In briefRADE: A Framework for Routing, Memory, and Escalation in Safe LLM Systems
SummaryThis paper treats a deployed language model system as a deployment policy that decides when to retrieve, remember, call a larger model, verify, or abstain. It introduces risk-adjusted deployment efficiency (RADE), a system-level objective that nets out inference, retry, oversight, and priced expected failure costs, while treating unpriced harms as hard budgets. The authors examine five deployment regimes and propose a reporting protocol based on risk-coverage behavior, end-to-end cost, memory exposure, and escalation load. They argue that frontier capability should be allocated as a resource rather than used by default.
In briefMeasuring AI-generated content in U.S. political campaigns and whether it is disclosed
SummaryThis study measures how much AI-generated content U.S. political candidates produce and share during campaigns, using text and images from 12,977 candidates' websites, X accounts, and emails from 2023 to 2025. Detectable AI-generated content rose significantly over the period, and campaigns that used it frequently often did not disclose it. Incumbents produced less of it than challengers, and candidates in states with disclosure or prohibition laws did not produce measurably less than those elsewhere.
In briefInside the Trading Loop: AI Agents Making Stock Trading Decisions
SummaryThis paper presents a hybrid trading system in which LLM-based company-news and macroeconomic agents send structured reports to a coordinator that sets stock weights and cash reserves, alongside momentum-based portfolio construction and rule-based execution. Evaluated on a broker-recorded paper account from March 11 to August 28, 2026, the system returned 34.5%, versus 14.4% for the S&P 500 fund SPY, though with a larger maximum drawdown. Design comparisons also cover cash authority and report formats.
In briefComparing simple generalized means with a large AI forecasting model for emergency calls
SummaryThis study compares generalized means (Hölder and Lehmer means) with a fine-tuned TimeGPT model for forecasting daily emergency call volumes over one to seven days. Both generalized means perform almost identically and outperform TimeGPT for short-term forecasts, while TimeGPT has lower error at seven days. Diebold-Mariano tests show no statistically significant differences between the methods at the evaluated horizons.
Feriel Fass, Djemel Ziou, and Marie-Flavie Auclair-Fortier
In briefOptimizing Multi-Tier Intermediation in Programmatic Advertising: A DAG-Based Mechanism for Supply Path Optimization
SummaryThis paper addresses efficiency losses in programmatic advertising caused by multi-hop supply chains, often called the “ad-tech tax.” It applies Directed Acyclic Graph intermediation models to supply chains traced through the OpenRTB SupplyChain object. The authors introduce multi-attribute utility functions and mixed-integer linear programming to handle substitutable routing and percentage-based fees. They also outline a computational method using public schain and sellers.json data to compare optimal Nash social welfare with current uncoordinated market outcomes.
In briefA trust region method for learning Schrödinger bridges, which are tools for transforming one probability distribution into another.
SummaryThis paper studies how to numerically approximate Schrödinger bridges when only an unnormalized target density is available rather than samples. The widely used iterative proportional fitting procedure and its variants often suffer from numerical instabilities. The authors propose a trust region variant that keeps each update at a fixed Kullback-Leibler divergence from the current iterate, which significantly improves stability. They show this method can be interpreted as a mirror descent scheme with adaptive step size, and demonstrate its effectiveness on several challenging sampling tasks.
Takeshi Koshizuka, Denis Blessing, Max Zimmer et al.
In briefPA-SaR: A Framework for Coordinating Multi-Robot Search and Rescue
SummaryThis paper presents PA-SaR, a framework that coordinates local control, learning-based exploration, supervisory task allocation, and adaptive victim guidance for autonomous search and rescue by multiple robots. Simulation studies show a trade-off between collision avoidance and search completion, and that supervisory allocation improves search completion beyond information sharing alone. The results also show that more exploration does not necessarily yield more rescues, while adaptive guidance increases rescues compared with continuous leading.
Jocubas Dimsa, Maria Grouev, Filip Keerberg et al.
In briefMICE: A foundation model for analyzing social behavior in mice from pose data.
SummaryThis paper introduces MICE, a hierarchical foundation model for quantifying mouse behavior from pose keypoints. It is pretrained jointly on six mouse-pose datasets covering one to four mice, combining per-mouse kinematic encoding with a social encoder that captures interactions. A single checkpoint works for any group size, and the model matches or exceeds prior keypoint-based baselines across four evaluation protocols.
In briefGaze-Triggered Digital Twins for Remote Collaboration
SummaryThis paper presents Feel My Focus, a mixed reality demonstration in which sustained gaze at a physical object triggers single-image capture, attention-guided segmentation, and one-shot 3D reconstruction. The result is a lightweight digital twin that a remote partner can inspect and discuss in situ. The system prioritizes responsiveness and experiential continuity over photometric fidelity. The authors report observations from a public demonstration, covering design trade-offs, failure modes, and user control considerations.
Shih-Yu Ma, Suibi Che-Chuan Weng, Ming-Yuan Pai et al.
In briefAn interpretable fuzzy tree predicting turbofan remaining useful life, with per-component diagnostics
SummaryThis paper presents an interpretable genetic fuzzy tree that predicts turbofan engine remaining useful life from cruise-condition sensor means in the N-CMAPSS dataset. The tree follows the engine hierarchy, moving from sensors to component health, bounded component damage, spool damage, and RUL, with each component's diagnosis inspectable rule by rule. On the native split it reaches 8.44 cycles RMSE on the multi-fault subset and 11.51 across all modes, using 333 free parameters.
