multimodal autoregressive models trained from scratch on GPIC, a dataset of 100M image-text pairs, without internet-scale language pretraining. The authors report that Halcyon-0.52B and Halcyon-1.6B outperform comparison models on gFD while using substantially less pretraining compute. Controlled experiments indicate that text-only pretraining does not improve text-to-image generation.
Keshigeyan Chandrasegaran, Yue Zhao, Armin W. Thomas et al.
In briefCan It Actually Work? Synthesizing Functional CAD Mechanisms
SummaryLLM agents can produce CAD assemblies that look plausible, but these designs often fail to perform their intended physical functions. The authors propose FunCAD-Harness, which turns abstract requirements into candidate-specific physical tests and uses simulation feedback to revise designs. They also introduce FunCADBench, a benchmark covering 46 mechanism families with multi-step operations, and report that the framework substantially improves functional success rates over simulation-augmented baselines.
In briefOpenVLHarness: A Framework for Visual Tool Use with Multimodal Memory
SummaryThis paper presents OpenVLHarness, a reasoning framework for multimodal models built around a capability layer and persistent multimodal memory. The capability layer provides unified access to heterogeneous tools and converts their outputs into observations the model can read, while the memory stores images, structured data, and web references and allows generated code to combine results across tools. Evaluated on 22 benchmarks with eight backbones, it improves average scores by 9.7 to 12.6 points over matched no-tool baselines.
In briefVideoMaMa++: Temporally Consistent Video Matting with Preserve-and-Refine Tokens
SummaryThis paper presents a diffusion-based video matting framework that extracts precise foreground transparency from video. Because diffusion backbones can only process 6 to 24 frames at a time, results often flicker and become discontinuous at chunk boundaries. The authors use separate RGB and mask token streams, learnable preserve and refine tokens, and a sliding-window scheme with a shared reference anchor to keep mattes consistent across chunks. The authors report better fine-detail accuracy and temporal consistency than prior diffusion methods on standard benchmarks and long real-world videos.
In briefReinforcement learning that corrects visual reasoning errors using visual evidence
SummaryThis paper studies reinforcement learning for multimodal visual reasoning, where token-level corrections from a privileged teacher can be contaminated by language priors, response style, or the teacher's view. The authors propose Visual-Attribution Gating (VAG), which admits only corrections supported by the contrast between evidence-present and evidence-degraded teacher views. Across six visual reasoning benchmarks, VAG reportedly outperforms GRPO and OPSD methods, achieving new state-of-the-art average performance.
In briefVitronEdit: One Framework for Editing Images, Videos, and 3D Scenes
SummaryThis paper presents VitronEdit, a single model that edits images, videos, and 3D scenes. It uses a shared sequence representation and progressive multimodal training to transfer editing knowledge from image and video data to the data-scarce 3D setting. It also adds geometry-aware post-training that uses a pretrained geometry reconstruction model to supply depth and camera-pose supervision. The authors report that applying geometric supervision only to 3D samples gives the best overall results, and image and video editing also improve.
In briefRopedia W0: A world model for locomotion and manipulation across human and robot bodies
SummaryThis paper presents Ropedia W0, described as the first egocentric world foundation model for combined locomotion and manipulation across humans and robots with diverse embodiments. The authors build Xp-Data, with 10k hours of human egocentric data and 3k hours of robot data, and use a unified kinematic representation. The model reports state-of-the-art results on five cross-embodiment benchmarks and supports real-time, minute-long rollouts after distillation.
In briefInfiniDream: Generating Open-Ended Photorealistic 3D Worlds with Procedural Methods
SummaryThis paper presents InfiniDream, a vision-language model (VLM) driven framework for open-ended procedural 3D world generation. It uses existing procedural algorithms to create novel photorealistic assets while preserving editability and seed-based diversity, and adds a text-to-world pipeline that combines semantic scene planning with frustum-guided grid placement. The authors report improvements across four perceptual metrics on 100 object and 20 scene generation tasks.
Yi-Chuan Huang, Hao-Jen Chien, Zheng-Hui Huang et al.
In briefCoopUQ: Multi-robot navigation among people that forecasts crowd occupancy and its uncertainty to stay safe.
