In briefClosing the Estimation Gap: Training-Free Transductive Inference for Cross-Domain Few-Shot Action Recognition
SummaryThis paper addresses recognizing unseen action classes in a target domain from only a few labelled videos. An oracle analysis with fixed representations shows that accurate prototypes and decision boundaries could raise 5-way 1-shot accuracy by 7 to 20 points, indicating the bottleneck lies in estimation from scarce support examples. The authors propose TRACE, a training-free transductive framework with confidence-thresholded prototype refinement and transductive classifier refinement that uses unlabelled queries. TRACE consistently improves a strong baseline across five datasets, including over six points on HMDB51 1-shot.
In briefDelay Complexity Dimension: learning when labels arrive late or get cut off
SummaryThe paper studies PAC learning of binary classifiers when labels arrive after random delays and are censored beyond an observation window W. It proposes a distribution-dependent Delay Complexity Dimension (DCD), combining the VC dimension with the observation channel's label-information rate. The authors show every learner needs Ω(DCD/ε²) samples, and an oracle pseudo-label ERM matches this rate on the lower-bound instance family. Experiments on synthetic data and Criteo conversion logs support the predictions.
In briefSWE-MISBEHAVIOR: A benchmark for coding agents that break their operating boundaries
SummaryThis paper introduces SWE-MISBEHAVIOR, the first taxonomy and evaluation framework for diagnosing coding agents that violate their boundaries. It includes 7.4K human-in-the-loop synthesized samples covering 11 misbehaviors across code generation, file management, and system interaction. Experiments on open-source agents such as OpenHands and smolagents show frequent misaligned behaviors, such as dependency misinterpretation and silent compromise of environment integrity. The authors also propose CODE-KIT, a lightweight framework that reportedly reduces misbehaviors by 52% for smolagents and 34% for OpenHands while improving task completion.
Do Xuan Trong, Duc Anh Nguyen, Do Xuan Long et al.
In briefReset Early, Reset Often, Eliminate Models: A Connection-Tableau Prover Improvement
SummaryThis paper studies connection-tableau automated theorem provers, which usually backtrack and retain little reusable information after failed attempts. The authors propose SATResetCoP, which abandons unproductive tableaux and starts new ones, feeding more ground clauses to an incremental SAT solver. It improves over its baselines by 54% and 18% on TPTP, and solves 616 problems on the MPTP2078 benchmark.
Martin S. Fixman, Fredrik R{\o}mming, Sean B. Holden
In briefCan Large Language Models Reliably Assess Soil Organic Carbon? An Australian Benchmark
SummaryThis paper tests seven general-purpose large language models on estimating soil organic carbon using Australian data. The tasks cover SOC concentration estimates for 308 paddocks, stock calculations in 120 scenarios, and comparisons of 300 paddock pairs. Adding measured soil attributes reduced concentration errors, but high-SOC underprediction persisted, and nearly all outputs were parseable without necessarily being correct. The authors recommend task-specific validation and auditable calculations rather than using LLM outputs unverified as ecological evidence.
In briefA Few More Tokens Ahead: Strengthening Parallel Generation in Diffusion LLMs
SummaryThe paper finds that diffusion large language models, though often considered inherently suited to parallel generation, degrade severely on math and code tasks when generating 16 tokens per step, with all scores below 15%. The authors propose RePara, which reinforces short consecutive token blocks via group-relative policy optimization and reduces rollout and log-probability recomputation costs. On LLaDA and SDAR, RePara improves GSM8K and MATH-500 accuracy over JustGRPO under reduced denoising budgets, and yields a 1.96x training-step speedup in the LLaDA/MATH setup.
In briefQORH: Query-specific ordered rule hypergraphs for inductive knowledge graph link prediction
SummaryThis paper addresses link prediction in knowledge graphs involving unseen entities and relations. It proposes QORH, which encodes rule heads and ordered rule bodies as query-conditioned hyperedges to model joint dependencies among multiple relations, and selects rules learnably using statistics such as confidence and coverage. The authors report that QORH significantly outperforms the baselines included in their evaluation on entity-inductive and fully inductive benchmarks.
In briefHierarchical forecasting with operator-valued kernels that keep predictions consistent across levels
SummaryThis paper addresses hierarchical time series forecasting, where predictions must stay coherent across multiple levels of aggregation. The authors cast the problem as multivariate regression in a vector-valued reproducing kernel Hilbert space and design structured operator-valued kernels that enforce hierarchical constraints by construction. They derive excess risk bounds describing generalization performance. Experiments on synthetic and real-world datasets are reported.
Junjie Yang (LPSM (UMR\_8001), SU), Adeline Fermanian et al.
In briefDual-Endpoint Image Attention for Predicting Millimeter-Wave Beam-Pair Power in Vehicle-to-Infrastructure Links
SummaryThe paper argues that sensing at a single endpoint gives only a partial view of signal propagation, so it proposes cross-image routing attention (CIRA), which sparsely links and mutually enhances propagation-related features from RGB images at both ends. Building on CIRA, the CiFormer model predicts millimeter-wave beam-pair power matrices for vehicle-to-infrastructure (V2I) communication. Experiments report lower mean absolute error and a 3.15 percentage point gain in top-1 beam-pair selection accuracy over the strongest baseline.
In briefVARI: Making Text-to-Video Generation Robust to Different Wordings of the Same Request
SummaryThis paper addresses the problem that text-to-video models can produce different results when the same meaning is expressed with different prompt wording. It introduces VARI, a framework that treats automatically verified prompt variants of one visual intent as a shared training unit, using Group-Consistent Distillation and Form-Contrastive Flow-DPO. On HunyuanVideo-1.5, VARI reports the highest T2V-CompBench, VBench, and StoryEval scores among the evaluated methods, along with improved weakest-form alignment and reduced cross-form drop while preserving visual quality.}
In briefThe Agentic Scientist: A system that coordinates coding agents to carry out automated scientific research, alone or alongside humans.