In briefHow Eye-Scanpath Classifiers Depend on Their Inputs: Perturbation Profiles Across Reading, Code and Video Tasks
SummaryThe paper proposes a perturbation profile for analysing which properties of eye-movement scanpaths a classifier relies on. Five transformations each destroy one property, the classifier is refit, and the drop in participant-held-out AUC is recorded. Profiling seven descriptors on twelve binary problems from three public datasets, the authors find that some apparent comprehension contrasts are confounded by stimulus layout.
In briefGenISys: Integrated Ship Design Using Subsystem-Conditional Diffusion
SummaryThis paper presents GenISys, a framework for generating integrated ship designs, demonstrated on early-stage bulk-carrier design. It uses a simultaneous conditional diffusion model (SCDDPM) that generates subsystems together so each responds to changes elsewhere during generation. In a study targeting lower cost per tonne of cargo for a 5,000 nautical mile voyage, SCDDPM met the engineering constraints and low-cost target in 55.70% of requests, versus 51.10% for a monolithic model and 47.54% for a sequential model.
In briefODSA: An LLM-directed framework that orchestrates test-time scaling methods
SummaryThe paper presents ODSA, a framework in which an orchestrator language model directs existing test-time scaling methods as actions, including discovering new solutions, comparing solutions to build better ones, and accepting and verifying candidate answers. Stateless workers carry out these actions while the orchestrator allocates computation and passes earlier results to later steps. On Humanity's Last Exam, ODSA improves accuracy by 4.7 to 18.0 points across three models compared with a single direct model call. On LiveCodeBench it improves accuracy by 0.6 to 14.0 points.
In briefForecasting Daily Emergency Call Volumes with Univariate and Multivariate TimeGPT
SummaryThis study tests TimeGPT, a pretrained time-series foundation model, for forecasting daily emergency call volumes using 2,101 daily observations. It compares univariate and multivariate settings, with the multivariate models including temporal, meteorological, demographic, and socioeconomic variables. Results show SARIMAX performs best in the univariate setting, while NBR and PR perform best in the multivariate setting. The authors conclude that large-scale pretraining does not necessarily improve accuracy for this task.
In briefHeterogeneous multistep workflows for autonomous microscopy with reward design and knowledge transfer
SummaryThe paper proposes a budget-constrained framework for autonomous microscopy that decides between imaging and spectroscopy, image size, scan speed, and when to recalibrate the instrument. It transfers pre-experiment knowledge into the online campaign and tests it on a real electron energy-loss spectroscopy dataset across two campaigns, one with instrument drift. The study finds that different surrogate models and acquisition strategies suit response discovery and structure-property learning, and it measures the spectroscopy throughput cost of explicit registration under drift.
In briefRethinking depth geometry for single-image 3D object detection
SummaryMonocular 3D detection often estimates depth with the formula z ≈ f·H/h₂D, but substituting the 2D box height recovers the depth of an object's visible front surface rather than its center, causing size-dependent underestimation. The paper proposes MonoPCA, where PCDepth removes this bias with an orientation- and size-aware correction derived from perspective geometry, and GeoAlign transfers geometric features from vision foundation models into a lightweight student model. The authors report state-of-the-art results across four benchmarks.
In briefWriting decision-making policy weights with language models and checking them with model checking
SummaryThe paper proposes having large language models write all the weights of small neural networks directly, without gradient descent, so that a decision-making policy can be created from a task description alone. It combines this with probabilistic model checking: given a Markov decision process environment and a PCTL requirement, the LLM writes the policy, which is checked by the Storm model checker. Storm computes the probability that the policy satisfies the requirement, so every returned policy is verified.
WhyThe authors have LLMs write policy weights directly, verified via PCTL model checking with Storm. They report optimal probability across 18 runs on three environments with two LLMs.
In briefMartine: A benchmark for multi-view 3D surface reconstruction under controlled viewpoints, resolution, and lighting
SummaryThis paper introduces Martine, a benchmark for evaluating 3D object surface reconstruction in controlled conditions. It contains 28 objects, including transparent materials and fine structures, captured from 84 views at 9568 x 6376 pixels under single-light and multi-light setups. The dataset provides accurate 3D ground-truth scans, camera calibration, image-to-geometry accuracy measures, and 3D property masks for analysing errors. Methods including photogrammetry, multi-view stereo, neural implicit surfaces, Gaussian splatting and multi-view photometric stereo are evaluated, with a best Chamfer distance of 0.058 mm.
Robin Bruneau, Baptiste Brument, Frederic Giraud et al.
In briefPerturbing inputs to improve language model distillation during pretraining
SummaryThis paper examines whether training a student language model to match a teacher's behavior on perturbed inputs, beyond the training text, improves distillation. During an early phase of pretraining, the authors apply three perturbations (additive token noise, blending two documents, blending neighboring tokens) and spend the remaining budget on standard distillation. The best method outperforms standard distillation at 135M and 0.5B parameters, with gains of 0.025 and 0.020 nats respectively, and shows modest zero-shot accuracy improvements.
Ziyang Cai, Yihe Dong, Dimitris Papailiopoulos et al.
