In briefGeneralized engression models: a unified method for modeling the joint conditional distribution of mixed-type multivariate outcomes
SummaryThe paper proposes generalized engression models, a unified nonparametric distributional regression framework for estimating the conditional distribution of multivariate outcomes that may be continuous, binary, categorical, ordinal, or rankings, and are mutually dependent. It extends engression, a scoring-rule-based deep generative model, with data-type-specific link functions and a stochastic perturbation that smooths the loss, enabling gradient-based training with discontinuous links. In simulations and two applications, a 242-species community ecology benchmark and a 17-dimensional mixed-type health outcome, the authors report results matching type-specific models on marginal scores, improving on the joint distribution, and matching or exceeding specialized joint species distribution models. Software is available in Python.
WhyExtends engression into a unified distributional regression framework for mixed outcome types; moderate novelty. Validated on 242 species and 17-dimensional mixed data against specialized models, with
In briefBeyond Linear Concepts: Finding and Aligning Non-Linear Concept Shapes in Language Models
SummaryThis paper adapts a non-linear concept discovery method from computer vision to analyze token-level activations inside large language models, modeling concepts as low-dimensional manifolds. The authors introduce a concept-based alignment score to compare these manifolds across layers and models without explicit feature matching. They report block structures in intermediate and late layers, varying English-Mandarin concept sharing across model families, and the highest alignment between adjacent Tulu-3 training stages.
Tido Specht, Elias Benedict Krey, Nils Neukirch et al.
In briefMECHVAR: Variance-Guided Selection of Experiments to Tell Apart Mechanisms in Machine Learning
SummaryThis paper studies how to identify which mechanism truly explains a benchmark improvement, given a finite library of candidate mechanisms and a limited experiment budget. The authors propose MECHVAR, which chooses the next probe by maximizing the posterior-weighted variance of predicted responses, a score that is cheap to compute and auditable. In a held-out Digits loop, its normalized mechanism-identification AUC is 0.8975, compared with 0.7825 for a score-greedy policy and 0.9092 for EIG. Full-library scoring is much faster than six-node quadrature EIG in the recorded environment.
In briefDebias Anything: Jointly improving fairness and diversity in diffusion models without supervision
SummaryDiffusion models produce high-quality images but can reproduce and amplify demographic imbalances in their training data. This paper proposes a method that addresses fairness and diversity together, applicable to any diffusion model and any sensitive attribute. An adapter links the frozen model to a pretrained vision-language embedding space, enabling guidance without sensitive-attribute annotations. Experiments show improved quality and diversity scores at comparable fairness levels.
In briefAI-assisted tool for more consistent and faster counting of mitotic cells in tumour slides
SummaryThe paper presents MitPro, an AI tool that directs pathologists to regions with the highest predicted mitotic activity and highlights mitotic figures, while pathologists retain control over region selection and the final count. In a retrospective paired reader study of 385 whole-slide images from 7 tumour types and 13 pathologists, AI assistance raised the intraclass correlation coefficient from 0.589 to 0.949. Median assessment time fell from 286.4 to 127.8 seconds on average.
WhyCombining AI region guidance with mitotic figure highlighting is moderately novel; a 385-slide, 13-pathologist paired study showing ICC rising from 0.589 to 0.949 gives strong evidence and clinical im
Simon Graham, Mostafa Jahanifar, Quoc Dang Vu et al.
In briefPhaseAT: Fourier Phase Adversarial Training for Generalizing Medical Image Models
SummaryThe paper proposes PhaseAT, a phase-aware adversarial training framework that helps medical image models generalize across scanners, sites, and protocols. It iteratively perturbs the Fourier phase with a bounded budget while keeping amplitude unchanged, applying changes only to the YCbCr luminance channel and focusing them with a phase-saliency mask. The authors report over 20% improvement in single-source domain generalization on two medical datasets, outperforming several state-of-the-art methods.
Ahmed Sharshar, Asif Hanif, Naveen Kumar Kummari et al.
In briefLineupRL: Training Time Series Captioners with Caption-to-Series Matching Rewards
SummaryLineupRL is a reinforcement learning method for generating natural-language captions of time series. Its reward comes from a frozen large language model that reads a generated caption and several candidate raw time series, then must pick the one the caption describes. Across two captioning benchmarks and forecasting and reconstruction tasks where only the caption is given, it outperforms supervised fine-tuning and other RL baselines on every metric.
In briefSkillEvoLean: Mutation-Enhanced Skill Evolution for Lean Theorem Provers
SummaryThis paper proposes SkillEvoLean, a framework that improves large language model agents for formal theorem proving in Lean without updating their parameters. It jointly evolves a high-level solving policy and its reference knowledge, using progressive updates from successful and failed attempts and mutation that samples mathematical concepts when no complete proof is found. The authors report higher proof success rates than baselines on MiniF2F, PutnamBench, IMO 2025, and USAMO 2026 under matched conditions.
In briefAdding muscle (EMG) and visual cues to VLA robot models beyond language
SummaryVision-Language-Action models mainly rely on language to describe tasks. This paper adds 8-channel EMG signals (EC-VLA) or visual segmentation annotations (VA-VLA) as extra task conditioning. In a cube-selection task with three participants, EC-VLA matched a language baseline in uncluttered scenes and clearly outperformed it in cluttered, out-of-distribution scenes. VA-VLA gave modest gains in-distribution and substantial gains in cluttered out-of-distribution trials.
Edward W. Staley, Connor O. Pyles, Rahul Hingorani et al.
In briefiADD: Better Alignment and Diversity in Diffusion Policy Optimization
SummaryThis paper studies reinforcement-learning post-training of diffusion models, such as DDPO, where reward optimization often reduces diversity and quality. The authors argue theoretically that updating only the later timesteps may harm diversity, and propose an incremental Feynman-Kac training method (iADD). Across three tasks compared with related approaches, they report gains in both alignment and diversity.
Ashok Prasad Neupane, Saugat Adhikari, Pramish Paudel et al.
In briefNot All Experience Belongs in the Weights: Component Routing for Self-Improving GUI Agents
SummaryThis paper studies how self-improving GUI agents should reuse their own trajectories, either by fine-tuning the weights or by retrieving them into the prompt. It splits experience into locators, procedures, state facts and lessons, and routes each component to context or weights. A rule based on recurrence and state-conditionality, measured before training, is reported to predict the destination correctly in 24 of 24 held-out cells. The authors say this routing outperforms whole-trajectory baselines and the better single destination for each backbone.
WhyProposes component-level routing of experience with a rule fitted on pre-training properties, validated across backbones, environments, and seeds.
In briefEstimating multiple neural timescales that switch with behavior
SummaryNeural activity often shows several timescales that change with behavior, and identifying them from recordings matters. This paper introduces MTS-SLDS, a framework that recovers regime-specific latent timescales from continuous or spiking neural observations. It combines multi-lag moment initialization with regime-conditioned Laplace-EM inference, and extracts timescales from the eigenvalues of learned transition matrices. In synthetic and neural experiments with Gaussian and Poisson spike observations, it accurately recovers timescales and switching structure.