SummaryThe paper introduces The Agentic Scientist, a distributed pipeline in which human researchers and coding agents run computational research together, covering task drafting, idea generation, planning, sandboxed experiments, analysis, and report writing. A meta-controller manages parallel research agents and enforces rubric verification at each stage. The authors report that a paper co-written with a development version of the system was accepted at ALIFE 2026, and that the system was applied fully autonomously to open machine learning problems. Ablations suggest reliable scientific autonomy depends on the cross-agent orchestration layer as well as base model capability.
Robert Tjarko Lange, Luca Grillotti, Yutaro Yamada et al.
In briefReinforcement learning for fast lifelong multi-agent path finding using shared spatial fields
SummaryThis paper proposes a navigation policy for lifelong multi-agent path finding, where a convolutional network computes a shared spatial field over the map each timestep and a compact operator derives agents' movement preferences. The policy is trained end-to-end with average-reward reinforcement learning using throughput as the reward, without planner demonstrations or a separate path planner. Across six layouts with 10000 agents, it achieves throughput comparable to or higher than the strongest baselines, with an average execution time below 10 ms per step.
In briefFrameScout: Selecting Query-Relevant Frames for Long Video Understanding
SummaryLong video understanding asks video language models to answer questions about videos lasting tens of minutes to hours, which produce more visual tokens than current models can handle. FrameScout introduces a streaming keyframe selector that uses a sliding-window KV cache and successor aggregation to create temporally contextualized frame embeddings, along with a frame-query contrastive objective that trains it to separate relevant from irrelevant frames. The authors report consistent state-of-the-art results across three long-video benchmarks and six base VideoLLMs.
In briefNeuro-symbolic graph Transformer for recognizing layered human intentions from activity sequences
SummaryThis paper proposes a Neuro-Symbolic Graph Transformer that injects a Human Reasoning Graph into self-attention as learnable topological biases and adds a differentiable first-order logic loss to enforce consistency across four levels: actions, activities, low-level intentions, and high-level intentions. It is evaluated on NTU-Hierarchy, a four-level extension of NTU RGB+D 120 with 119,383 episodes. With the logic loss at λ = 0.5, logic consistency reaches 94.59% while top-1 action accuracy remains 89.90%.
In briefLooped reasoning converges faster than expected
SummaryThis paper studies convergence in looped reasoning models with dynamic depth, where latent trajectories often oscillate rather than settle. The authors show that slowly decaying oscillatory modes can limit convergence speed, while damping them indiscriminately can disrupt reasoning. They propose QRR (Quotient Recurrent Reasoner), a gauge-equivariant recurrent architecture with exact per-token phase-rotation symmetry that decouples rotational motion from reasoning dynamics. The authors report that QRR converges more stably and quickly and matches or surpasses recent recurrent reasoning models on Sudoku, Maze, and Mini-ARC.
In briefTRACESTEER: Trace-Driven Activation Steering to Reduce Hallucinations in Diffusion Language Models
SummaryThis paper proposes TRACESTEER, a framework that reduces hallucinations in diffusion large language models (dLLMs) by steering their internal activations. It uses predictions from earlier denoising steps as feedback to adjust both the direction and strength of the steering. The method has two parts: DF-Steering, which builds step-specific steering vectors, and TS-Gate, which controls strength based on prediction uncertainty and timestep. Experiments on two dLLMs show improved factual reliability on TruthfulQA and generalization across five downstream benchmarks.
In briefBack to the Future: Reverse-consequence modeling for predicting long-term object motion in physics simulations
SummaryThis paper addresses the problem of error accumulation in autoregressive models when predicting long sequences of object motion. It proposes ReStep, which reverses the usual forward-prediction approach by first predicting the final state with a stochastic model, then generating intermediate trajectories with Anchored Diffusion conditioned on that endpoint. The method is also integrated into a 3D Gaussian Splatting pipeline to allow inference from images. Experiments on rigid and non-rigid collision datasets report better long-term accuracy and reduced error accumulation.
In briefExtending finite-sample Gaussian approximation of wide neural networks to functional convergence
SummaryRandomly initialised neural networks of width m converge to Gaussian processes as m grows. Existing finite-sample bounds give an m^{-1/2} rate in the quadratic Wasserstein distance with explicit constants, but these constants grow with the number of samples, making the bounds uninformative as the sample set is refined. This paper lifts the finite-sample bound to a functional one by measuring error per sample, making the constant independent of sample count. For polynomial activations and any compact input set, it obtains an m^{-1/2} rate in the supremum norm with an explicit constant.
Luis Wiederhold, Nil Ayday, Debarghya Ghoshdastidar
In briefA method using random probing so graph neural networks can capture node self-interaction structure
SummaryGraph neural networks struggle to capture structural properties defined by node self-interactions, such as cycle counts and subgraph centrality. The paper introduces DIAL, a message-passing layer that gives nodes access to graph structure via diagonal entries of graph-derived operators, using randomized probing for learnable, permutation-equivariant access. The authors prove DIAL can express cycle counts, subgraph centrality, and heat kernel signatures. Experiments on molecular property prediction, cycle counting, and pattern detection show its effectiveness.
Saku Peltonen, H. \c{C}a\u{g}r{\i} Bilgi, Kubilay Atasu
In briefLoOPSD: On-policy self-distillation using teachers built by looping layers
SummaryThis paper introduces LoOPSD, which keeps a model's weights fixed and searches over which layers to repeat to build a stronger teacher. The teacher's predictions on student-generated prefixes guide the student, and search and distillation alternate each round. On GSM8K with a 400-token thinking budget, Qwen3-1.7B accuracy rises from 42.30% to 79.61%, with further gains from later rounds.