In briefBeyond Parameters: Syntax-Augmented Prompting (UD-PrMPT) for Low-Resource Machine Translation
SummaryThis paper studies machine translation for very low-resource languages using large language models. Prior lexicon-based prompting hits a ceiling when models must infer sentence meaning from combinations of word senses. The authors propose UD-PrMPT, which uses Universal Dependencies syntactic analyses as prompt components, and test it on five languages: Beja, Chintang, Chukchi, Coptic, and Xibe. The approach yields additive gains on top of lexicon-based prompting, larger with higher-quality manual parses and consistent with predicted parses.
In briefLAION-Audio-XL: A Web-Scale Audio Dataset and Scaling Study for Language-Audio Models
SummaryThis paper collects 1.48 billion audio samples from academic, internet, web, Common Crawl, and YouTube sources, and cleans them with a multi-stage pipeline. The authors also generate short and long captions for samples that lack them. Using this dataset, they study how contrastive audio-language models scale, identifying trade-offs between model size and training data under a fixed compute budget. The resulting CLAP2 models outperform prior audio-text models of similar size on the MAEB and HEAR benchmarks, and the dataset, captions, checkpoints, and code are released.
Gijs Wijngaard, Felix F\"orster, Ross Wightman et al.
In briefFinCutBench: A Continually Updated Benchmark for Measuring LLM Financial Knowledge Cutoffs
SummaryThe paper introduces FinCutBench, a monthly updated benchmark that measures how far a fixed model's financial knowledge extends over calendar time. The current dataset contains 9,000 Chinese and English four-option multiple-choice questions from January 2023 to September 2026, covering Chinese and U.S. markets and four financial information categories. The authors define the FinCut Boundary as the first month when the fitted knowledge profile reaches a lower level, and they report how stable the estimated date is under resampling.
In briefFastGRT: Faster Training for 3D Gaussian Ray Tracing
SummaryThe paper studies how Gaussian-based ray tracing (GRT) runs on GPUs and finds that much of its high training cost comes from how the computation is organized. It introduces FastGRT, which separates ray traversal from blending, reuses forward-pass work in the backward pass, and adds fast gradient accumulation and a hybrid BVH update strategy. Implemented in Vulkan, it is described as the first Vulkan-based GRT implementation supporting both training and rendering. Compared with 3DGRT, training time improves by 1.9× to 4.97× while keeping the original rendering formulation, and the code is open source.
In briefNightSteps: A Night-Time Dataset for Predicting Pedestrian Crossings
SummaryThis paper presents NightSteps, a dataset of night-time dashcam footage from six cities, with pedestrian location data, track IDs, and crossing ground-truth labels. Gemini, ChatGPT, and human assessors were used to predict crossings and compare against the labels. Human input reached 88.17% accuracy and an 80.21% F1-score, while Gemini and ChatGPT reached 60% and 79.13% accuracy respectively. The authors suggest the dataset can be used to evaluate pedestrian crossing intention prediction models at night.
In briefSame Error, Different Recovery: How Feedback Delivery Shapes LLM Agent Behavior
SummaryThis paper studies how tool-using language models respond after a failed call, such as retrying, correcting, or stopping to tell the user. The authors vary only how the error feedback is delivered, including its placement, guidance, retained history, and turn structure, across 13 open models from 3B to 70B parameters. They report that a new user turn retaining the failed exchange avoids a severe stopping failure seen in Llama models. On ToolMaze the delivery layer improves success on explicit errors, while on Loud/Quiet it shows no measurable gain.
In briefQubit-to-Motion: Traceable Robot Action Records in Hybrid Quantum Reinforcement Learning for Bin-Packing
SummaryThis paper presents Qubit-to-Motion, a framework for support-constrained object sequencing and simulated robotic placement, where four quantum readouts are the only learned input to the actor and critic. Each decision is linked through step indices and SHA-256 fingerprints to a route, robot waypoints, joint motion, and MuJoCo contact replay. Across five deterministic 30-object scene variants, all 150 placements satisfied support, containment, and settling criteria.
In briefA new exact method for building optimal classification trees faster
SummaryThis paper presents ISCG-OCT, an incumbent subtree column generation algorithm for optimal classification tree problems. It formulates the problem as a mixed set-packing and set-partitioning model and uses a recursive structure in which fixing a branching decision splits the pricing problem into independent subtree subproblems solved by dynamic programming. Experiments on 27 benchmark datasets report competitive accuracy and substantially shorter solution times on many instances.
In briefContinuous Modeling of Manipulative Participation for Stance Analysis: Online Debate on Türkiye's 2024 Street Animal Law
SummaryThis paper proposes modeling manipulative online participation as a continuous degree rather than a binary label, and integrates this with stance analysis. Using online discussion of Türkiye's 2024 Street Animal Law, the authors collected 7,320,574 tweets from 537,994 users and had experts score 301 selected users on a 10-criterion rubric. The framework estimates both manipulative participation and user stance accurately and generalizes well to an independently annotated user set.
Ulya Bayram, Ebru Belk{\i}s G\"uzelo\u{g}lu, Emel Ku\c{s}ku \"Ozdemir et al.
In briefBeyond the Transcript: Improving Audio Models' Use of Sound
SummaryThis paper argues that audio language model benchmarks can confuse acoustic understanding with language skill, since some tasks can be solved from the transcript alone. The authors separate lexical tasks from tasks requiring attributes such as speaker and emotion, or structure such as stress and intonation. They evaluate 19 open models and Gemini 3.1 Pro on 92 tasks from 6 benchmarks. They train LoRA adapters in early layers of Gemma-4 12B and Qwen2.5-Omni-7B, which raises attribute and structure accuracy by 12.8 to 44.4 points while keeping lexical accuracy within 1 point of the base model.