In briefVETO: A Plug-in That Cuts Visual Tokens for Faster Long-Video Understanding in Vision-Language Models
SummaryThe paper introduces VETO, an optional plug-in that reduces visual tokens when vision-language models process long videos. It compresses redundancy in two steps: merging similar tokens within each frame, then merging temporally redundant frames. The authors report up to 45% faster inference on LLaVA-OneVision-7B and 55.7% accuracy under a 10% token budget. They also state that zero-shot accuracy was preserved or improved across LLaVA-OneVision, InternVL-2.5, and LongVA.
Gueter Josmy Faure, Hao Ping Wang, Min-Hung Chen et al.
In briefModel-free Q-learning for reachability in MEC-free MDPs (Quasar)
SummaryThis paper presents Quasar, a model-free reinforcement learning algorithm for reachability in Markov Decision Processes that have no non-terminal maximal end components (MECs). It uses classical Q-learning with temporal-difference updates and converges to an optimal policy without learning transition probabilities. On the Quantitative Verification Benchmark Set, it needs orders of magnitude fewer samples than the previous model-based state of the art, and its memory use drops from O(|S|^2|A|) to O(|S||A|).
WhyFirst model-free asymptotically convergent reachability learner, reducing memory to O(|S||A|); orders-of-magnitude sample savings on QVBS benchmarks.
In briefRealCompanion: A Benchmark for Understanding People from Months of Real AI Companion Conversations
SummaryThe paper releases RealCompanion, a dataset of ten real relationships between people and an AI companion, with 27,218 messages over up to 120 days, plus profiles, personas, and chat and question test items with evidence-linked labels. The authors find that only 3.4% of messages depend on earlier content, and that when they do, the source is often far back. They also report that tested AI systems barely beat chance at detecting when the past matters, and that agent systems tend to infer more about a person than was actually shared.
WhyIntroduces a benchmark from ten real long-term human-AI conversations, revealing rare callbacks and poor past-relevance detection; well-evidenced, with impact mainly for research.
In briefGIFTBench: A Benchmark for Testing and Improving Image Forgery Localization Generalization
SummaryThe paper introduces GIFTBench, a benchmark of 115,013 manipulated images with pixel-level labels across manipulation source, semantic target, editing operation, and composition complexity. Its diagnostic studies reveal asymmetric cross-source transfer and recall-dominated failures. The authors also propose ForenScope, a detection and localization framework that improves cross-dataset localization while keeping image-level detection capability.
In briefCODesign: Keeping Sequences and Structures Consistent in All-Atom Protein Binder Design
SummaryThe paper presents CODesign, a framework for de novo protein design that generates sequences and structures together while keeping them consistent. The authors build about 105,000 consistency-distilled dimers to improve data consistency and use a multimodal joint flow model with a consistency-aware resampling strategy to refine sequences and side chains. Reported experiments show state-of-the-art in silico success rates for protein and ligand binder design, and an ablation indicates the distilled dataset raises performance by 70.9%.
In briefCONTRA: Finding and Filtering Clarifying Questions That Change Code Behavior in LLM Code Generation
SummaryThis paper introduces CONTRA, a training-free method that helps coding agents ask selective clarifying questions when requirements are underspecified. It generates candidate questions, filters out irrelevant or already-resolved ones, and keeps those whose two plausible answers produce stable behavioral differences in generated programs. On the ClarifyCodeBench benchmark, CONTRA achieves the highest F1 across four coding agents, outperforming Claude Code and OpenHands. The authors also provide a Claude Code plugin implementation.
In briefA Matryoshka Hierarchical RAG for Efficient Multi-Hop Question Answering
SummaryThis paper presents MatRAG, a hierarchical retrieval-augmented generation framework that combines RAG with Matryoshka Representation Learning to answer multi-hop questions more efficiently. It organizes documents into a directed acyclic graph of clusters, with each level indexed by a lower Matryoshka dimension, and uses top-down traversal with entity-driven hop control and re-ranking. The authors report that on three standard benchmarks against seven baselines, MatRAG outperforms its strongest competitors in retrieval quality. They also state that it lowers indexing costs by avoiding knowledge graph construction and LLM-based summarization, and reduces query-time costs through dimension-aware similarity.
Gianluca Bonifazi, Christopher Buratti, Michele Marchetti et al.
In briefVideoEvolve: Automatically Evolving Agent Harnesses for Locating Events in Videos
SummaryThis paper addresses video temporal grounding, which locates events in videos from natural-language queries. It introduces VideoEvolve, a framework that automatically evolves the agent harness built around frozen video-language models, using a Cloze-Structured Harness Representation and Branch-Guided Harness Evolution. Experiments report improved grounding performance across multiple benchmarks. Component analyses find that instruction refinement is a consistent source of gains.
In briefPhysics-refined forecasting for hydrologic graphs with open boundaries
SummaryThis paper addresses instability in spatiotemporal forecasting on hydrologic graphs with open boundaries, where external inputs such as upstream inflows or tidal signals are unavailable at prediction time. The authors propose learning ghost node proxies to approximate these missing boundary forcings, along with two physics refiners that limit error accumulation during autoregressive rollout. Experiments on two real-world hydrologic graphs show higher accuracy and long-horizon stability than learning-based and physics-informed baselines.
In briefTopK-Guided: A Training-Free Way to Speed Up LLM Inference via Adaptive Activation Sparsity
SummaryThis paper presents TopK-Guided, a training-free method that speeds up large language model inference by zeroing out unimportant activations. It combines bounded token-level sparsity adaptation with sensitivity-aware budget allocation across transformer blocks. On Llama-2 and Llama-3 models, it improves perplexity and downstream accuracy over TEAL and WINA at essentially the same projection compute as WINA, with the largest gains at high sparsity. Ablations indicate the two components provide complementary improvements.
In briefOneStreamer: A Model Unifying Perception, Memory, and Timely Response in Streaming Video
SummaryOneStreamer is a streaming video language model that records evidence from video as it arrives and decides when to respond once enough evidence is available. Its Proactive Hierarchical Caption Memory stores time-grounded captions and summaries of completed events, while Proactive State Transition Learning reduces the dominance of repeated waiting states. The authors also release OneStreamer-1M, a dataset of over one million streaming interaction records, and report that their 4B model achieves the best results across eight streaming video benchmarks.
In briefPhysDEM: Physics-guided diffusion model that generates physical fields from sparse measurements
SummaryPhysDEM is a diffusion framework that combines governing equations with spatially sparse measurements to generate multiple plausible spatiotemporal physical fields. It reweights a measurement-conditioned Gaussian reference by PDE residual energy and reduces denoising to supervised learning of an energy-induced mean correction. Experiments on synthetic PDE systems and real-world-informed applications report coherent field recovery and efficient sampling under the tested noise levels.