SummaryThis paper proposes CoopUQ, a framework that lets several mobile robots navigate together in environments shared with people. It fuses information across robots to predict future occupancy grid maps and uses these forecasts and their uncertainty to guide a constrained multi-agent reinforcement learning policy that penalizes paths through predicted or highly uncertain regions. The authors report better prediction and safety than representative baselines in Unity3D simulations, including out-of-distribution scenarios, and show robustness in real-robot experiments.
In briefManga2Animate: A multi-agent system that turns manga into faithful animation via script and storyboard planning
SummaryManga2Animate is a multi-agent framework for adapting existing manga into animation that stays faithful to the source. It splits the task into two intermediate outputs, an animation script and a visual storyboard, each produced by a creator agent and checked against the source panels by a reviewer agent. These are refined iteratively before the final animation is synthesized. The authors report that experiments and a user study show more source-faithful and shot-consistent results than state-of-the-art methods.
Sheng-Wei Zhang, Hsi-An Chen, Yan-Ting Liou et al.
In briefGenOlympiad: A benchmark that stress-tests top image generators on their hardest tasks
SummaryThis paper introduces GenOlympiad, a benchmark of very challenging text prompts across nine capability areas, used to stress-test leading text-to-image models. The prompts are built so that whether an image matches them can be checked visually, such as by counting objects. The authors also propose an automated grading framework that splits each prompt into easy and hard questions and uses drawn guide lines to ground evidence. Among more than 20 models evaluated, the best, Nano Banana Pro, scores only 23.9 out of 100 on the strict measure.
In briefWhere Physical Self-Improvement Ability Comes From: An Observational Scaling Law for LLMs as Policies
SummaryThe paper systematically studies the mechanisms, scaling behavior, and generalization of physical recursive self-improvement in vision-language models. It introduces PhysRSI-eval, a five-axis metric, and evaluates several VLMs on 18 manipulation tasks under a fixed interaction budget. The authors propose a Physical Barrel Hypothesis and analyze scaling by predicting performance from public capabilities and early physical feedback, plus workspace interventions showing which gains persist.
In briefWeAct: Letting Agents Switch Capability Adapters as Actions
SummaryThis paper proposes WeAct, which trains separate LoRA adapters for different capabilities on a frozen backbone and adds adapter selection to the agent's action space. This lets the agent compose capabilities on demand during long tasks. Across eight deep-research benchmarks, the authors report that Qwen3-32B with WeAct improves average accuracy by 6.2% over full fine-tuning and by 27.5% over the strongest static LoRA-merging method. The approach also supports modular continual adaptation.
In briefTraining an Agent to Imagine Action Outcomes Before Acting
SummaryThe paper introduces Agentic Imagination (Agima), a framework that trains an agent to propose actions, generate images of their likely outcomes, and reason over those imagined futures before acting. It first uses supervised fine-tuning to unify policy and world modeling in one omni model, then uses agentic reinforcement learning to teach effective planning with imagined futures. Across Maze2D, Sokoban, Pacman, and PushT, Agima generalizes to harder settings unseen during training.
In briefAnyGoal: Testing whether world models generalize beyond expert demonstrations
SummaryThis paper studies whether world models can recombine familiar state and action factors into valid environment dynamics for new goals. The authors introduce AnyGoal, a benchmark of goal-shifted tasks built from common world-model environments, along with an open-source library. Experiments across several environments and model families show that current models consistently fail under these recombinations, and the paper analyzes where generalization holds and where it breaks down.
Heejeong Nam, Divyat Mahajan, Quentin Le Lidec et al.
In briefLatComm: Reusable context compression for passing information between language models
SummaryThe paper presents LatComm, which encodes a shared context once into a fixed-size packet using a frozen language model, then passes it to a different frozen model that selects what to read per question via cross-attention. Across 12 multiple-choice question-answering datasets, it improves average accuracy over single-vector baselines by up to 10.2 points. On QuALITY, it cuts the receiver's input by 57 times and answers questions about an article 3 times faster.