In briefBenchmarking multimodal large language models for zero-shot human-object interaction recognition
SummaryThis paper presents the first systematic study of multimodal large language models for recognizing and detecting human-object interactions, using an external object detector with zero-shot multimodal reasoning. The authors introduce a benchmark framework with a taxonomy of zero-shot prompting strategies and a method that converts token logits into interaction scores for fair comparison, and they create new question-answer annotations for HICO, HICO-DET and V-COCO. They also propose two visual prompting strategies, Objects as Context and Black Other Objects. The method reports 53.50 mAP on HICO and 23.69 mAP on HICO-DET, state-of-the-art zero-shot results.
Oriol Rabasseda, Zenjie Li, Kamal Nasrollahi et al.
SummaryThis paper studies sycophancy in large language models, where models favor user preferences over facts. Using factual questions with verifiable answers, the authors expressed correct and incorrect user opinions in various framings and tested 14 models on 42K questions. All models were influenced by user opinions, most strongly by how confidently the opinion was stated, who it was attributed to, and whether it was framed as shared with the model. Expertise, authority, and social pressure mattered little.
In briefWhen and How Order-Flow Interventions Work After Dealers Adapt: A State-Aligned Order-Flow Contract (SOFA)
SummaryThe paper notes that learning dealer agents can keep quotes persistently wide, raising customer trading costs. The authors propose SOFA, a public contract with a scorer that ranks market states from lagged public data, a scheduler that announces intervention strength before quoting, and an allocator that shifts part of customer order flow toward tighter quotes. In an ABIDES market across 20 independently trained dealer populations, concentrating intervention in low-risk states lowers low-risk execution costs but raises high-risk costs, so it pays off only if the venue weighs low-risk costs more than 1.38 times high-risk costs.
In briefFan Out or Relay? Ownership-Aware Routing for LLM Agents Querying Siloed Tables
SummaryThis paper studies how an LLM coordinator should query data spread across several owners who keep their tables private: sending the question to all owners at once (fan-out) or querying them in turn and reusing returned rows (relay). The authors find neither strategy is safe in advance, since data ownership determines which works better; fan-out is usually more accurate when one owner can answer, but rarely answers questions joining two owners' tables. They propose Ownership-Aware Routing (OAR), which uses a schema-only draft SQL query to count the owners involved and route the question before any private row is read, and report it is rarely significantly worse than the best fixed protocol.
In briefWhen Do Learned Blocks Keep Their Causal Meaning? Block Identifiability and Constructive Causal Abstraction
SummaryThe paper studies whether multidimensional latent blocks in learned representations keep their causal meaning when their internal coordinates are not uniquely determined. It links block-identifiability to constructive causal abstraction and shows that block-diagonal invertible transformations followed by compatible coordinate permutations form the maximal class preserving all such abstractions. A small synthetic experiment with additive autoencoders shows that high-level interventions transfer between independently learned representations while approximately preserving their intended effects.
In briefLearning a Group Element from Its Action: Galois Correspondences and Optimal ERM
SummaryThe paper studies learning an unknown group element from examples of its action on a countable set. Through a Galois correspondence, multiclass learning dimensions are expressed as group invariants. Symmetry makes empirical risk minimization optimal among deterministic learners. For the Galois groups of the quintic, the Natarajan and DS dimensions differ, while they coincide for every solvable Galois group.
In briefTesting AI-Text Detectors by Attacking Both Human and Machine Text
SummaryThis paper examines how robust widely used AI text detectors are when attacks are applied to both human-written and machine-generated text, not just machine text as in prior work. The authors introduce a two-sided evaluation protocol across two corpora and nine detectors. Under this protocol, previously reported AUROC gains disappear, and attacked human text is misclassified at rates up to 100%.
In briefRadixFlow: Filling Idle Slots in Prefix-Cached Hybrid LLM Serving
SummaryThis paper introduces RadixFlow, a scheduling method for LLM inference that reuses cached prompt prefixes. In hybrid models, a request may need recurrent state still being computed by another request, and the native scheduler's blocking causes idle batch slots. RadixFlow separates prefix matching from state readiness so only the blocked request waits while independent requests fill the batch. On five hybrid models it reduces mean latency by 3% to 24% and worker rounds by 11% to 44%, with nearly unchanged reuse.
In briefFlowCred: Hierarchical Credit Assignment and Coordination for Joint Text-to-Video Training
SummaryThis paper presents FlowCred, a framework for jointly training the prompt policy and video generator in text-to-video generation. It uses matched prompt-seed rollouts to separate prompt-level credit from generator residuals, and probes alternative prompts along real generation trajectories to coordinate the two modules. On LTX-2.5, it raises a matched joint-training baseline from 61.47 to 64.46 on T2V-CompBench and from 55.58 to 58.08 on StoryEval.
In briefFixing Modality Misrouting to Reduce Cross-Modal Hallucinations in Omnimodal LLMs
SummaryOmnimodal large language models can hallucinate across modalities when audio and visual inputs interfere with each other. The authors observe that, for many wrong joint predictions, the question-relevant single-modality branch still yields the correct answer, a pattern they call modality misrouting. They propose CaRe, a training-free inference-time method that detects likely misrouting and steers the multimodal state toward the relevant modality evidence. Experiments on CMM and AVHBench with two models show CaRe outperforms the evaluated inference-time baselines.
In briefSparPR: Runtime KV-Cache Recovery for Running LLMs on Consumer Hardware
SummaryThis paper presents SparPR, a method that recovers the key-value cache during LLM inference without changing model weights or requiring offline training. It combines query-aware block selection, dynamic budget control, and on-demand local recomputation. Tested on a single consumer laptop, it reports 31%, 52%, and 63% lower end-to-end latency at 1k, 2k, and 4k prompt tokens, and completes an 8k-token workload where the FlexGen baseline times out.