In briefGenCOPE: Synthetic-to-real category-level object pose estimation for robotic picking
SummaryThis paper proposes GenCOPE, a method for estimating the 3D pose of objects within a category using a model trained only on rendered synthetic data, to be applied directly in the real world. It uses 2D and 3D semantic consistency constraints to learn features that are less sensitive to domain differences, and an end-to-end 2D-3D cross-consistency pose regression framework that relies only on global features. The authors report experiments on the REAL275 and Wild6D benchmarks and real robotic manipulation scenes, showing strong synthetic-to-real generalization.
WhyProposes 2D and 3D semantic consistency constraints with 2D-3D cross consistency learning, validated on REAL275, Wild6D, and real robot scenes for Syn2Real generalization.
In briefCog-VADU: A Training-Free Reasoning Framework for Video Anomaly Detection and Understanding
SummaryThis paper presents Cog-VADU, a fully training-free framework that treats video anomaly detection as a sequential cognitive reasoning task. Its CoADTP method unrolls a vision-language model into a recurrent reasoning chain across video segments, and a cross-modal re-ranking stage aligns textual rationales with visual embeddings. The authors report competitive zero-shot performance on multiple public benchmarks and consistent gains across different models.
In briefEvidence-Gated Research: Statistically Controlled Model Adoption in Adaptive Search
SummaryThe paper argues that in adaptive model search, adopting a new model makes it the reference for later candidates, so a statistically unsupported replacement can distort hypotheses not yet proposed. It introduces Evidence-Gated Research (EGR), a statistical layer that freezes each challenger, builds anytime-valid evidence across environments, and feeds a persistent e-value to an online controller to control false discovery rate. In a 5,000-trajectory benchmark, the audited EGR variants showed no persistent false-adoption path, and in matched replay EGR used 56.1% less decision evidence at the representative threshold.
In briefFFBL-Coop: Cooperative 3D Multi-Object Tracking with Decoupled Association
SummaryThis paper presents FFBL-Coop, a framework for tracking 3D objects across multiple cooperating agents by fusing observations first and assigning identities afterward. It separates instance admission from identity management using three components: Confidence-ranked Slot Admission, Unified Representation Aggregation, and Cooperative-Priority Identity Anchoring, plus a shared codebook to reduce transmitted data. The authors report AMOTA/AP of 0.611/0.548 on V2X-Seq and 0.688/0.653 on Griffin-25M, and state that code will be released.
In briefDesigning dashboards for uncertain interpretation of topological learning analytics
SummaryThe paper presents TopoLA, a dashboard that applies Zigzag Persistent Homology to learning management system data to study temporal dynamics. It proposes three early design principles for supporting interpretation of analytics from methods without established norms: separating measurement from interpretation, graduated disclosure from metrics to reflective prompts, and acknowledging methodological uncertainty. The system uses a modular three-stage pipeline that can be extended to other analytical methods.
In briefComparing End-to-End and Modular Designs for Self-Driving Cars
SummaryThis paper compares two design approaches for autonomous driving: end-to-end learning, which maps raw sensor inputs directly to driving controls, and modular architectures, which split the system into perception, planning, and control subsystems. It evaluates both on interpretability, scalability, robustness, and real-world applicability. The authors note that hybrid approaches combining both paradigms are a promising direction for future research.
In brief3DROID: A renderable 3D Gaussian dataset with per-scene reliability scores for robot manipulation
SummaryRobot manipulation models usually reason from 2D views, even though they act in 3D space. This paper studies how camera extrinsic reliability and pose conditioning affect feed-forward 3D Gaussian Splatting, and proposes a calibration-aware pipeline that anchors reconstructed scenes to the robot's metric workspace. The authors find that pose conditioning improves novel-view fidelity, but its geometric benefit depends on how reliable the injected extrinsics are. They release a renderable, metric-pose-anchored dataset that includes scene-level reliability information.
In briefWorld Motion Models: flexible modeling of 3D object movement sequences
SummaryThis paper proposes World Motion Models (WMMs), which represent dynamic 3D scenes using sparse SE(3) pose trajectories covering objects, human bodies, camera motion, and robot states and actions. The joint distribution of these entities is modeled as a sequence using flow matching with per-token noise levels and a context token mechanism. Different tasks, such as future prediction, motion infilling, and policy learning, are handled by applying different masks to the same network, and experiments on six 3D vision and robotics applications are reported.
In briefATI-VLA: Action-focused predictive robot vision-language-action model (paper)
SummaryThis paper proposes ATI-VLA, a framework for robotic manipulation that predicts future observations to guide action. The authors argue that existing predictive vision-language-action models underperform direct action models due to modality misalignment and conflicting joint optimization. The method first aligns observation and action representations in a shared discrete codebook, then injects predictive latents into action decoding through a lightweight adaptive side-path. The authors report state-of-the-art results in simulation and real-world tasks, with faster convergence.
In briefFixed-point neural samplers on discrete spaces: a data-free method for sampling from unnormalized discrete distributions
SummaryThis paper introduces the Discrete Gibbs Iterative Neural Sampler, a fixed-point neural sampler for drawing samples from discrete, unnormalized distributions without data. It builds on masked diffusion and extends to transport between pairs of distributions. The authors report that the method scales to high-dimensional systems, supports amortized sampling across conditions, and accurately estimates alloy phase diagrams.
In briefParticipation-Sensitive Convergence and the “Fragment First, Converge Later” Pattern in Online Learning
SummaryThis paper studies how asynchronous online learning communities change structurally, using the number of disconnected behavioral clusters, β₀, from Zigzag Persistent Homology. Across 22 OULAD courses with over 22,000 learners and 857 week-pairs, the authors test whether apparent β₀ convergence reflects genuine behavioral alignment or learner dropout. They find β₀ changes strongly co-vary with active learner counts, identifying it as a participation-sensitive indicator. Assessment deadlines were followed by fragmentation and often a fragmentation-convergence cycle.
In briefVerifier-guided reinforcement learning that teaches models to write checkable math code
SummaryThis paper proposes FSG-RL, which links subproblem graphs with Python implementations and uses multi-verifier feedback for mathematical reasoning. The policy is first trained with supervised fine-tuning to generate code from function graphs, then optimized with GRPO using answer-gated rewards and span-level credit assignment. On a benchmark built from GSM8K, MathQA, MATH and Omni-MATH, GRPO raises final-answer accuracy from 43.25% to 67.50% and full solution success from 32.25% to 52.25% over SFT.
In briefUsing skip connections to remove false minima in planar feature learning
SummaryThis paper studies the Gaussian population loss of shallow, bias-free ReLU networks in a teacher-student setting to understand loss landscapes. It shows that adding a learned linear skip connection removes all spurious local minima with non-negative student output weights when the student is at least as wide as the teacher. Without the skip, spurious minima can persist even under arbitrary overparameterization.
Jakob Paul Zimmermann, Andrei Balakin, Moritz Grillo et al.
In briefHow variation in call timing affects deep-learning classification of bat species
SummaryThis study examines whether the time gaps between consecutive bat calls help identify species, and whether transformer models use this timing information better than convolutional networks. Researchers compared European bat recordings with original call timing against versions where call onsets were normalized to 50-ms intervals. Effects differed by model, and overall support for the hypotheses was limited. Models trained on normalized recordings performed worse when tested on natural recordings.