In briefActions Should Earn Their Credit: Test-Time Adaptation for Frozen LLM Agents
SummaryThis paper introduces EARN, a method that helps frozen LLM agents improve their decisions over repeated attempts without updating model parameters. It lowers credit for actions that repeatedly reach the same score and gives more credit to action segments that reach higher scores. In the final episodes, EARN replays stored effective action sequences when the context matches. Across Jericho, ScienceWorld, and Crafter, it achieved the best average score over the last four episodes in 31 of 33 task-budget settings, including ties.
In briefFrom Sound to Occupancy: Privacy-Aware Audio Sensing for Buildings
SummaryThis paper explores using sound to detect whether a room is occupied, as an alternative to motion sensors that miss still occupants. A two-stage design encodes each second of audio into a short on-device embedding, and a server-side temporal model converts these into an occupancy state. Results show better detection than log-mel features and much less leakage of spoken words, while speaker identity can still be recovered and remains an open privacy issue.
In briefHydroTuring: a physics-check benchmark testing whether hydrologic models conserve mass, energy and momentum
SummaryHydrologic models are usually judged by how well simulated streamflow matches observations, but a close match does not guarantee they obey basic conservation laws. This paper introduces HydroTuring, a benchmark of 37 physics probes that test mass, energy and momentum conservation using pass or fail criteria. Among 32 evaluated models, only the 3 exact reference models pass every scored probe, and process-based and AI-based models pass similar shares of their probes. About 49% of model-probe pairs remain unscored because models omit required variables.
In briefUTOPI: user-guided token pre-compression for efficient egocentric long-video understanding on AR devices
SummaryThis paper introduces UTOPI, a plug-and-play pre-compression module for egocentric long videos on augmented reality platforms. It uses user motion signals to estimate inter-frame overlap and prune redundant frames with a viewpoint-shift-invariant algorithm. It also uses eye-tracking to identify fixations and preferentially retain attended foreground content. The authors report up to 95% compression while preserving performance across multiple benchmarks and model backbones.
In briefLiveAsset: learning user intent from corrections to guide AI agents at runtime
SummaryThis paper proposes LiveAsset, a framework that learns reusable intent assets from users' multi-turn corrections to an AI agent's outputs. It infers user intent by comparing each correction with the agent's original and revised results, and uses the learned requirements both to guide task execution and to verify replies before delivery. The authors evaluate it on several benchmarks against memory, skill, and rule-based baselines.
In briefInfraRSI: Using recursive self-improvement to optimize LLM training infrastructure
SummaryThe paper presents InfraRSI, an agentic framework that turns open-ended human commands into verified optimizations for distributed LLM training. It works in five stages, including locking the editable scope, profiling baselines, co-optimizing launch settings and code patches, and logging results to a cross-hardware experience bank. Across four dense and Mixture-of-Experts LLM workloads, it reports a speedup of up to 2.75× and raises mean MFU from 17.4% to 30.0%.
In briefKimodo-Policy: Learning Motion-Based Action Models for Humanoid Movement and Manipulation
SummaryThe paper introduces Kinematic Action Models (KAMs), which treat whole-body motion as a direct pretraining target and use a pretrained motion generator as the main action backbone. It presents Kimodo-Policy, combining a frozen text-to-motion backbone with lightweight ControlNet-based visual conditioning. The policy achieves the highest overall success on multi-task HumanoidArena (72.14%) and 60% mean success across eight real-world tasks, with 678M parameters (92M trainable).
In briefWhen Should Diffusion Language Models Reconsider Their Commitments?
SummaryMasked diffusion models generate text quickly by decoding tokens in parallel, but independent sampling can lock in inconsistent predictions. This paper proposes PURL, a training framework that jointly learns parallel unmasking and selective revision of earlier tokens within one model. Across logic puzzles, math reasoning, and code generation, PURL improves accuracy at matched unmask budgets while reducing the number of decoding rounds.
Haoyang Zheng, Yidong Ouyang, Blake Jianhang Chen et al.
In briefNot All Branches Are Equal: Adaptive Semantic ID Construction for Generative Recommendation
SummaryGenerative recommendation models represent items as sequences of discrete tokens called semantic IDs. The paper argues that semantic IDs should behave as hierarchical paths, with each token's meaning depending on its prefix. It proposes Adaptive Silhouette Tree (AST), a top-down divisive clustering method that chooses the number of branches at each node by maximizing the silhouette coefficient. Experiments on five Amazon Review datasets show AST significantly outperforms strong baselines.
In briefCausal Attribution for Diabetic Retinopathy Without Actually Dropping Tokens
SummaryThis paper examines whether attention-based attribution maps in a vision transformer offer causal explanations for diabetic retinopathy classification. The authors split attention into two maps, contribution and pull, and test their causal faithfulness on the MAPLES-DR dataset by dropping and inserting tokens. The pull map achieves the highest causal alignment and is far more computationally efficient than occlusion.
In briefSafety Safeguards for AI Agents in Radiology: Comparing Rule-Based and LLM-Based Checks
SummaryThis simulation study examines how often tool-using large language model agents record clinically relevant false or unsupported information in simulated radiology documentation. It compares three conditions: permissions only, deterministic rules, and rules plus a locally run LLM checker, across 48 base scenarios and 91 misleading-information variants. Layered safeguards reduced but did not prevent high-severity errors, and errors after misleading information were attributed to misjudgment rather than missed information.