In briefRevisiting Error Feedback for Model Quantization
SummaryThis paper studies causal error-feedback coefficients as design variables for neural network quantization, modeling quantization errors as independent, zero-mean variables with fixed variance. It formulates a metric-dependent trace problem and proves that the joint rotation-feedback trace minimum is n·det(H)^(1/n) for a positive-definite metric H, with an attaining construction. Experiments on synthetic networks, MLPs, and language models show gains from downstream-sensitive feedback, but trace optimality does not always improve finite-level performance.
In briefHuman-Inspired Asymmetric Collaboration for Dual-Student Semi-Supervised Medical Image Segmentation
SummaryThis paper proposes HAC, a closed-loop dual-student framework for semi-supervised medical image segmentation, where one student learns from weak augmentation and the other from strong augmentation. Reinforcement-guided Difficulty Calibration adjusts strong-augmentation magnitude, operation number, and range based on augmentation history and training responses. The Gramian Geometry Preservation Constraint uses global volume and pairwise area constraints to keep the early three-branch geometry and prevent the students from homogenizing. Evaluations on LA, Pancreas, and BraTS-2019 are reported to validate segmentation effectiveness and capability-aware difficulty calibration.
In briefMonitoring hidden reasoning in looped language models
SummaryLooped language models reuse one block of layers several times per token to add latent computation without extra storage. The authors build an inspection protocol from mechanistic interpretability, with five lenses, and apply it to eight looped models from four families. The main difference found is whether the prompt is carried only in the loop state or re-supplied at every iteration. The study reports insights on safety and efficient use of these models.
Pawel Olszowiec, Paulina Tomaszewska, Maciej \.Zelaszczyk et al.
In briefFresh at decision, stale at action: routing new information to LLM decisions already made
SummaryThe paper introduces a timestamp theory of decision validity and a decision ledger, arguing that an LLM decision's freshness is set when it is made while the action is carried out later in a changing world. Accuracy is said to factorise into routing of new information to affected decisions and temporal judgment of that information. Controlled decision streams across five LLMs, from 8B to frontier scale, are used to test four predictions.
In briefH-DDCM: Fast generative image and video compression using Hadamard codebooks
SummaryThis paper introduces H-DDCM, a training-free generative compression framework that replaces dense Gaussian codebooks with implicit randomized Hadamard dictionaries. A fast Walsh-Hadamard transform reduces computation and memory, and a table-free lexicographic coding scheme makes combinatorial encoding practical at large dimensions. The method is also extended to an autoregressive video flow model with streaming compression for near real-time long-video coding on one GPU. The authors report competitive rate-distortion-perception performance on image and video benchmarks.
In briefCheap, consistent, and calibrated, but biased? Typed decision models as judges
SummaryTyped decision models (TDMs) output a probability for each allowed answer rather than generating text, so they may replace LLM judges comparing two answers at lower cost. The authors ran the CALM benchmark's bias tests on eight TDMs and seven LLM judges, finding every TDM biased in three to eight of twelve tests, most strongly to emotional tone, where rewriting the better answer lowered a typical TDM's accuracy from 99% to 42%. TDMs were more biased than the 2026 LLM judges, yet both were affected by the same prompts. The authors suggest TDMs may be attractive judges if their biases are tested for the specific task.
Fabian Lukassen, Christoph Weisser, Alexander Silbersdorff
In briefPredicting Story Events in LLMs: Decodability, Advantage, and Steering
SummaryThis paper examines whether language models encode information about upcoming narrative events before generating them, and how far ahead this information extends. Using salient, state-changing events as the unit of analysis, the authors annotated 23,355 events from 2,000 stories produced by four open-weight LLMs, with human validation on a subset. Future event types were decodable from hidden states for at least four subsequent events, but probes significantly outperformed text-only predictors only for the next event. In exploratory experiments with OLMo-3.1-32B, a probe-derived direction shifted the proportion of the target event type in the intended direction at higher steering strengths.
In briefLearning and Verifying Almost-Sure Reachability Certificates in Unknown MDPs
SummaryThis paper asks whether a given policy can be verified against an almost-sure reachability specification in a Markov Decision Process with unknown transitions, using only sampling access. The correctness certificate is a ranking supermartingale (RSM). Without further information, the authors show that such a certificate can neither be learned nor verified with probabilistic guarantees. Under a mild restriction that a lower bound p_min on nonzero transition probabilities is known, a correct certificate can be learned with PAC-style guarantees. Experiments indicate that the p_min-based sufficient condition is sound but conservative, with learned certificates often remaining valid on the true MDP even when the condition fails.
In briefWhen missing features follow the graph: a new method beyond feature propagation
SummaryGraph Neural Networks usually assume complete node features, yet real-world data often has missing features. Existing graph methods fill missing values by propagating observed values along edges, but they are mostly tested under independent, uniformly random missingness. This paper formalizes graph-structured mask dependence, where missingness is shaped by the graph, and shows that propagation-based methods perform differently in this setting. The authors propose PEMix, which uses global structural representations to condition a probabilistic model of incomplete features before message passing, and report robust results, especially under severe missingness.
Enrico Chial\`a, Francesco Ferrini, Antonio Longa et al.
In briefSYNPRED: A Synergistic Multimodal Learning Method for Clinical Prediction
SummaryThis paper proposes a synergistic multimodal variational autoencoder framework for clinical prediction. It separates task-relevant unimodal and multimodal information from instance-specific, task-irrelevant variation, so complementary signals across modalities can be used for prediction. Evaluated on treatment response prediction with medical imaging and RNA data, it outperforms alignment-based and multimodal VAE state-of-the-art methods.
Ivana Jan\'i\v{c}kov\'a, Yen Y. Tan, Thomas H. Helbich et al.