Welmoed R. Eversteijn, Burooj Ghani, A. Leonie Baier et al.
In briefRelICL: Training-free relational learning with tabular foundation models
SummaryThis paper proposes RelICL, a method for learning from relational data without additional training. It propagates and fuses information step by step through the schema graph, using the same tabular foundation model that is used for prediction. The authors identify two problems with deep feature synthesis (DFS): feature explosion as schemas grow more complex, and column-wise aggregates that ignore feature interactions. In experiments on RelBench tasks, RelICL performed on par with the strongest DFS-based approach.
In briefReducing memorization in diffusion models by reinforcing text conditioning
SummaryText-to-image diffusion models can memorize and closely reproduce training images. This paper proposes a training-free method that uses Gaussian smoothing to redistribute cross-attention, then strengthens content-token contributions and weakens padding-token contributions. On Stable Diffusion v1.4 and v2.0, all tested smoothing widths lie on the empirical Pareto frontiers for training-image similarity versus prompt alignment and image preference. A configuration chosen on Stable Diffusion also reduces template reproduction in DeepFloyd IF.
In briefDetecting and locating anomalies in the pantograph-catenary system
SummaryThis paper studies health monitoring of the pantograph-catenary system, which links trains to overhead power lines. Existing industrial video-based methods track contact wire height and stagger but ignore the train route's geographic location. The authors propose a framework that aligns these measurements with nominal GPS coordinates of the reference route and performs collective anomaly detection. They evaluate it on real industrial data from several train journeys on Italian railway routes.
In briefvFedProtoQNAS: Prototype-guided personalized quantum architecture search for federated learning
SummaryThis paper addresses the problem that devices in quantum federated learning have different capabilities, so one shared quantum neural network architecture does not suit every client. The proposed vFedProtoQNAS method lets each client search for and train its own network without sharing model parameters, collaborating instead by sharing class-wise prototypes. Experiments reported in the paper show a 3.70% accuracy improvement over FedAvg and better class-consistent representation alignment.
Seok Bin Son, Samuel Yen-Chi Chen, Soohyun Park et al.
In briefArchitecture Without an Architect? Global AI Governance in a Divided World
SummaryThis review essay examines Matthijs Maas's book on global AI governance, which frames the field through sociotechnical change, governance disruption, and regime complexity. It argues that institutional design cannot be separated from the distribution of power. The author notes that states and a small number of technology companies hold different forms of power over frontier AI decisions, shaping an emerging but fragmented governance architecture.
In briefComparing Two Attention Designs for Learning Nonlinear Tasks From Examples Alone
SummaryThis paper studies in-context learning, where a pretrained model infers a task from examples without updating its parameters, focusing on nonlinear single-index tasks. It compares two one-layer attention architectures: a kernel learner that applies a fixed nonlinear feature map before linear attention, and a feature learner that applies attention first and then a learned nonlinear readout. The authors derive memorization and generalization error predictions using the replica method, and these closely match numerical experiments.
In briefCoEvolve: Locating Described Objects by Building Then Editing Visual States
SummaryCoEvolve is a framework that splits visual grounding, locating an object described in language with a bounding box, into explicit state construction and state editing. Its Region-Evolution Reinforcement module builds a step-by-step semantic-to-spatial trajectory of candidate regions, and its Bidirectional Denoising Refiner adjusts the coordinates while keeping the reasoning text fixed. The authors report that a 9B-parameter backbone rivals models up to 241B parameters in accuracy, and that a single refinement pass under controlled corruption raises mean box overlap by over 27 percentage points.
In briefMEGA: Extracting Object-Level Meshes from 3D Gaussian Splatting via Spatial Visual Distillation
SummaryThis paper presents MEGA, a framework that extracts watertight, object-level meshes from complex 3D Gaussian Splatting (3DGS) scenes using a segment-then-mesh approach. Its core, Spatial Visual Distillation, treats the 3DGS model as a teacher that renders segmented objects from diverse camera poses, and uses these views to train a mesh reconstruction model with photometric supervision. The authors report state-of-the-art accuracy in recovering object-level 3D occupancy on several benchmarks. They also note that combining the meshes with 3DGS rendering enables physical interactions.
In briefTask-Oriented Rank Adaptation for Continual Learning in Text Classification
SummaryThis paper addresses catastrophic forgetting and negative transfer in continual text classification. It proposes TORA, a method that uses the low-rank structure of LoRA adapters to decide whether to transfer knowledge from the most compatible expert (Boosting) or isolate a new task (Shielding). Evaluated on 15 text classification benchmarks, TORA reportedly improves compatible tasks while reducing training time, and protects structurally distant tasks from interference.
Rey Sanchez Lopez, Eduardo Morales Manzanares, Hugo Jair Escalante
In briefLearning PDE dynamics between submanifolds using Green's observation operators
SummaryMany physical systems are driven and observed only on low-dimensional submanifolds, while their dynamics are governed by the surrounding medium. The paper introduces the Green's Observation Operator (GObO), which maps the ambient medium once to the Green's kernel of a linear PDE restricted to the source and observation submanifolds, so new sources need only one lower-dimensional integral. On three-dimensional heat conduction and advection-diffusion problems, GObO predicts responses to moving sources zero-shot with 4 to 8 times lower error than black-box surrogates, at 1.4 ms per query.
Jan Tauberschmidt, Jephte Abijuru, Naukshatro Bose et al.
In briefAcmite: Reducing Gender Bias in LLMs Using Concept-Guided Mutual Information
SummaryThis paper addresses social stereotypes that large language models reproduce from their training data. It proposes Acmite, a lightweight framework that represents stereotypes as semantic concepts, selects diverse ones with maximal marginal relevance, and approximates the dependence between outputs and these concepts using token-level KL divergence. A LoRA adapter is activated at inference only for inputs similar to stereotype-related concepts. Experiments on BBQ, CrowS-Pairs, and StereoSet show reduced gender bias across three LLMs, while performance on ARC-Challenge, GSM8K, and PIQA remains competitive.
In briefArtifact Annotations Partially Substitute for Per-User Calibration: SAFE-EDA Wrist EDA Affect Recognition
SummaryThe paper introduces SAFE-EDA, a compact convolutional network pretrained on expert artifact annotations from 43 subjects, and compares it with a network trained from scratch on the WESAD dataset using leave-one-subject-out evaluation. When normalization statistics come only from training subjects, pretraining raises macro-F1 by 0.078 to 0.227. When statistics come from the held-out user's full recording, the gain shrinks to 0.020 to 0.050 and is no longer significant. The authors argue that reporting the source of normalization statistics is needed to separate first-use performance from post-calibration performance.
In briefCompound meaning as analogy: the Compound Analogy Model (CAM)
SummaryThis paper proposes the Compound Analogy Model (CAM), a parameter-free method that predicts a compound's embedding by adding its constituent embeddings and the average shift vectors of their compound families. On Mandarin Chinese compounds, CAM outperformed the CAOSS model in prediction accuracy except for three-character compounds. CAM-derived measures also better predicted visual lexical decision response times for two-character compounds, supporting local analogical generalization as a plausible account of compound comprehension.