Semil Eminovic, Robin Schmidt, Bogdan Levita et al.
In briefOstinato: A fast CPU-GPU engine for running large mixture-of-experts language models on local agents
SummaryThis paper presents Ostinato, an inference engine that runs large mixture-of-experts language models split between CPU and GPU memory. It keeps a stable set of frequently used experts on the GPUs during every forward pass, without changing the model's outputs. Tested on real coding and customer-service agent sessions with DeepSeek-V4-Flash on four RTX 4090D GPUs, it cuts the median wait after tool outputs by 22-55% compared with llama.cpp and KTransformers, and finishes sessions 1.4-2.5 times sooner.
In briefDriveFoundry: Co-evolving Data and Model for Faster Autonomous Driving Iteration
SummaryThe paper presents DriveFoundry, a closed-loop framework that improves how efficiently autonomous driving systems iterate under data and compute limits, combining causal data refinement with adaptive inference. Its model, Xiaomi-Drive-o1, handles perception, causation, action, world, and value tasks in a shared vision-language backbone. The authors report benchmark results on PhysicalAI and AlpaSim, and state that adaptive inference matches full reasoning using only 28% of tokens.
In briefProbing which layer and module of transformer models best represents information
SummaryThis paper studies which internal representation of transformer foundation models should be probed. Through more than 144,000 linear probing evaluations on vision and language models, the authors find that the best layer-module pair outperforms the standard final-layer residual stream. They also propose two simple fixed probing strategies that improve over the baseline in most evaluations.
Ambroise Odonnat, Theo Gnassounou, Laetitia Chapel et al.
In briefAttacks that manipulate data in machine-learning-powered cyber-physical systems
SummaryThe paper notes that cyber-physical systems often combine several machine learning models, such as vision and planning, to make decisions, and studies how an attacker could exploit this composition. The authors present a framework for building an attack policy that maps system states to data manipulations, along with a basic threat model and methods for computing an optimal policy. This is a work-in-progress paper intended to spark discussion and collaboration on the security of compositional learning.
In briefA stability study of the AdamW optimizer, showing some equilibria are unstable and training may not converge.
SummaryThis paper analyzes the AdamW optimizer, widely used in deep learning, using tools from dynamical systems. It shows that AdamW's fixed points match the stationary points of a constrained, regularized optimization problem, but not all of them are stable, and stability depends sensitively on curvature, weight decay, and momentum parameters. Even in one dimension, equilibria can be unstable and trajectories can oscillate persistently instead of converging. The authors derive sufficient conditions for local stability of the continuous-time dynamics and propose a convergent modification that keeps AdamW's fixed points.
In briefBoundary Mass in Feature Space: A Label-Budget Diagnostic for Representations
SummaryThis paper proposes effective boundary complexity, a diagnostic for frozen representations and linear probes that estimates how many labels are needed to learn the downstream decision boundary. It measures the mass of samples in a thin strip around the learned boundary, normalized by class balance. Experiments report strong agreement with label demand and backbone rankings.
In briefMusic inpainting using similarity-guided latent diffusion
SummaryThis paper proposes a music inpainting method that combines similarity matching with a diffusion-based generative model to replace large segments of audio recordings. The hybrid approach, called SimDPS, retrieves candidate replacement segments from the wider signal based on contextual similarity and uses them to guide the diffusion process toward coherent reconstructions. The framework is model-agnostic and works with any pre-trained diffusion model in the waveform or latent domain. Objective and subjective evaluations show that SimDPS outperforms RePaint and DPS on acoustic and neural-embedding metrics and improves perceptual plausibility, while remaining robust to poorly matched candidates.
In briefK-PWM: Structured Bilinear World Models for Control Under Partial Observation
SummaryThe paper introduces K-PWM, a probabilistic world model that combines Koopman-bilinear latent dynamics with a nonlinear observation decoder and infers latent states via extended Kalman filtering and smoothing. On partially observed Gymnasium MuJoCo systems with limited training data, it reports lower long-horizon prediction error than the evaluated baselines and much shorter training time than the strongest prediction baseline. On high-dimensional MuJoCo Menagerie robots, it achieves competitive tracking accuracy with low per-step computation times.
Santosh Mohan Rajkumar, Sriram Narayanan, Samuel E. Otto et al.
In briefHydra: A Transferable Multi-Task Learning Model for Temporal Graphs
SummaryThis paper introduces Hydra, an architecture that combines local connectivity features from temporal graph neural networks with a spectral module that captures global connectivity patterns, learning both under one multi-task objective. In multi-task classification, Hydra improves AUC by 8.9% relative to the strongest competitor. In multi-task regression, it is competitive on all three tasks and best on two, with an 8.2% relative MAE gain over the strongest baseline. Training time is reduced by 22% compared with temporal transfer models.
Kiarash Shamsi, Farimah Poursafaei, Tran Gia Bao Ngo et al.
In briefInfCoiL: Jointly Training a Planner and Controller in a Closed Loop for Physics-Based Human-Object Interaction
SummaryThe paper presents InfCoiL, a closed-loop framework that makes physics-based characters perform text- and goal-conditioned human-object interactions. It connects a flow-matching planner and a multi-morphology controller through a shared interface and fine-tunes both with reinforcement learning from the same physical rollouts. A new method, Semantic Flow Policy Optimization, is introduced to optimize the planner online. On held-out OMOMO sequences, InfCoiL outperforms open-loop hierarchical baselines in goal-reaching success and physical tracking accuracy.