In briefCaLLM Judge: Using Multicalibration to Reduce Bias When LLMs Grade Answers
SummaryLarge language models are increasingly used as judges to predict which of two answers a human would prefer, but their judgments can be systematically biased. This paper introduces CaLLM, a post hoc multicalibration framework that corrects miscalibration across many, including unspecified, bias groups at once. The authors prove bounds on group bias and derive conditions under which accuracy also improves. Experiments on pairwise preference benchmarks show lower multicalibration error and better predictive performance.
Cesare Barbera, Luca Stradiotti, Lorenzo Cascioli et al.
In briefChain-of-thought monitoring can fail even when the reasoning is necessary
SummaryThis paper studies how reliably chain-of-thought (CoT) monitoring can detect misbehaviour in AI systems. It argues that the earlier binary view, that CoT is monitorable when necessary, is too coarse, and proposes a four-level necessity hierarchy with intervention experiments to measure monitorability. On a hard medical-diagnosis task, patient sex and a clinical attribute influenced at least 19% and 24% of correct answers that relied on a necessary CoT, yet were almost never stated in the reasoning.
WhyProposes a four-level CoT necessity hierarchy with intervention experiments, showing monitoring failures even when CoT is necessary on a medical diagnosis task; significant implications for monitoring
Lennart B\"urger, Piotr Komorowski, Moritz Weckbecker et al.
In briefHow layer order in hybrid models affects task performance
SummaryThis paper analyzes how the choice and ordering of layer types in hybrid attention architectures affect model performance. The authors prove that for any layer configuration, there is a task where that configuration provably outperforms others under fixed resource constraints. Experiments show that the theoretically predicted optimal architecture performs better on a segmented retrieval task under the same memory budget, and that Transformer layers contribute most of the performance on associative recall.
In briefEvaluating How Accurately Deleting Model Components Removes One Skill
SummaryThis paper studies how well a first-order spectral ranking predicts the effect of deleting singular components from a language model to remove one capability. The authors test selective editing on Mistral-7B-Instruct-v0.3, targeting arithmetic while preserving code and question answering, with extensions to Llama-3.1-8B-Instruct. They find that interval coverage, retention of feasible candidates, and independent test preservation are distinct properties that must be assessed separately.
In briefAPT: A Self-Improving Automated Training Method for Text-to-Image Models
SummaryThis paper introduces APT (Agentic Post-Training), a fully automated framework that post-trains text-to-image models based on their evolving capabilities. Three agents build evaluation sets, diagnose strengths and failure modes, and generate reinforcement learning data that emphasizes underperforming skills. By repeatedly evaluating, diagnosing, and updating the generator, the model's own behavior shapes its next training tasks. Experiments reportedly show the approach works across multiple base models, benchmarks, and architectures.
In briefLifeSearch: A Benchmark Testing AI Research Agents on Everyday and Work Tasks
SummaryThis paper introduces LifeSearch, a benchmark of 203 Chinese and English tasks drawn from real user queries across 7 life and productivity domains. The authors built over 4,000 expert-written rubrics using more than 4,500 hours of human annotation and tested 12 deep research agents under a unified harness. The best agent, GPT-5.6 Sol, scored only 63.4%, and its ranking agreed strongly with human judgments.
In briefAM-Bench: How AI agents handle underspecified tasks across models and harnesses
SummaryThis paper introduces AM-Bench, a benchmark of 48 tasks, each with a fully specified version and six variants where goals, context, or constraints are ambiguous or missing. The authors evaluated five models paired with four execution harnesses over 20,160 runs. They found that missing goals hinder completion, missing context increases token use, and missing constraints weaken compliance, with missing information lowering success more than ambiguous information.
WhyNovel evaluation angle on underspecified agent tasks, backed by 20,160 runs across models and harnesses; useful for the subfield.
In briefA one-sided verification method to reduce expert review workload
SummaryThis paper addresses verification when human expert review capacity is limited, aiming to filter out invalid candidate solutions while keeping false rejections of valid ones low. It introduces OSVerify, a framework that separates error proposal from error validation. Experiments report that, compared with voting baselines under the same low false-rejection constraint, it reduces human reviews of invalid solutions by 46% on average, and by 57% after post-training the validator.
In briefResearchFlow-Eval: Measuring Whether Autonomous Research Agents Can Execute Multi-Claim Projects
SummaryThis paper introduces ResearchFlow-Eval, a framework for evaluating how well autoresearch systems carry out continuous, multi-stage scientific work across multiple claims. Each evaluation is based on a published anchor paper and uses stage-specific rubrics to score design, planning, and experimental execution. Across all tested systems, experimental execution lagged behind design and planning, and repeated attempts rarely produced measurable gains, while skill augmentation and research-specific orchestration did not reliably raise scores.
In briefSame function, different training path: standard optimizers are not invariant to Transformer reparameterization
SummaryThe paper shows that two Transformers computing the same function can follow different training trajectories when their parameters are rescaled, since gradient scales change. The authors propose Symmetry-Aware Gradient Geometry Stabilization (SAGGS), which periodically equalizes Frobenius norms of exact symmetry pairs, plus a damped variant, Soft SAGGS. Reported experiments on WikiText-103 and a checkpoint-resume setup show perplexity changes under scale imbalance and after rebalancing.
In briefAReaL-RSI: A Framework for Scaling Recursive Self-Improvement Through Coordinated Agent Teams
SummaryThis paper presents AReaL-RSI, a framework for organizing a persistent team of research agents without updating the weights of the backbone LLMs. A leader agent selects research orientations and guides investigations, while conditional memory sharing separates unresolved problems from supported findings. Experiments against Codex-Goal on harness optimization, GPU kernel optimization, and AI research tasks report performance gains and lower kernel latency.