In briefNewton-based methods for bilevel optimization without lower-level strong convexity
SummaryThis paper studies bilevel optimization, a framework used in tasks such as hyperparameter tuning, data reweighting, and meta-learning. It proposes two Newton-based methods that work without lower-level strong convexity, with convergence guarantees and improved complexity bounds compared with GALET and sl-BAMM. It also introduces GNBO, a generalized framework, and evaluates the methods on synthetic problems, data hyper-cleaning, and GPT-2 data reweighting.
In briefOntology-free medical record modeling via a shared discrete code vocabulary
SummaryEHR foundation models struggle because the same clinical concept can appear under different codes, names, or free text, and standard tokenizers treat these as unrelated. The paper introduces MEDIATOR, a discrete tokenizer that maps any text-expressed clinical concept into a shared vector-quantized codebook without manual mappings or a fixed vocabulary. The authors report a 4.0-point mean AUPRC gain over the strongest prior tokenizer on 20 MIMIC-III tasks and an 8.6-point gain in zero-shot transfer to 14 eICU tasks.
WhyProposes text-based discrete ontology-agnostic tokenizer for cross-hospital EHR transfer, with thorough MIMIC-III and eICU comparisons and same-encoder ablation controls.
In briefSpecKCast: Spectral Kalman Forecasting to Speed Up Diffusion Models
SummaryThis paper proposes SpecKCast, a training-free method that speeds up diffusion transformers used for image and video generation. It applies a 2D discrete cosine transform to the model output and tracks each frequency coefficient with a lightweight Kalman filter, running the transformer only when the forecast uncertainty exceeds a threshold. Experiments on FLUX.1-dev, Qwen-Image, Wan2.1-14B, and HunyuanVideo report a better balance between acceleration and fidelity than existing caching and forecasting methods, with up to 6x speedup and less than 0.05 GiB of extra memory.
Javad Rajabi, Hakki Can Karaimer, David B. Lindell et al.
In briefSame Simulation, Different Evidence: When Retrieved Tool Output Lowers LLM Agent Accuracy
SummaryResearchers tested LLM agents answering 158 multiple-choice questions with the USGS geochemistry simulator PHREEQC, comparing reading its report in context against retrieving sections through an index. Under retrieval, some GPT models lost accuracy, often skipping the report and answering from their own output files before rounding to a neighboring option. The authors call this failure evidence substitution.
Ke Zhang, Sahchit Chundur, Jerry Peprah Owusu et al.
In briefKronecker Delta Rule for Stable Long-Context Linear Attention
SummaryThis paper proposes KronLSTM, a linear attention architecture built on the Kronecker Delta Rule, a new update rule for associative matrix memories. The rule revises memory only along a single direction, reducing interference with stored information compared with delta-rule updates. The authors report that the memory dynamics form a Markov chain converging to a steady state, with a uniformly bounded expected memory norm. Experiments indicate competitive or better results than existing linear attention models, especially on long-context benchmarks. Code and models are open source.
Anamaria-Roberta Hartl, Levente Z\'olyomi, Michael List et al.
In briefCompeting Learning in Hybrid Language Models: Recurrence and Attention Racing to Learn Retrieval
SummaryThis paper studies hybrid language models that combine linear-time recurrence layers with softmax attention. It finds that different mixer types compete to learn the same function, and once one layer learns it, the learning signal for the others is suppressed. Under typical data distributions, recurrence learns retrieval faster than attention, so attention often fails to learn retrieval. The authors propose a staged pre-training procedure that stops recurrence weight updates midway, reporting gains on models up to 1B parameters in language modeling, common-sense reasoning, and retrieval tasks.
In briefProteinGeoBench: Frontier Structural-Biology Agents Are Fluent at Lookup but Lost at Geometry
SummaryThis paper introduces ProteinGeoBench, a suite of 131 protein-structure tasks used to evaluate four frontier models and 26 models overall. Most frontier agents retrieve structural metadata better than they compute geometry from coordinates. The authors find that rankings and scores also depend on evaluation settings such as the number of runs, the harness, and network access. They recommend reporting results by task family along with these configuration details.