In briefArchitectural Sampling: Diverse Computation Paths for Better Answers from Frozen Vision-Language Models
SummaryThis paper proposes architectural sampling, a training-free method that generates candidate answers by reusing selected blocks of decoder layers, so each candidate follows a different forward computation path. The method updates no weights and adds no parameters. Across five Qwen checkpoints and twelve multimodal benchmarks, it improves pass@9 by 6.58 percentage points on average over standard temperature sampling at the same nine-candidate budget. Reusing early layers gives the largest gains, and the coverage improvement persists under greedy decoding.
In briefMiLoop: Memory-Propagating Reinforcement Learning for Combinatorial Optimization
SummaryThis paper proposes MiLoop (Memory-in-the-Loop), a purely reinforcement learning framework that builds solutions step by step for combinatorial optimization problems. It fuses current embeddings with historical memory and applies adaptive gated updates, allowing a shallow policy to learn dynamic embeddings without solution labels or training-time search-space pruning. Experiments across four problem types report high-quality solutions on instances from 100 to 10 million nodes, indicating strong generalization.
Changliang Zhou, Yuanyao Chen, Rongsheng Chen et al.
In briefMuLoRA: A Spectrally Balanced Low-Rank Adaptation Method for Continual Learning
SummaryThe paper identifies spectral plasticity collapse, where LoRA updates concentrate in a few singular modes during sequential adaptation, leaving much low-rank capacity unused. It proposes MuLoRA, which uses historical whitening to select input directions responsive to the current task and approximate polar orthogonalization of momentum updates to reduce spectral concentration. Across five class-incremental benchmarks and eight incremental settings, MuLoRA reports the highest mean accuracy in 15 of 16 metrics.
In briefBeyond Leaderboard Scores: A Deployment-Focused Protocol for Tracking Evaluation
SummaryThis paper proposes a tracker-only evaluation protocol for pedestrian-tracking robots that uses shared detections to isolate tracker behavior. It assesses initialization, continuation without detector support, identity recovery, close-neighbor association, and runtime on an embedded NVIDIA Jetson Orin. Tested on the JRDB dataset with six open-source trackers and the PedRefTrack baseline, results show that after 1.0 s without detections, no non-GT tracker keeps correct, same-identity output in more than half of cases, and several trackers fall below 10 Hz in crowded frames.
WhyProposes a deployment-oriented tracker protocol separating initialization, detector-gap continuation, identity, and runtime, validated on JRDB with six trackers.
In briefWhen Text-to-Image Helps Editing: Conditioning Effects During Denoising
SummaryUnified models are trained for both instruction-based image editing and text-to-image (T2I) generation. This paper asks whether editing can benefit from T2I, observing that in pure editing, source attention declines for some edits during sampling. The authors propose task switching, which temporarily uses the T2I task for bounded intervals. Across three unified editors and four benchmarks, this improves edit quality while mean perceptual preservation stays close to pure editing, and the switch timing sets the quality-preservation balance.
Lidia Troeshestova, Alexander Ustyuzhanin, Sergey Kastryulin
In briefInvent a Dataset: Testing how well AI can create training data from scratch
SummaryThis paper addresses the common case where practitioners have no data at all for a capability they want to teach an AI. It introduces Invent-A-Dataset, a prompt-based system that turns a dataset description into realistic, large-scale post-training data. Across eight task types and up to 20K samples, it reportedly beats five frontier model APIs on both quality and diversity, with the diversity gap widening as dataset size grows. Models fine-tuned on its data also rank higher than those fine-tuned on other generators.
Shivalika Singh, Andrija Djurisic, Gbemileke Onilude et al.
In briefDo MLLM Judges Really Evaluate the Edit? Auditing Bias with Verified Quality Preservation
SummaryThis paper introduces EditJudgeBias, a counterfactual benchmark of 1,196 real image-editing samples with 13 cues injected at four evaluation sites. The authors verify that these cues preserve the requested edit's quality using calibrated multimodal validators, controls, and human inspection. They then audit five multimodal large language model judges for cue invariance, agreement with human judgments, and stability of pairwise preferences. Results show that quality-preserving cues shift every judge beyond its own noise floor, and swapping candidate order reverses up to 60.9% of pairwise decisions.
In briefpCoMole: Pareto-constrained editing of biomolecule sequences using discrete flows
SummaryThis paper introduces pCoMole, a framework that edits biomolecular sequences to improve multiple properties while enforcing hard feasibility constraints. It builds on discrete flow matching to steer a pre-trained Edit Flow toward user-specified preferences. The authors test it on shortening GFP, Cas9 orthologs, and peptide binders, and report that two 229-residue eGFP variants retained green fluorescence in BL21 cells in wet lab tests.
WhyCombines discrete flow matching with Doob-h transform for constrained multi-objective sequence editing, with wet-lab GFP validation; impact is mainly within biomolecular design.
In briefLower bounds for variance-reduced stochastic first-order methods in nonconvex-concave minimax optimization
SummaryThis paper establishes complexity lower bounds for stochastic first-order algorithms solving nonconvex-concave minimax problems, including algorithms that use variance reduction. It proves a bound for a zero-respecting algorithm class under an unbiased oracle with bounded variance and mean-square smoothness, showing dependence on accuracy, dual-domain radius, and noise. Complementary bounds are also given for the nonconvex-strongly-concave setting.
In briefDiVid: Diagnosing where video generators lose diversity, dimension by dimension
SummaryVideo generation models often produce very similar outputs when repeatedly sampled from the same prompt. The authors introduce DiVid, which splits video diversity into six interpretable dimensions (Semantic, Style, Subject, Scene, Motion, Camera) and measures each with a reproducible computer-vision pipeline. They find diversity is highly dimension-specific, with Motion and Camera collapsing most, and that temporal factors show larger faithfulness losses when controlled through text alone.
In briefAutomatically refining structured robot policies by having an LLM analyze their rollouts
SummaryThis paper proposes a closed-loop method in which an LLM analyzes logged policy rollouts, stored as tabular data, and generates diagnostic code to iteratively correct structured policies without human instruction. On car racing and door opening tasks, the approach improves imitation learning performance by up to 15% over zero-shot LLM-generated structures. It also reports 75% less compute needed to reach the same reinforcement learning performance.
In briefCrossGMN: Graph metanetworks for transforming neural network weights across different architectures
SummaryThe paper introduces CrossGMN, a graph metanetwork that operates directly on neural network parameters to map a trained source network into a target network with a different architecture, such as for model compression. It jointly processes the source network and a target initialization through symmetry-preserving cross-network message passing. The authors report that CrossGMN speeds up knowledge distillation by up to 8.89x across 2-D and 3-D INRs and image classifiers, and that it transfers across datasets.
WhyProposes a new equivariant cross-architecture formulation with a universality proof, validated on INRs and image models, achieving up to 8.89x distillation speedup.