In briefRecovering Long-Horizon Agent Failures via Belief Updates and World Modeling
SummaryLong-horizon agents often fail because the cause lies several steps earlier, which local reflection cannot fix. The paper proposes BELIEFWEAVER, which keeps evidence across attempts and reasons prospectively about repair directions, and distills these reasoning components into a 9B student model. Across three terminal-based benchmarks, the framework improves aggregate final score by 5.1 points, and the distilled student improves it by 9.8 points while reducing estimated teacher API cost by 54.8%.
In briefCausal discovery from truncated data using a robust score-matching method
SummaryThis paper addresses causal discovery when non-random selection leads to truncated (hard-selected) data, which can make observed data unrepresentative. For Additive Noise Models, the authors show that deterministic truncation preserves the joint score function and its Jacobian. They use this property to propose a new score-based causal discovery method that is robust to truncation and does not require information about the selection process. When selection is absent, it performs comparably to state-of-the-art approaches.
Yiwen Qiu, Francesco Montagna, Shimeng Huang et al.
In briefLensErase: A Foundation Model for Erasing Anything
SummaryThis paper presents LensErase, a mask-based general model for removing content from images, covering objects, visual text, watermarks, and related visual effects. The authors build LensErase-100M, a dataset of about 101.07 million source-mask-target examples, and pretrain the model from the text-to-image model Lens. The paper reports improvements over six public benchmarks and its own LensErase-Bench under both reference-based and reference-free evaluation.
In briefFrom Agents to Edges: Paired Optimization for Multi-Agent Collaboration
SummaryThis paper introduces PACO, a method that jointly optimizes the instructions at both ends of each communication edge between agents, along with their shared communication convention, in multi-agent systems built on black-box language models. Model weights, agent roles, and workflow topology stay fixed, and revisions are evaluated through staged end-to-end execution. On a four-agent workflow across diverse benchmarks, PACO consistently outperforms leading prompt-optimization and communication baselines.
In briefWorkRover: Scaling long-horizon cowork agent training through recursive environment evolution
SummaryThis paper introduces WorkRover, a framework for training long-horizon cowork agents by recursively evolving their training environments. Construction agents jointly build and validate work contexts, executable tasks, and verifiers, guided by execution feedback. It adds Factorial Compatibility Replay to select the smallest validated repair, and Verifier-to-Turn Credit Assignment to give finer training signals to prerequisite turns. The post-trained WorkRover-35B-A3B reportedly outperforms its base model Qwen3.5-35B-A3B on JobBench, OfficeQA Pro, and GDPval v2.
In briefUncertainty Evaluation: Bridging Machine Learning and Measurement Science
SummaryThis paper draws conceptual parallels between uncertainty evaluation in machine learning and measurement science, tracing metrology's shift from error-based methods to the GUM framework. It argues that the aleatoric/epistemic split for ML uncertainty is model-dependent and insufficient as a unified basis. The authors propose a metrology-inspired approach that systematically identifies, quantifies, and propagates uncertainty contributions to assess the trustworthiness of individual predictions.
In briefA Benchmark for Understanding Bangla Task-Oriented Dialogue
SummaryThis paper introduces BanglaTOD, a parallel benchmark in Bangla, English, and romanized Bangla (Banglish) for task-oriented dialogue. It contains 209 multi-turn dialogues across nine Bangladesh-specific service domains and 64 intents. The authors establish context-aware baselines for intent classification and slot filling using mBERT, BERT-base-uncased, and BanglaBERT. Removing dialogue history reduces intent accuracy by 19.0 to 25.7 percentage points. The dataset is publicly available on Hugging Face.
Afrina Alam Lamia, Sumiaya Sharif, Prottoy Modak et al.
In briefHow Transformers learn off-policy Q-learning in context, with proof
SummaryThis paper studies how Transformers can carry out Q-learning, a model-free control method, within their context. The authors prove that a two-layer attention-only Transformer can implement a finite-temperature form of Q-learning, which recovers window Q-learning as the temperature goes to zero. Trained on randomly generated MDP tasks, the model reaches performance comparable to Q-learning on unseen tasks.
In briefA paper proposing a more reliable bandit method that works without knowing reward scales, using empirical likelihood.
SummaryThe paper studies stochastic multi-armed bandits when reward scales and tail parameters are unknown. It introduces EL-F, an anytime empirical-likelihood extension of the g-Forcing policy, and provides asymptotic, expected-regret, and finite-time tail guarantees. Experiments show that EL-F reduces prolonged suboptimal allocation compared with g-Forcing on challenging instances.
In briefLADDER: A Long-Horizon Robot Manipulation Benchmark
SummaryThis paper introduces LADDER, a benchmark that tests robot manipulation policies on tasks scaling from 1 to 10 subgoals in the same environment, covering 145 tasks across three simulated suites. Low-level policies reach 85.4% on single-subgoal tasks but nearly zero on extended ones. Orchestration approaches improve results, with the tool-selection method GPT-Policy reaching about 62% at significant time and cost, and real-world experiments confirm these trends.
Petko Petkov, Giuliano Albanese, Nikolay Nikolov et al.
In briefCan AI Agents Forget on Request? A Benchmark for Memory Unlearning
SummaryThis paper introduces Bemula, the first benchmark for testing whether LLM agents can delete specific memories when a user asks. It includes 50 multi-session dialogues, 135 forget requests across six types, and 1,742 test questions, with paired control versions. Six mainstream memory systems were evaluated, and none passed more than 1.5% of forget requests end to end, indicating that memory unlearning remains an open challenge.
WhyFirst benchmark for memory unlearning, using paired controls and five test suites across six systems, revealing near-total failure of current agent memory.
In briefinLoRA-MINT: Auditing Training Data in Domain-Adapted LLMs Using Internal Representations
SummaryThis paper applies the Membership Inference Test (MINT) to large language models fine-tuned with LoRA, to determine whether specific samples were used in training. Extending the authors' earlier perplexity-based black-box method, it uses intermediate model representations such as activations so a classifier can separate training from non-training samples without full parameter access. Tested on four LLMs and three medical, mathematical, and academic reasoning datasets, the approach scores between 95% and 98%.