In briefTraining computer-use agents in synthetic, controllable software environments
SummaryThis paper presents WildAWM, a scalable pipeline that synthesizes controllable environments and tasks for training general computer-use agents. The environments jointly support graphical interfaces, model context protocol services, and command-line interfaces, and are generated entirely with open-source models. The pipeline produces 341 long-horizon environment-task pairs, of which GPT-6-Astra solves only 60%. Two Qwen models trained exclusively in these synthetic worlds show substantial gains across 8 out-of-distribution benchmarks.
WhyUnifies GUI, MCP and CLI into synthesized executable environments with execution-based quality gates; gains across 8 out-of-distribution benchmarks give solid evidence and moderate influence on comput
In briefKEEP-Bench: Testing Whether Context Compression Keeps Key Evidence
SummaryThis paper introduces KEEP-Bench, a diagnostic benchmark of 4,102 cases that directly checks whether compressed context remains faithful, covering plausible reference resolution, explicit qualifiers, and minority perspectives. Even the strongest models were unreliable, with the best open-weight model (31B) preserving all required evidence in 66.8% of cases and the best proprietary model in 61.6%. The authors also propose an Evidence Preservation Return (EPR) metric, which peaks at moderate model scales (4B-8B) where evidence recall remains low.
In briefRefCraft: A Benchmark for Visual-Reference-Based AI Agent Web Development
SummaryThis paper introduces RefCraft, a benchmark of 300 realistic multimodal webpage-development tasks that pair natural-language instructions with screenshots of one to five reference webpages. AI agents must selectively use these references while rendering, visually checking, and iteratively refining their output. The generated webpages are assessed with independently verifiable keypoints and both programmatic and vision-language model judges. Across nine frontier (M)LLMs, the best system scores only 52.12 out of 100, indicating the task remains difficult for current models.
In briefMemT-1.0: A memory foundation model for memory-native evolution
SummaryThis paper introduces MemT-1.0, a foundation model architecture that treats memory as an intrinsic part of model computation, built on Metis with Gated DeltaNet-2 and Mixture-of-Storage to scale native memory. It is trained in two stages, continued pre-training and post-training, using 34B tokens, and released as MemT-1.0-Nano and MemT-1.0-Flash. The authors report that it outperforms parametric memory baselines on average, is competitive with retrieval-augmented methods, and at 128K history length, Nano achieves about 1.86x inference speedup and 66x less persistent storage than full context KV cache.
In briefSpecMRL: Redesigning Matryoshka Representation Learning with a Parametric Spectrum
SummaryThis paper proposes SpecMRL, an extension of Matryoshka Representation Learning (MRL) that reorganizes the feature space using a parameter-derived spectral basis ordered by feature importance. Prefix readout is then performed in this learned spectral space. Experiments on MTEB show SpecMRL outperforms MRL at all evaluated reduced dimensions, with larger gains under stronger compression. It also surpasses full-dimensional MRL while using only half the embedding dimensions.
In briefAfriBED: A Speech Recognition Bias Evaluation Dataset for Low-Resource African Languages
SummaryThis paper introduces AfriBED, a dataset for evaluating bias in speech recognition systems for two low-resource African languages, Bemba and Nyanja. It is designed to detect gender and speaker-type bias. The Bemba portion contains about 9 hours of speech from 80 speakers, and the Nyanja portion contains about 18 hours from 127 speakers. Baseline experiments show the dataset can be used to examine gender and speaker-type bias in these languages.
In briefOpen-Form Geometric Style Transfer: applying a learned geometric style to new, unrelated shapes
SummaryThe paper studies how a geometric design style learned from one shape family, such as stylized Latin letters, can be applied to substantially different forms, including Greek glyphs and human poses. Target forms are represented by structural sketches, and the method separates structural adherence from style realization using a Geometry LoRA and a style-specific LoRA. Experiments and ablations indicate that this disentanglement and a universal geometry prior built from diverse synthetic sketches enable transfer to unseen forms.
In briefASTOR: A grammar-guided method and benchmark for autoformulating soft-constraint optimization problems
SummaryReal-world decision problems include hard constraints that must be satisfied and soft constraints (preferences) that may be violated. The paper introduces ASTOR, an inference-time framework that maps a natural-language decision problem to a grammar-guided abstract syntax tree and then deterministically compiles it into executable optimization code. The authors also present SoftOR, a benchmark of natural-language problems focused on soft-constraint formulations.