In briefA cloud-fog-edge system idea for smart buildings that learns locally
SummaryThis paper presents a vision for a decentralized system in which the building itself acts as the data center, learning from real-time data inside the building to protect privacy and reduce reliance on external cloud infrastructure. It describes a lightweight, Kubernetes-like framework for deploying AI services and reports progress in running AI algorithms on low-power microcontrollers such as Arduino boards. The authors also evaluate two new online algorithms and present experimental results using real-world data traces.
In briefMulti-Metric and Intersectional Fairness Evaluation in MIMIC-IV ICU Mortality Prediction
SummaryThis paper compares several fairness interventions for ICU mortality prediction on MIMIC-IV, using predictive-utility and subgroup-error metrics. It also introduces a lightweight method that balances ethnicity, gender, and insurance representation without conditioning on mortality outcomes, evaluated at marginal and three-way intersectional subgroup levels. The results show that intervention assessments vary substantially across accuracy/AUROC, sensitivity, and false-positive rate, and that marginal demographic summaries can hide heterogeneous error profiles within intersections.
Abdullah Al Noman, Fahmid Al Rifat, Tahrima Hashem et al.
In briefMCIR: A dependence-aware method for explaining model features with reliability guarantees
SummaryMachine-learning models often have correlated or redundant features, which can make standard explanation methods like SHAP and LIME give unstable rankings. This paper proposes MCIR-M, a global importance method that measures each feature's unique predictive information beyond its strongly dependent neighbours, producing a normalized score. It also introduces a lightweight estimator that uses a fraction of the data. Experiments show clearest gains on synthetic data with injected near-duplicate features, while comparisons on real data are mixed.
Poushali Sengupta, Sabita Maharjan, Frank Eliassen et al.
In briefNot All Errors Fall With Scale: Where Scaling Stops in Vision-Language Inference
SummaryThis paper examines how vision-language models should split a fixed compute budget between a larger language backbone and more visual tokens for high-resolution inputs. The authors propose a Separable Law that describes performance as a function of backbone size and visual token count, fitted to 26 InternVL and QwenVL models on four high-resolution benchmarks. They report that some questions respond to scaling in a way predictable from the skill required, while a substantial fraction does not respond at all.
In briefFedSAP: Federated Learning with Structured Adaptive Partitioning for Heterogeneous Edge Devices Across Domains
SummaryThis paper proposes FedSAP, a federated learning framework for edge devices with unequal resource budgets and domain-shifted local data. It treats structured pruning as budget-constrained tri-state channel allocation, placing stable channels in a global pool, useful domain-sensitive channels in pseudo-domain private pools, and low-utility channels in a dropped state. On Digits and Office-Caltech, FedSAP reaches mean global accuracies of 76.00% and 72.67% across three random seeds, exceeding the strongest baseline by 1.70 and 4.92 percentage points, while supporting client pruning ratios of up to 80%.
In briefVietnamese visual question answering using a multi-layer fusing Transformer that combines image and text features
SummaryVisual Question Answering (VQA) asks computers to answer natural-language questions about images, but very little work has been done for Vietnamese. This paper proposes a Multi-layer Fusing Transformer that uses a cross-attention module to merge image and text features from different layers into one representation, moving from low-level to high-level information. Through experiments and ablation studies on the ViVQA dataset, the authors report that the model achieves promising results against competitive baselines.
In briefYo-ByT5: An efficient, high-fidelity model that restores Yorùbá tone marks
SummaryYorùbá is a tonal language whose diacritics distinguish word meanings, but they are often omitted in writing, which hampers NLP tasks. The authors introduce Yo-ByT5, a byte-level diacritic restoration model fine-tuned from ByT5-small. On the YAD benchmark it is compared with five public Yorùbá models and one open-weight large language model. It matches the strongest existing model, mT5-base, with a DER of 10.14% and a CER of 3.48%, while using about half the parameters and showing better text fidelity.
Ahmad Samuel Gali (1), Shamsuddeen Hassan Muhammad (2, 3) ((1) University of Lagos et al.
In briefNeural network generalization viewed through magnitude potential
SummaryThis paper introduces magnitude potential, a quantity based on metric magnitude theory that measures how well a point is represented by a given set. The ratio of class-level to full-data magnitude potential, computed at the logit layer, correlates with Feldman memorization scores for individual training points. Aggregated across points, the ratios detect changes in decision boundaries and serve as a geometric indicator of grokking in modular arithmetic. The authors report that this ratio stays informative even when neural collapse is explicitly suppressed.
In briefMaintaining Learned Cooperation in Multi-Agent RL via Gradient Routing
SummaryThis paper studies whether cooperative behavior learned by multi-agent reinforcement learning agents survives continued training. It compares three matched setups: value-loss gradients updating shared actor features, a critic kept but its gradients blocked, and a critic-free reference. Higher reward scales selectively made cooperation less stable when value gradients reached shared actor features, while blocking those gradients kept it near the ceiling. The authors conclude the risk is conditional on direct value-gradient routing rather than a general failure of critics.
In briefWhat makes a question harder? Natural-language explanations of difficulty
SummaryThis paper proposes a data-driven method that automatically generates and validates natural-language hypotheses explaining why one question is harder than another. Difficulty is first estimated from responses of many LLMs using Item Response Theory, then an LLM proposes explanations by contrasting easy and hard questions, which are validated on held-out items. Across mathematical, logical, and commonsense reasoning datasets, the hypotheses are interpretable and predictive, and they also improve existing difficulty regressors.
WhyProposes automatically generated, validated natural-language difficulty hypotheses, with causal checks via question edits across three datasets, giving fairly strong evidence.
In briefMeasuring whether injected action errors shrink or grow under action chunking
SummaryThe paper studies why action chunking improves behavioural cloning policies by injecting small action errors at each state and measuring their propagation under open-loop and closed-loop execution. Across twelve manipulation tasks from three benchmark suites, confidently stable states are rare and error amplification is common. The measured propagation rate also depends strongly on the fitting horizon.
In briefBeyond Domain-Level Adaptation: Margin-Oriented Semantic-Appearance Correction for Personalized Federated Vision-Language Models
SummaryThis paper studies domain heterogeneity across clients in federated parameter-efficient fine-tuning of vision-language models. The authors argue that treating each domain as a class-agnostic transformation is insufficient, since only some class-domain residuals harm the image-text decision margin. They propose MOSAIC, which uses a decision-aware harmfulness score, a low-rank residual adapter with globally shared class factors and client-private domain factors, and an image-conditioned gate. Experiments on Office31, OfficeHome, and DomainNet100 report consistent gains in macro-client top-1 accuracy across the evaluated settings.
In briefFedLore: Efficient Federated Learning with Shared Low-Rank Gradient Projection
SummaryThis paper addresses high memory and communication costs when federated training large foundation models. It proposes FedLore, which shares a low-rank optimization basis across clients within each round and refreshes it between rounds, aiming to remove the projection bias caused by independently chosen client subspaces. Experiments on vision and language tasks, including federated pre-training, report that FedLore outperforms the evaluated low-rank adapter baselines and matches or exceeds full-parameter training while reducing communication and optimizer-state memory.