In briefDetecting common golf swing faults from video using AI and biomechanical features
SummaryThis paper presents a method that detects named golf swing faults from ordinary video without markers or calibration. It tracks body pose and the clubhead in a body-centered frame and classifies biomechanical quantities a coach would measure using shallow boosted trees. Tested on over-the-top and early-extension faults, it achieved cross-validated ROC-AUCs of 0.975 and 0.911.
In briefBayesian experiment design that uses extra prior knowledge at deployment without retraining
SummaryThis paper introduces IMBUE, a framework for amortized Bayesian experimental design that enables real-time design strategies. It accepts user-specified prior knowledge and observations from earlier experiment instances through a unified in-context interface, without retraining. A shared Transformer processes these inputs with accumulated observations, while a learned reliability filter excludes inconsistent external tokens. Across four benchmarks, the method speeds up early-stage information acquisition when external knowledge is reliable and stays close to the no-external baseline when it is not.
In briefReaching is not finding: a map-anchored visual prompt for the final step of drone navigation
SummaryAerial vision-language navigation asks a drone to find an object described in natural language, usually relative to landmarks like streets and buildings. The authors find that current navigators often reach the target area but stop at the wrong place, and they learn landmarks rather than the described object. They propose MAP-VP, a plug-and-play last-step module that draws the instruction's landmarks onto the drone's view and uses a fine-tuned 4B vision-language model to locate the target relative to them. On CityNav, it improves all six navigators tested without retraining them, roughly doubling the success rate of the two strongest trained navigators.
In briefReviewing Episodes, Refining Lessons: Simple Memory Evolution for Robot Agents
SummaryThis paper introduces SimplEvo, a method that lets a multimodal large language model improve by recording its own execution experience as text memories, without weight updates or privileged information. The model writes a task memory from one demonstration and maintains a working memory refined after each training episode, with both fixed at test time. On EBench, it raises GPT-6-Astra's success rate from 46.7% to 81.7%, and it also reaches 69.63% on the RoboDojo leaderboard.
In briefConvLUT: Convolutional Designs for Ultra-Fast Lookup-Table Inference
SummaryThis paper presents ConvLUT, a family of fully unrolled convolutional architectures that can be mapped to lookup tables (LUTs) for ultra-low-latency inference on FPGAs. It targets the MLPerf Tiny benchmarks and includes a standard convolutional model built on a hierarchical LUT tree and a depthwise separable model. Evaluated on CIFAR-10, Visual Wake Words, and Keyword Spotting, it achieves competitive accuracy with fully unrolled inference latencies below about 50 ns.
In briefAdapting foundation models to data for causal effect estimation
SummaryCausal inference cannot observe counterfactual outcomes, so training goals and inference targets differ. The paper finds that geometric properties of data embeddings, specifically the variogram, signal whether a given causal estimation strategy suits a dataset. The authors propose MoE-Learner, a mixture-of-causal-experts framework that uses the variogram as a counterfactual-free router to assess each expert's validity. The method achieves the best average rank on the IHDP, ACIC, and Lalonde benchmark datasets.
In briefAgentCT: A Vision-Language Agent for Universal CT Imaging
SummaryAgentCT is a vision-language agent that identifies the type of degradation in CT images (low-dose, sparse-view, or metal artifacts) and routes each image to a task-specific restoration model. A LoRA-adapted Qwen3-VL-4B controller predicts degradation type and severity, selects a specialist, and reassesses the restored image. On patient-disjoint Mayo 2020 abdominal CT data using ground-truth routing, AgentCT achieved an overall PSNR/SSIM of 46.18 dB/0.984, outperforming AMIR and TAMP and reaching the highest PSNR in all nine settings.
In briefWhether Merge Forms a Hopf Algebra: Coproducts, Traces, and Algebraic Minimalism
SummaryThis paper examines the claim by Marcolli, Chomsky, and Berwick (2025) that the syntactic operation Merge can be formalized as a Connes-Kreimer-type Hopf algebra. The authors argue that neither coproduct used in that work satisfies the required properties, so the algebra is not a Hopf algebra and the Birkhoff factorization follows only from weaker conditions. They conclude that the current formulation cannot both keep the Hopf structure and remain compatible with Heim-Kratzer formal semantics, and they outline three directions for algebraic Minimalism.
In briefForeAct: Geometry-Guided Robot Action Learning from World Representations
SummaryForeAct repurposes a pretrained video world model as a single-pass representation backbone for low-latency robotic control, avoiding future-frame synthesis and iterative denoising. It uses learnable tokens supervised by next-frame relative depth to make the shared video representation future-aware, with a geometry readout used only during training and removed at deployment. The paper reports 96.6% average success on LIBERO, 63.5% on SimplerEnv-Bridge, and 74.4% across 90 real-world rollouts.
In briefJoMoHa: Jointly Optimizing Language Agent Models and Harnesses via Reinforcement Learning
SummaryThis paper introduces JoMoHa, a framework that uses reinforcement learning to jointly train an executable harness generation policy and a task execution policy for language agents. The authors report that naive alternating training of the model and harness leads to growing incompatibility and performance degradation. On ALFWorld, WebShop, and DBBench, JoMoHa achieves higher success rates than the evaluated weight-only and harness-only baselines, averaging 70.3% and reaching 97.0% on ALFWorld.