In briefReuF2F: A Benchmark Testing Coding Agents on Formalized Undergraduate Research Problems
SummaryThis paper introduces ReuF2F, a Lean 4 benchmark of 244 formalized open problems drawn from Research Experiences for Undergraduates across many mathematical subfields. Each problem and its logical negation are placed in a Lean project inside a Docker container, and a unified harness called Leanimum-agent, built on mini-SWE-agent, enables fair comparisons across models. The best-performing model resolves only about 40% of the problems, indicating substantial room for improvement on formalized open research problems.
In briefNext Best Action: Causal Constrained Offline RL for Targeting Business Actions
SummaryThis paper formulates Next Best Action (NBA) systems, which choose customer interventions, as a Constrained Markov Decision Process. It integrates off-policy causal correction of customer dynamics into a linear programming formulation to maximize long-term value while enforcing business constraints. The method is validated on a rideshare coupon targeting case study using RFM customer states and model-based off-policy evaluation. The optimized policy improves customer lifetime value by 19% over the estimated behavior policy and 8% over a myopic baseline.
Surya Ramachandiran, Yiming Li, Juan Mancilla-Caceres et al.
In briefReal-time radar scene understanding for segmenting objects
SummaryThis paper presents StreamRPS, a framework for streaming panoptic segmentation of FMCW spinning radar data. It combines an automatic labeling pipeline based on LiDAR scene completion with a real-time encoder-decoder network built on a radar-specific hierarchical Transformer. Experiments on the SAX and BOREAS datasets validate the method, which processes sectors at over 400 FPS in a 4-sector streaming setting. The authors also release manual annotations for 100 SAX radar scans.
In briefPICA-WAM: Making robot world-action models ignore irrelevant visual changes
SummaryWorld Action Models predict future video and generate robot actions, but they can be disturbed by visual appearance changes unrelated to the task. The paper proposes PICA-WAM, a self-supervised framework where an EMA teacher sees the original observation and an online model sees a perturbed one, aligning its shallower action representation to the teacher's deeper target. The authors report improved robustness under visual perturbations across four WAMs on RoboTwin 2.0, LIBERO-Plus, EBench, and a real robot, with nominal simulation performance preserved and higher real-robot success rates.
WhyCasts perturbation invariance as cross-appearance self-alignment for WAMs, validated across four models, multiple benchmarks, and a real robot.
In briefState World Models: Rule-Governed Video Generation with Hidden State
SummaryThis paper presents Stateful Harness, a framework that separates evolving world-state updates from video rendering so that actions follow persistent rules even when outcomes depend on hidden state. Each turn, it resolves the user's action, compiles the resulting event into an instruction, renders it with a frozen video model, and uses the clip to update the state. On WorldState-Bench, a set of 50 counterfactual twin pairs, the full harness reaches 92% rule compliance with the same frozen model, compared with 32% for rule-only prompting.
In briefPhysical AI: Concepts, Techniques, and Challenges
SummaryThis monograph defines physical AI as artificial systems that perceive, model, and act on the physical world through continuous feedback under real-world constraints. It distinguishes physical AI from digital AI, embodied AI, cyber-physical systems, and digital twins. The work reviews methods such as world models, embodied foundation models, motor control, and multi-agent coordination, along with applications in robotics, manufacturing, and infrastructure. It also examines challenges including sim-to-real transfer, safety, and governance.
In briefSafety Mid-Training: Building LLM Safety into Core Reasoning Ability
SummaryThis paper proposes a "safety mid-training" stage between pre-training and post-training, where safety and general reasoning are learned together through a shared reasoning structure. Compared with Qwen3-1.7B, 4B, and 8B base models, the method reduces average attack success rate by 72% to 81% while also improving general capability. The authors report that these benefits persist after subsequent post-training.