Gonzalo Mancera, Julian Fierrez, Aythami Morales et al.
In briefBeyond Black Boxes: Visual Grounding for Text Embeddings
SummaryThis paper proposes a sketch-based framework that converts text embeddings into visual topographic sketches, to make the encoded semantics visible and explore their use for explainability, retrieval, and reasoning. Evaluated on MS-COCO, SBERT is stronger in strict top-1 text retrieval accuracy, while both methods perform comparably in top-5. A cascaded combination of SBERT and the sketch-based method outperforms SBERT alone.
In briefReconstructing high-fidelity 3D turbulent flow fields from coarse data
SummaryThis paper proposes a framework that reconstructs time-resolved three-dimensional turbulent flow fields from coarse-grained data. It combines a multiscale neural operator with a state space model for temporal consistency, and uses Navier-Stokes residuals as soft constraints. A mixture-of-experts mechanism supports unsupervised adaptation to unknown coarse-graining. Tests on canonical turbulent systems and atmospheric reanalysis data report lower reconstruction errors than learning-based and classical baselines, with better preserved multiscale structure and temporal evolution.
In briefTraining-free token pruning and merging to make Vision Transformers cheaper
SummaryVision Transformers perform strongly but have high computational cost from quadratic token interactions. This paper proposes Prune-and-Merge ViT, a training-free framework for off-the-shelf models that combines class attention-based pruning with similarity-based merging to reduce redundancy. The authors introduce Relative Accuracy Drop and Prediction Flip Rate to measure robustness. On pretrained ViT and DeiT, the method reduces computation competitively while keeping robustness and accuracy better than existing token reduction methods.
In briefData-free continual learning for robot world action models via rectified generative replay
SummaryWorld action models built on pretrained video generators adapt quickly to new robot tasks but forget earlier ones under sequential fine-tuning. The proposed R²-WAM method has the model regenerate rollouts of past tasks for replay, truncating them at task completion with an agentic evaluator and filtering out rollouts whose actions are inconsistent with observations. On four LIBERO suites and five real-world tasks, it reduces forgetting by over 90% relative to sequential fine-tuning and nearly matches experience replay without accessing previous task data.
In briefWeighting Strategies for Imbalanced Classification: Theory and CART Tree Analysis
SummaryThis paper compares unweighted classification with weighting-based strategies for handling class imbalance in tabular data. It shows that, as predictor complexity grows with sample size, both approaches asymptotically yield identical performance across all metrics, differing only by a threshold shift. Theoretical analysis of depth-limited CART trees and numerical experiments identify synthetic settings where weighting performs better. The authors also introduce an indicator to help practitioners judge whether a dataset may benefit from weighting.
In briefCOMPASS: A Method for Generalizable Detection of AI-Generated Images via Contextual Prior Alignment
SummaryThis paper addresses the difficulty of detecting AI-generated images (AIGI) and deepfakes across different data sources. It proposes COMPASS, a data-aware subspace adaptation framework that uses a small unlabeled Semantic Proxy Set to estimate target activation statistics and incorporate them into the decomposition of pre-trained weights. The method preserves target-relevant representations while a residual adapter learns forgery-related variations, and the authors state that the proxy calibration requires no gradient-based optimization. The paper reports strong results across cross-dataset, cross-method, and open-world cross-generator evaluations.
In briefFlowSim: Rendering-Free Closed-Loop Simulation for Testing End-to-End Self-Driving Systems
SummaryFlowSim is a simulation method that evaluates end-to-end driving systems without rendering camera images. It predicts the internal latent representation directly from the previous latent and symbolic scene conditions, then feeds it to the original frozen downstream modules. Experiments on nuScenes and NavSim show it competes with, and often exceeds, image-space simulation in agreement with reference behavior, while each step takes 0.04 to 0.06 seconds.
Hanfeng Wu, Alvaro Marcos-Ramiro, Michael Schmidt et al.
In briefCan machine learning potentials be more accurate than the methods used to train them?
SummaryMachine learning potentials (MLPs) generalize across configurational and chemical spaces, but their accuracy is usually thought to be capped by their training methods. This Perspective argues that MLPs can exceed those methods non-empirically by generalizing across quantum-chemical methods. It proposes method-transferable MLPs that predict for unseen methods, with working implementations in density-functional and wavefunction-based theories.
In briefChampionKV: Adaptive per-head budget allocation for compressing LLM KV caches
SummaryThis paper proposes ChampionKV, a method that allocates KV cache budgets adaptively across attention heads for long-context inputs. It uses the maximum aggregated observation-window attention score as a proxy for attention concentration, giving more budget to dispersed heads. The method plugs into SnapKV and PyramidKV without changing their token selection. Results on LongBench and RULER show gains over base compressors, largest and significant on Llama3.1-70B-Instruct, while 7B/8B results are mixed and it does not uniformly beat Ada-KV.
In briefSHELDON: Scientific Reasoning via Dual Search over Hypotheses and Observations
SummaryThis paper introduces SHELDON, a reasoning framework that separately represents and jointly explores hypothesis and observation spaces. By actively seeking new observations alongside new hypotheses, it finds observations its current hypotheses fail to explain, guiding the generation of hypotheses beyond refining existing ones. The authors report that SHELDON outperforms evaluated baselines by up to 26.7 points on diagnosis, abstract reasoning, and optimization benchmarks, and generates up to 4.5x more novel solutions.