In briefMeasuring the Hidden Energy Cost of Deep Learning Audio Development
SummaryThis paper argues that studies of deep learning's environmental impact have focused on training and inference while overlooking the development phase, which involves architecture prototyping and intensive experiments. The authors propose a methodology to estimate these costs using activity logs from the Grid5000 shared computing platform used by the LORIA laboratory. Their case study of four audio projects from the Multispeech team finds that development energy was 3 to 256 times greater than the energy needed to train the best-performing model alone. The authors call for more systematic reporting of energy consumption across the full life cycle of deep learning audio projects.
In briefAgents are configurable systems, not fixed models; evaluation should reflect this
SummaryThis paper argues that agent evaluations usually treat the agent as fixed, although its performance depends on configurable choices. The authors test five configuration factors (task information, reasoning, self-verification, time budget, and backbone model) on a new benchmark of four scientific tasks. They find that repeating the same configuration accounts for about 54% of outcome variance, and that the information given to the agent has the largest effect, exceeding time budget and model size.
WhyFrames agents as configurable systems, with extensive configuration experiments and trajectory analysis showing information and verification tools matter more than prompting or time.
Luis Wiedmann, Leander Girrbach, Cordelia Schmid et al.
In briefCan LLMs reliably annotate bioassay metadata to improve data readiness?
SummaryThe paper finds that much of PubChem's roughly 2 million bioassays lack annotations such as assay format and detection technology. The authors use seven open-source and proprietary LLMs to predict and audit these metadata fields from assay text, comparing them with existing silver labels. Recall is at least 0.96 for biochemical and cell-based formats. The authors conclude LLMs can support large-scale annotation and auditing, but per-class reliability checks and human review are still needed before use in downstream machine learning pipelines.
Laura van Weesep, Riccardo Tedoldi, Jens Sj\"olund et al.
In briefOneira: Moving Video World Models from Open-Ended Generation to Open-World Interaction
SummaryThis paper presents Oneira, an interactive video world model that links generation and interaction through an explicit world state managed by a coding agent. Newly generated objects are incorporated into the actionable world, and interaction outcomes are written back so they persist into later video segments. Experiments show direct and consistent interaction with newly generated objects and preservation of prior interaction effects over long horizons.
In briefHob-VL: A Benchmark for Image-Based Boolean Reasoning
SummaryThis paper introduces Hob-VL, a benchmark for testing whether vision-language models can perform Boolean reasoning grounded in images. It has two tasks: judging whether a Boolean rule holds in an image, and identifying the unique object matching a Boolean description. The dataset includes 6,000 human-verified Yes/No questions and 1,000 object-identification questions across generated scenes and labeled photographs. Across eight model configurations, Boolean accuracy ranged from 48.52% to 50.57% and identification accuracy reached at most 43.0%, with substantial remaining errors and inconsistencies.
In briefOrder-robust decision models that score each candidate independently
SummaryThis paper studies decision models that score a variable-sized set of candidate actions encoded in one sequence, noting that standard causal attention can make scores depend on serialization order. The authors propose candidate-independent block-causal attention, which keeps causal computation within the shared context and each candidate while blocking cross-candidate information flow and resetting candidate positions. Experiments on Gemma 3 1B, Qwen3 1.7B, and Qwen3 4B show reduced permutation sensitivity while retaining competitive decision quality. Ablations suggest candidate isolation is the main source of the effect.
In briefHow the STORM optimization method converges on different types of problems
SummaryThe paper studies convergence of the STORM (stochastic recursive momentum) method for nonconvex, convex, and strongly convex objectives without assuming average smoothness. An auxiliary sequence is used in the analysis to show a rate of O(T^{-1/4}) for nonconvex objectives. For convex and strongly convex cases, the authors prove averaged and last-iterate bounds with optimal rates. All results use the same recursion with different hyperparameter choices.
In briefReinforcing Temporal Representations at Inference Time in Video LLMs
SummaryThe paper finds that Video Large Language Models often give the same answer even when video frame order is reversed, revealing weak temporal reasoning. The authors define a temporal divergence vector that peaks at intermediate layers and then fades toward the output, suggesting temporal information is lost late in the network. They propose Temporal Activation Injection (TAI), a training-free method that reinjects the peak temporal signal into later layers, reporting consistent gains across three VideoLLMs and four benchmarks with negligible impact on non-temporal tasks.
In briefChoosing the best LLM via online active selection with limited labels
SummaryThis paper proposes ONLINE LLM PICKER, a framework for actively selecting the best large language model in online settings. Given a stream of queries and a limited annotation budget, it chooses the most informative prompts to annotate and identifies the top candidate model. Across tasks covering 10 datasets and over 130 language models, the authors report annotation cost savings of up to 71.67% while reliably finding the best or near-best model.
Alessandro Turrin, Patrik Okanovic, Torsten Hoefler et al.
In briefTwo Routes to the Middle: Placement and Brain Signals Point to Where Continual Learners Should Specialize
SummaryThe paper studies where to place task-specific adapters in a few blocks of a vision transformer to limit storage growth in continual learning. An exhaustive search over contiguous four-block placements finds accuracy peaks at intermediate depth, varying by up to 3.5 percentage points, while cheap weight- or activation-based criteria favor the deepest blocks. The authors propose LS-B, which uses a frozen fMRI encoding model to pick specialization blocks from the first tasks without labels or backpropagation, reaching within 1.5 points of full accuracy using 60% of the adapter storage on Split ImageNet-R.
In briefPAGER: Aligning partial camera views with global 3D features using geometric and relational distillation
SummaryThis paper shows that pretrained 3D encoders, trained on globally reconstructed scenes, perform very poorly on partial, camera-coordinate observations; a frozen Sonata encoder drops from 72.47 mIoU on full ScanNet scenes to 2.57 mIoU on single-frame inputs. The authors propose PAGER, a label-free method that trains lightweight adaptation modules using paired partial and global geometry, aligning partial-view features with a frozen global semantic space. PAGER outperforms label-supervised parameter-efficient fine-tuning on Sonata and Concerto, and in zero-shot ScanNet to ScanNet++ transfer it surpasses fully fine-tuned Sonata (53.93 vs. 48.09 mIoU).
Akira-Miranda Adeyomi Adeniran-Lowe, Binod Singh, Lars Arnold Dethlefsen et al.
In briefContinual Reinforcement Learning via Neuroevolution: Searching Weight Space for Changing Tasks
SummaryThe paper notes that reinforcement learning methods struggle to balance adapting to new tasks with retaining earlier skills when tasks keep changing. The authors instead use neuroevolution, which searches weight space directly through mutation and selection over populations of neural networks, and compare evolution strategies and genetic algorithms against continual RL and population-based RL. Evolution strategies most consistently achieve a good stability-plasticity trade-off, while genetic algorithms are the most plastic but forget more.