In briefTID System for Speech-COSER 2026 ASR Track: Telefónica's Speech Recognition Entry
SummaryThis paper describes the system Telefónica Innovación Digital submitted to the Speech-COSER Challenge 2026 ASR track, which uses spontaneous speech from elderly speakers in rural Spain. The system uses CTC-based pseudo-forced alignment to score audio-text match quality, guiding a two-stage LoRA curriculum, followed by post-hoc magnitude pruning of LoRA matrices. An ensemble of seven systems combined with ROVER achieves 36.25% unnormalized and 22.17% normalized word error rate.
In briefRegret analysis for bandits with reward-dependent delays and no optimal-arm feedback
SummaryThis paper studies stochastic bandits where reward reports arrive with delays that depend on the reward and omit the originating round. It proposes Balanced Historical Optimism (BHO), whose delay term is controlled by suboptimal-arm quantiles without requiring optimal-arm reports. The authors provide lower bounds and a worst-case regret bound for a family with growing arm count, and analyze a pending-triggered balancing mechanism and its exploration coefficient.
In briefCapacity-aware discharge policies to reduce biased hospital readmission prediction
SummaryThis paper studies readmission risk predictors used to guide hospital discharge decisions, which can become biased against vulnerable patients due to censoring when they are repeatedly retrained on discharge-generated data. The authors propose capacity-aware discharge-up-to policies that diversify patients' lengths of stay, allowing consistent risk estimation without explicit fairness constraints. In synthetic experiments and simulations based on partner-hospital data, these policies reduced learning error relative to threshold-on-risk policies and narrowed readmission cost disparities across demographic groups.
In briefCallCleaner: A lightweight tool that removes background noise from animal sound recordings
SummaryField recordings of animal calls often contain background noise from weather, vehicles, and other animals. The authors distill the general-purpose source separation model SAM-Audio into CallCleaner, a 19M-parameter denoiser that needs no detector and runs about 90 times faster than real time on a laptop CPU. CallCleaner leaves 11 dB less noise around calls than Biodenoising, a leading neural bioacoustic denoiser with public weights, while preserving call fidelity better than other fast systems.
In briefTHEATRON: A Benchmark for Testing Persona Persistence in Long-Term AI Decisions
SummaryThis paper introduces THEATRON, a benchmark that uses a financial replay setting to test whether language models keep a fixed behavioral-finance persona over multi-month sequential decisions. Models receive shared daily market information and account feedback while making trading actions under zero-shot, one-shot, and chain-of-thought persona prompts. The results show that personas produce systematic behavioral shifts, but these shifts do not reliably form coherent persona-specific decision policies.
In briefMicroSim: Graph neural networks that detect similar compiled code
SummaryThis paper presents MicroSim, a system that detects similar binary code by representing each function as a graph built from IDA Pro Hex-Rays microcode and comparing these graphs with a contrastive Siamese graph neural network. It is evaluated on the TREX benchmark of 849,072 function-level microcode graphs across optimization levels -O0 to -O3, using a split where no function key appears in both training and test sets. The authors report a ROC-AUC of 0.9972 and MCC of 0.9762, outperforming retrained Gemini, SAFE, jTrans, and a bag-of-words baseline.
Moussa Moussaoui, Tarik Houichime, Abdelghani Souhar et al.
In briefOne-shot federated hierarchical clustering for analyzing distributed non-IID biomedical data
SummaryThis paper proposes a one-shot hierarchical federated clustering framework for analyzing biomedical data held by multiple institutions. Each client explores its local data distribution and sends only prototype-level information one way to a central server. The server then fuses these local clusterlets using a multi-granular competitive learning mechanism to recover global distribution structure. The authors report experiments on public datasets showing a favorable balance among clustering quality, privacy preservation, and computational efficiency.
In briefOne Small Embedding Model for Tajik, Kazakh, Uzbek and Kyrgyz
SummaryThe paper reports that the released EmbeddingGemma-300M ranks the correct Belebele passage first for only 51.7% to 60.8% of questions in Tajik, Kazakh, Uzbek and Kyrgyz. The authors fine-tune their Tajik model on additional Kazakh, Uzbek and Kyrgyz question pairs generated by their own language model, plus English and Russian replay data. The resulting single model reaches 75.2% to 77.7% recall@1 on Belebele in these four languages, but still trails the larger 560M-parameter XLM-RoBERTa-based embedders on Kazakh, Uzbek and Kyrgyz.
Tohir Saidzoda, Nekruz Maxudov, Bedil Karimov et al.
In briefGOPR-CAD: Step-by-step reasoning over CAD operation graphs for generating 3D models from text
SummaryThis paper presents GOPR-CAD, a pipeline that generates CAD models from text descriptions. A language model acts as an agent that progressively builds an executable graph of encapsulated CAD operations, which a deterministic compiler then turns into the final model, requiring no additional training. The authors also introduce MHCAD, a benchmark of 300 standard mechanical part requirements derived from Machinery's Handbook, enabling reference-free STEP/B-rep structural evaluation. GOPR-CAD reaches 82.7% Structural Yield on MHCAD and 87.2% on an external CADPrompt subset.