WhyIntroduces safety mid-training between pretraining and post-training, jointly modeling safety and reasoning; novel framing, with experiments across model sizes and follow-up training, though details a
In briefWhat language adds to pathology vision-language models, and at what cost: four controlled tests
SummaryThis paper tests how much language contributes beyond the image-processing part of pathology vision-language models, reusing released checkpoints and public data and changing one input at a time. Report text adds some prognostic signal across 1,633 TCGA patients, while prompts add no measurable inference accuracy and slide question answering is weakly tied to the slide. The authors propose minimum controls for future evaluations to expose these failure modes.
In briefScaling parameter-efficient fine-tuning for predicting brain responses to stimuli
SummaryThis paper applies parameter-efficient fine-tuning (PEFT), which trains only a small number of parameters, to adapt large pretrained models for brain encoding, predicting neural responses to stimuli. The authors predict multiple timepoints in one pass using the model's causal structure and train low-rank adapters (LoRA) on multimodal backbones. The method reportedly achieves the best published performance on the Algonauts 2025 Challenge.
In briefLearning to Refer: Collaborative Inference Across Decentralized Federated Learning Clients
SummaryThis paper addresses decentralized federated learning, where clients have complementary expertise but each query is only handled by the client receiving it. It proposes EdgeDRM, which lets queries be referred between neighboring clients under a budget, using a learned density-based routing potential to choose the path. Training and inference use only neighbor-local information. Experiments on benchmark and medical data show it outperforms graph-matched decentralized baselines in most settings, with the clearest gains under sparse connectivity.
In briefTesting whether LLM pipelines built in English work across other languages
SummaryThis paper examines whether large language model pipelines developed and validated in English can be reused in other languages, measuring accuracy, transfer of reasoning-effort settings, and behavioural stability. It uses three tasks: cheque extraction, Mintaka question answering, and a FinanceBench pilot on SEC 10-K filings. The authors find accuracy is broadly reusable, but per-domain and per-language validation remains necessary.
Joong Ho (Joshua) Choi, Sherine George, Jiayang Zhao et al.
In briefTwo hidden layers suffice: the Caterpillar Net, a construction for maxima of ReLU-style functions
SummaryThe paper studies how to represent continuous piecewise linear functions, which are central to understanding the expressiveness of ReLU networks. It presents the Caterpillar Net, a nonrecursive construction that computes the maximum of N inputs using two hidden layers of widths O(N²) and 2^{O(N log N)}. The authors note that the construction has a direct geometric interpretation.
In briefDoes More Training Data Help Tabular ML Generalize? A Study at Scale
SummaryThis paper studies how training-data size affects tabular machine learning. The authors assemble eight real-world tabular datasets with up to about ten million training examples, keeping the natural distribution shifts and using random-split evaluation as a control. They find that GBDTs and MLP-based models show broadly similar scaling trends, while tabular foundation models vary more, with recent versions benefiting more consistently from larger contexts. Under distribution shift, in-distribution and out-of-distribution performance mostly improve together at smaller scales but can diverge at larger scales.
In briefSOMA: A Framework Where a Fast Model Monitors and Steers a Slow Reasoning Model
SummaryThis paper proposes SOMA, a framework in which a fast System One model continuously monitors a slower System Two reasoning model and intervenes when needed. It has three layers (scenario, awareness, and control) and requires no training, sharing one S1 model and controller across tasks. The paper reports that it reduces reasoning tokens on overthinking problems while improving average accuracy, and raises abstention on unanswerable questions.
In briefTraining models on a mix of true and machine-labeled data using stochastic first- and zeroth-order optimization
SummaryTraining machine learning models usually needs ground-truth labels, which are costly and hard to scale when produced by humans. This paper considers a setting where only some samples have true labels and the rest are labeled by an ML predictor that may assign incorrect labels. The authors propose first-order and zeroth-order stochastic optimization methods for training on such mixed datasets. Experiments range from linear regression to fine-tuning a large language model (LLM) to show the methods' effectiveness.
In briefLG-Bench: A graph-structured benchmark for testing AI on life sciences knowledge
SummaryThis paper introduces LG-Bench, the first graph-structured evaluation benchmark for life sciences, with over 10,000 multiple-choice questions spanning medicine, biology, and chemistry. The questions are linked in a weighted graph based on their semantic relationships. Two new metrics, the Global Coherence Score and the Knowledge Balance Score, use this graph to assess model consistency and reveal conceptual blind spots.