In briefA Five-Layer Protocol for Checking Physical Consistency in Generative Driving Scenario Models
SummaryThis paper introduces a layered evaluation protocol for generative AI models that produce autonomous driving scenarios. The first four layers examine internal representations through kinematic alignment, statistical baseline comparison, latent controllability, and activation analysis, while the fifth checks outputs against vehicle-dynamics constraints such as lateral jerk thresholds. The authors demonstrate the protocol on a VAE-based scenario generator, showing that realistic-looking outputs do not guarantee consistency with vehicle dynamics. They also apply it to other generative models.
Manasa Mariam Mammen, Zafer Kayatas, Stefan Wagner
In briefBeyond Pointwise Error: Multi-Metric Evaluation of Climate Downscaling
SummaryThis paper examines how to evaluate climate downscaling, which reconstructs fine-scale spatial fields from coarse-resolution inputs. It benchmarks five downscaling methods on ERA5 temperature, wind, and precipitation using five metrics covering pointwise error, structure, distribution, spectra, and gradients. The authors find a systematic trade-off between spatial fidelity and fine-scale variability, and method rankings change across metrics and variables. They conclude that no single method is best and that multi-metric evaluation is essential.
In briefTheory explaining why resampling masks helps masked autoencoders learn better features
SummaryThis paper studies masked autoencoders in a high-dimensional model and shows that masked linear reconstruction can recover latent features where unmasked reconstruction fails. It quantifies the statistical benefit of resampling masks and shows that standard random cropping and flipping can hide this advantage. Without those transformations, dynamic masking outperforms static masking for CNN autoencoders and vision transformers.
WhyProves masked linear reconstruction recovers latent features at linear sample complexity where unmasked PCA fails, and quantifies mask resampling's advantage with supporting experiments.
Jorge Medina Moreira, Lorenzo Bardone, Lenka Zdeborov\'a
In briefOptimal momentum methods for stochastic multilevel composition optimization: faster convergence for nested non-convex objectives
SummaryThis paper studies stochastic multi-level optimization, where the objective is a nested composition of several smooth non-convex functions and only noisy estimates of gradients and function values are available. The authors use a mini-batch momentum estimator to track each level's function values and build momentum gradient estimates, achieving an optimal sample complexity of O(ε^-4) for ε-stationary points without the stronger average smoothness assumption. They also present a batch-free variant and validate the methods on risk-averse portfolio optimization and hierarchical tilted empirical risk minimization.
In briefManaging Context and Communication in Distributed Agentic UAV Swarms
SummaryThis paper studies drone swarms in which each UAV runs its own small language model, raising the problem of managing information as histories grow and communication spreads. The authors propose an architecture with an event-driven reason-act-observe loop, structured memory split into core, local, and peer-specific layers, and an interest-aware gossip engine that shares notes based on recipient-specific novelty and recency. In a simulated search-and-rescue mission with ten UAVs, the approach completed all runs, while unrestricted flooding completed only 70-85%. Compared with flooding, it roughly halved inference-token consumption, reduced transmitted data, and achieved lower survivor-count error.
Andrea Iannoli, Ivan Zyrianoff, Angelo Trotta et al.
In briefTowards Optimal Policy Improvement: A New Greedification Operator Under Approximate Evaluation
SummaryThis paper studies policy improvement from first principles, defining optimal policy improvement as the best policy attainable in a single update under given constraints. It shows that optimal improvement restricted to a set of states is equivalent to solving an induced MDP, and formulates greedification under approximate evaluation as probabilistic decision-making under uncertainty. The authors derive a novel operator optimal for this objective, and experiments report that the operator and its gradient-based approximations improve aggregate performance across several RL methods.
WhyFormalizes optimal policy improvement, derives a new operator, and shows aggregate gains across several RL algorithms and action settings.
In briefAURAL: Faster Adaptive Latent Reasoning for Speech Language Models
SummaryThis paper introduces AURAL, which reasons in latent space by modeling multiple plausible reasoning paths and jointly predicting chunks of future states to cut sequential passes and latency. The authors build AuralReason-683K, about 1,000 hours of bilingual speech with concise chain-of-thought for emotion recognition, empathetic dialogue, and general reasoning, to supervise latent reasoning. On Qwen2.5-Omni, time to the first answer token drops from 1.22 s to 0.10 s, an 11.8x reduction.
In briefCompletion Aware Guidance for World Action Models: a training-free sampling method that reduces task-incomplete imagination
SummaryThe paper observes that World Action Models (WAMs), when adapted for short-chunk control, can produce plausible, action-consistent predictions that omit the transition needed to finish a task. The authors introduce Completion Aware Guidance (CAG), a training-free sampling method that steers generation toward task completion. The reported results on representative WAMs show success rising from 64% to 70% on a RoboTwin 2.0 subset and from 69% to 75% in zero-shot simulation, with task-incomplete imagination falling from 79% to 40%.
In briefQK-Wanda: Pruning Language Model Attention Weights Jointly
SummaryThis paper introduces QK-Wanda, a pruning method for large language models that scores query and key weights together, letting them share one pruning budget instead of being scored independently as in Wanda. Its closed-form scores need no gradients or weight updates, and full pruning takes only slightly longer than Wanda. Tested on 15 models from 0.5B to 72B parameters, it lowers query-key reconstruction error on average, but downstream quality gains depend on the model, and some models show higher perplexity despite lower error.
In briefNeural Parser Builds Patent Invention Graphs for Prior Art Search
SummaryPatent search involves documents often exceeding tens of thousands of tokens, which most neural retrieval methods cannot fully process. The paper proposes a neural parser that adapts biaffine attention from dependency parsing to predict invention graphs directly from patent text, using a sliding window to reduce complexity. Trained on short sequences, it generalizes to documents over 40,000 tokens without retraining. Distilled from one million rule-parsed documents, it outperforms its teacher at lower inference cost, improving citation recall by 0.5% on short queries and 1.1% on full documents in a downstream Graph Transformer system.
In briefRange-GRPO: policy optimization using pairwise comparisons of reward ranges
SummaryThis paper addresses post-training of large language models when reliable supervision is scarce. It uses LLM-as-a-Judge to produce conformally calibrated reward ranges for unlabeled responses instead of single point scores. The proposed Range-GRPO method compares these reward ranges pairwise within the GRPO framework, so interval uncertainty affects both the magnitude and direction of learning signals. The authors report the highest average in-distribution and out-of-distribution performance among the semi-supervised methods evaluated, while requiring fewer training resources.
In briefReinforcement learning that speeds up the PDHG method for linear programming
SummaryThe paper introduces GALLOP, which uses reinforcement learning to jointly learn continuous parameters and discrete restart decisions for primal-dual hybrid gradient (PDHG) solvers, without differentiating through the solver. Evaluated on six linear program families and a public benchmark, it reduces iteration counts and achieves up to a 16.0x wall-clock speedup over MPAX. One policy trained per family generalizes without retraining to within-family LPs up to 400 times larger than the training instances.
WhyUses reinforcement learning to jointly tune PDHG parameters and restarts with a generalized accelerated update. Reports 1.9–5.6x fewer iterations and up to 16x speedup across six LP families, with sol