In briefRealEmbedBench: A Benchmark Testing LLMs on Real-World Application-Layer Embedded Software Tasks
SummaryThis paper introduces RealEmbedBench, a benchmark of 5,142 samples covering code generation and completion, state machine modeling and implementation, and cross-OS code translation. Six representative LLMs were evaluated, showing substantial differences across tasks and models. Local code context strongly affects code generation, and translation from Harmony to RT-Thread is especially difficult, with an average pass rate of 0.1667.
In briefWhen Normality Shifts: Multi-layer WavLM for Open-World Audio Deepfake Detection
SummaryThis paper studies an attack-free approach to audio deepfake detection, training a normality model only on bona fide speech and scoring audio by its deviation from authentic representations. The proposed MLN-WavLM applies shrinkage-Mahalanobis modeling to pooled hidden states of a frozen WavLM Base+ model. It reaches an 8.45% Equal Error Rate on ASVspoof 2019 LA, close to a fully supervised baseline at 7.86%, but rises to 22.12% under the ASVspoof 2021 DF domain shift. Threshold recalibration under that shift causes low-FPR recall to collapse entirely.
Daniel Niekurzak, Daniel Kozuch, Weronika Folwarska et al.
In briefOff-policy RL with online inference tuning for the CoRL 2026 RoboEval challenge
SummaryThis paper describes the author's solution to the CoRL 2026 RoboEval challenge, which has eight bimanual tasks scored on task stage completion, safety, efficiency, and arm coordination compared against human demonstrations. The method is off-policy reinforcement learning on a π0.5-based policy that also serves as its own value function, with each rollout doubling as a Thompson-sampling trial over inference parameters. The submitted policy reaches 0.990 RDS on held-out seeds and 0.987 on the official seeds.
In briefAligning unpaired model representations using a shared pullback distance on the data manifold
SummaryThe paper introduces Neural Gromov–Monge Alignment (NGMA), a method that aligns two models' representations without paired data or labels. It uses the pullback of the data metric as a distance shared by any two models of the same data-generating process, under a precise manifold hypothesis. The alignment map is identifiable exactly when the data distribution breaks every symmetry of the data manifold, which the authors show holds generically. NGMA outperforms unpaired and label-informed baselines on image and single-cell models.
Stas Syrota, Albert Lopez I Serrano, Johanne Franck et al.
In briefMOMENT-DP: Momentum-Aware Evaluation of Correlated Noise for Private Training
SummaryThe paper introduces MOMENT-DP, a framework for evaluating temporally correlated noise used in differentially private optimization. It models how privacy noise propagates through momentum-based optimizers, accounting for curvature, learning-rate schedules, momentum, and noise covariance. The authors show that changing only the momentum can reverse the ranking of privacy noise mechanisms, indicating that optimizer state is an important factor in evaluating correlated privacy noise.
In briefSemantically aligned KV cache compression for efficient agentic video generation
SummaryThis paper introduces SemanticKV, a training-free method that compresses the key-value cache in agentic autoregressive video generation. It uses semantic constraints from the planner, identified through text cross-attention, to score cache entries and keep the high-scoring ones under a memory budget. On Causal Forcing at a 0.7 compression ratio, it reaches a VBench Total Score of 84.10, above Full KV's 83.99, with a 1.22x speedup and KV-cache memory reduced from 6.04 GB to 1.81 GB.
In briefHow much train-inference mismatch on-policy distillation can handle
SummaryThis paper studies how much mismatch between the training and rollout inference engines on-policy distillation (OPD) can tolerate. The authors measure all mismatch sources on one scale, testing quantization across three student-teacher pairs and weight staleness and injected noise on one pair. OPD tolerates substantial mismatch without meaningful quality loss but degrades beyond a threshold. They also propose a lightweight admission check against a calibrated budget, along with model-specific FP8 recipes and asynchronous rollouts that improve efficiency.
In briefLearning How to Rank Constraint Propagators for Faster Solving
SummaryThis paper proposes learning instance-dependent priorities for propagators, the components that narrow down variable domains in constraint programming solvers, instead of using fixed manual heuristics. A self-attention model scores propagators and is trained on pairwise preferences derived from solver runtimes under randomized priority assignments. On 698 difficult XCSP3 instances, compared with Choco's default priorities, the learned priorities raise solved instances from 634 to 654 and cut median runtime by 27.93%.
In briefSWHForest: A structural layer for filtering large code corpora without redistributing source code
SummaryThe paper presents SWHForest, an open-source framework and public dataset that adds a structural filtering layer to code corpora published only as file pointers. It parses each file once into an abstract syntax tree, discards the source text, and keeps only structural summaries so that matching files can be selected before downloading. On the ten largest languages of Stack-Edu (52.1M unique files, 341M definitions), queries answer in milliseconds and cut downloads by 46-72%.
Federico Pennino, Andrea Gurioli, Maurizio Gabbrielli et al.
In briefBenchmarking general AI agents for understanding surgical scenes
SummaryThis paper evaluates general-purpose multimodal language models on surgical scene understanding, which is hampered by scarce data and inconsistent evaluation methods. It introduces SurgKnowBench, a diagnostic benchmark built from harmonized public surgical datasets, showing models are weak in relational, procedural, and workflow understanding. It also proposes SurgAgentGym, an environment where agents use bounded knowledge search and visual grounding at inference time without surgical training. Results indicate agentic inference helps mainly perception-heavy tasks, while fine-grained localization and expert-style assessment remain bottlenecks.