In briefRESIN: Multi-chain reasoning using entity-structured indexing to locate evidence
SummaryThis paper introduces RESIN, a method for questions that require multiple chains of reasoning. It splits a question into focused one-hop needs, retrieves a small set of source-grounded evidence for each hop, and places this evidence in a compact reasoning buffer until final composition. The authors report higher context recall and answer relevancy than published baselines on the full HolisQA benchmark.
In briefV-JEPA Hand: a fast model that recovers 3D hand and camera paths from first-person videos
SummaryThis paper presents V-JEPA Hand, a feed-forward video model that jointly reconstructs both hands' motion and camera motion from egocentric video clips. It uses a self-supervised V-JEPA 2.1 encoder and a query-based predictor to output per-frame hand poses, hand presence, and camera motion, which place hand trajectories in a shared world frame. On ARCTIC, HOT3D, and EgoStandard, the single model reports the lowest coverage-aware mean per-joint position error among comparable baselines, and it runs at 93.4 FPS on an RTX A6000.
In briefRobustness of internal signals for detecting hallucinations in vision-language models
SummaryThis paper examines whether methods that detect object hallucinations in vision-language models using internal signals are robust. The authors find that a simple baseline based on a generated word's token position in the response performs competitively, and some detection methods appear to rely on this position as a shortcut. Strong detection results are often driven by language-based hallucinations, and methods relying on token position are less visually grounded. Combining different methods in an ensemble still yields performance gains.
Allison Chen, William Yang, Salma Abdel Magid et al.
In briefTopological Sheaves of Irreducibility: a theory of obstructions in multi-agent tasks
SummaryThe paper introduces Topological Sheaves of Irreducibility (TSI), a sheaf-theoretic framework that separates task-level compatibility obstructions from a system's capability. Nonzero cohomology defines a structural signature, and registered interventions test whether a system can meet the task. In experiments on finite one-round models, an all-agent additive primitive solved decision and repair exactly on 15,625 degree-1 inputs, while on a degree-2 task both tasks became exact only when a four-agent primitive was added.
In briefMonsoon: A Framework for Collecting Spontaneous Speech Data for Low-Resource Speech Recognition in 50 Languages
SummaryThe paper presents Monsoon, a framework for collecting large amounts of unscripted spontaneous speech to improve automatic speech recognition in low-resource languages. It elicits conversations between paired contributors, filters recordings with automated classifiers, corrects transcripts through five levels of human review, and selects segments for lexical and acoustic coverage. The authors report collecting speech in 50 languages and analysing seven corpora of 3,500 hours each.
WhyProposes a scalable spontaneous speech collection framework with multi-level quality review, evaluated across 50 languages and seven corpora against baselines, showing practical gains for low-resource
In briefHarnessPro: Cost-Aware Self-Improving AI Agent Harnesses via Pareto Selection
SummaryThis paper presents HarnessPro, a label-free method that evolves AI agent harnesses to improve task performance while reducing execution cost. Specialized agents diagnose execution trajectories and propose updates, and the system returns lite, medium, and heavy harness variants selected by normalized weighted distance to suit different budgets. On Terminal-Bench 2, SWE-Bench Pro, and ProgramBench, the heavy mode with GPT-5.5 and Codex reportedly gains 2.0 percentage points in pass rate and cuts per-task cost by 26.1% relative to RHO.
In briefVerifierBench: Testing whether AI agents can write checkers for their own task answers
SummaryThis paper introduces VerifierBench, a benchmark of 222 tasks from six agent benchmarks with 4,993 hidden test cases, to evaluate whether LLM agents can write executable verifiers that accept valid solutions and reject invalid ones. Across 19 models, normalized balanced accuracy ranges from 10.6 to 62.1, and generated verifiers reject invalid solutions more reliably than they accept valid ones. Most false rejections come from verifiers that accept no tested valid solution, and the authors suggest testing both correct-solution acceptance and acceptance of other permitted solutions.
WhyIntroduces a new verifier-generation benchmark; evidence from 222 tasks, 4,993 hidden tests, and 19 models supports its analysis of overly restrictive verifiers.
In briefChannel-level mask guidance for slimmable image compression, a method letting one learned image codec run at flexible complexity and bitrate.
SummaryThis paper addresses the lack of scalability in learned image compression for devices with different computing and storage limits. It proposes a learnable channel-wise mask with sparsity regularization to select non-contiguous top channels, then converts the trained model into a standard slimmable network. On Kodak, narrow models show large rate-distortion gains at low bitrates, and a single variable-rate model replaces 40 independent models.
In briefUsing Simulation Priors to Speed Up Model-Based RL Adaptation for Autonomous Racing
SummaryThis paper introduces Prior-Informed Adaptation (PIA), a framework that pretrains a world model on simulated trajectories and then adapts it with limited real-world interaction using uncertainty-regularized planning. PIA treats simulated data as biased priors rather than requiring an accurate simulator. In simulation, it adapts faster than strong baselines across track and surface shifts, with lower lap times and fewer failures. On a full-size autonomous race car, it achieves reliable few-shot adaptation with less than one hour of on-vehicle interaction.
In briefTokenGate: Faster, Lower-Memory Long-Input Processing via Learned Token Pruning
SummaryTokenGate is a small, plug-in component that lets existing trained language models learn to prune less important input tokens end to end, speeding up prefill on long inputs while using less memory. To handle the non-differentiable top-K selection, it introduces token-gated attention, inspired by Mixture-of-Experts, which reweights attention using token importance scores. The paper reports 1.4 to 11.3 times faster time-to-first-token and 1.5 to 3.4 times lower peak memory across tasks with no accuracy loss. It works with both standard and hybrid-attention models such as Qwen3 and Qwen3.5.
In briefBeyond Omni: A Unified MLLM for Understanding Embodied Sensory Data
SummaryThis paper proposes Unisense, a unified embodied multimodal large language model that handles diverse physical sensory modalities within one shared architecture. Training has two stages: first aligning sensory representations with language, then tuning modality-specific experts inside the shared LLM. The authors build Unisense-1M, a corpus of 1 million samples across 8 sensory modalities, and evaluate on 12 benchmarks covering depth, point cloud, IMU, tactile, thermal, radar, motion, and smell. Unisense performs strongly and outperforms both generalist and specialized models on most benchmarks.
In briefTrajectory Replay: Continuous-Time Model Learning Without Forgetting
SummaryThis paper proposes trajectory replay, a method that helps pre-trained continuous-time generative models learn new data sequentially without forgetting earlier tasks. Instead of generating synthetic samples at every step as generative replay does, it trains an updated velocity field to match a frozen historical velocity and the new task's velocity, weighted by a stability-plasticity coefficient. It outperforms generative replay and distillation baselines on class-incremental image generation. In two biological case studies, it extends generative perturbation models to new contexts and batches while preserving treatment-effect prediction.
Alessandro Palma, Charles Jones, Tristan Cinquin et al.
In briefGradient descent reaches any accuracy in a fixed number of large steps for multiclass logistic regression
SummaryThe paper shows that gradient descent with a sufficiently large step size can reach any chosen optimality gap on the multiclass logistic regression loss within a number of iterations that does not depend on that gap. The step size depends on the target accuracy and properties of the linearly separable data, while the iteration count depends only on the data. The proofs rely on the known fact that large-step gradient descent behaves like the perceptron algorithm.
In briefrecAMPLIFY: Large-Scale Pretraining of Recursive Protein Language Models
SummaryThis paper adapts recursive pretraining to masked protein language models, using a weight-tied Transformer core that refines residue hidden states over an adjustable number of cycles. The resulting recAMPLIFY model has 155M unique parameters, trained on the same 2T tokens as AMPLIFY-350M, and uses 56% fewer parameters while improving CASP contact recovery and three of four supervised tasks, including a 22.5% gain in fold accuracy. The authors also introduce Gated Mixture Recurrence (GMR), which uses learned residual gates and probability-weighted hidden states to reduce expected layer depth by up to 37%, with accuracy trade-offs that vary by task.