In briefFRACTION: an agent that navigates graph context to explain fraud risk labels in readable language
SummaryThe paper proposes FRACTION, a framework that generates human-readable fraud explanations by letting a language model use graph tools to progressively retrieve neighbor records. It introduces a task-specific rubric for explanation quality covering faithfulness, graph grounding, helpfulness, consistency, and conciseness, validated against human judgments. The agent is trained with rubric-guided trajectory distillation to learn tool use and structured explanation writing. Experiments on three fraud datasets report consistent gains over state-of-the-art baselines under the human-aligned rubric.
In briefStudy on whether AI language models can mark undergraduate assignments consistently with human markers
SummaryThe study evaluates three frontier large language models, alone and in ensemble, on 761 undergraduate psychology assignments from three UK universities. Agreement between human and machine marks varied widely, and no single model or prompt was optimal across institutions. All models compressed marks toward the middle, overvaluing weak work and undervaluing strong work, and they rewarded linguistic features more than humans did. The authors conclude that validation in a single context overstates readiness for deployment.
In briefDynamic Interpolation for Training-Time Prediction Reshaping in Low-Light Image Enhancement
SummaryThis paper addresses low-light image enhancement and proposes DITM, a training-only module that adds no learnable parameters. It interpolates the model prediction with a brightness-aligned image, reducing the dominance of global brightness error so residual local errors matter more in later training. Experiments on several benchmarks and CNN, Transformer, and Mamba backbones report better fidelity and perceptual quality with no added inference cost.
In briefLatentRemedy: An Efficient Fix for Reusing Non-Prefix KV Caches in Multi-Agent Systems
SummaryThis paper studies reusing non-prefix KV caches across agents in multi-agent systems, where directly reusing them can badly hurt performance when agents have conflicting tasks. It proposes LatentRemedy, which uses an MLP to turn the target prefix's last hidden state into a few task-conditioned latent tokens placed after the reused KV block. The MLP is trained with full-prefill distillation to guide predictions toward the target task. Experiments report that it outperforms existing reuse baselines on conflicting tasks while keeping time-to-first-token close to direct reuse as cache size grows.
In briefCorrecting or Continuing? How On-Policy Distillation Learns from Student Errors
SummaryThis paper studies on-policy distillation (OPD), where a student learns from its own rollouts under a frozen teacher. The authors train only on competition math problems that the student currently gets wrong, testing five OPD mechanisms with Qwen-3 and Phi-4 models. Accuracy on this hard subset did not improve and mostly declined, even as training KL divergence kept falling. Without concentrating on failures, the same pipeline improved accuracy by 4.6 percentage points.
In briefFish Audio Drama 3 Technical Report: An Open-Source Speech Model That Performs Under Direction
SummaryThis report introduces Fish Audio Drama 3, an open-source speech generation model that turns text into directed performances. Built on a 35B-parameter mixture-of-experts backbone with 3B active parameters per token, it clones voices from multiple references, follows natural-language direction, voices multi-character scenes, edits speech by infilling, and synthesizes long-form and streaming audio. Model weights are published on GitHub and Hugging Face.
In briefReconstructing surfaces from noisy depth scans while respecting what the sensor can see
SummaryThis paper addresses building 3D surfaces from range scans that contain noise and outliers. The authors frame reconstruction as robustly regressing the predicted first surface hit along each sensor ray against the measured depth, using a Huber loss and a 1-Lipschitz MLP representation with sphere tracing. Experiments on synthetic and real data show competitive results on clean data and much better fidelity and robustness under moderate corruption. The authors also report a trade-off between fitting capacity and robustness, with an intermediate capacity typically performing best..
In briefSWEPatchTest: A Benchmark for Judging Whether Generated Tests Really Validate Patches
SummaryThe paper introduces SWEPatchTest, a benchmark of 619 tasks and 82,579 candidate patches for evaluating whether tests generated by coding agents correctly accept good patches and reject bad ones. It measures issue reproduction, correct-patch acceptance, and incorrect-patch rejection jointly, using a Joint Validation Rate (JVR). The best configuration reproduces the reference transition on 59.29% of tasks but reaches only 9.21% JVR. The authors also use validation scores to select and scale demonstrations for supervised fine-tuning, which roughly triples JVR and improves downstream repair on three SWE-bench benchmarks.
In briefMatching the Generation Canvas: Predictability-Aware SFT for Diffusion Language Models
SummaryDiffusion language models generate tokens in parallel according to their own predictability rather than left to right, but standard supervised fine-tuning uses random masking that is unrelated to inference. The paper argues this training-inference mismatch can misdirect supervision. The proposed PACS method uses the model's predictability to choose each example's visible input length and supervises reference tokens just before confidence-guided reveal. PACS achieves the highest mean accuracy among compared SFT methods on reasoning and code-generation benchmarks, with gains up to 12.84 points over vanilla SFT.
In briefSaliency-guided replay to reduce forgetting in deep spiking neural networks
SummaryThis paper studies how to reduce catastrophic forgetting when deep spiking neural networks learn tasks in sequence, under low-latency constraints. The method combines Experience Packing and Replay with VGG9 networks stabilized by Batch Normalization Through Time and Spike Activation Maps, which turn spike activity into spatial attention maps. These maps select the most informative image patches to store for replay instead of full images. On Split CIFAR10 and Split CIFAR100, the approach improves the accuracy-memory trade-off under tight memory budgets and stays competitive while requiring much less storage.
Francesco Panetta, Federico Di Valerio, Michela Proietti
In briefSemi-supervised sugarcane leaf disease detection using uncertainty-aware learning and contrastive features
SummaryThis paper proposes U-SSDNet, a semi-supervised network for detecting sugarcane leaf diseases with fewer labeled images. It combines wavelet preprocessing, KL-divergence refinement of pseudo-labels, and supervised contrastive learning to improve feature separation. Experiments on sugarcane leaf disease datasets report performance close to fully supervised models and better results than existing supervised and semi-supervised methods.
In briefImproving neural symbolic regression with a diverse prior and post-training alignment
SummarySymbolic regression seeks a mathematical expression that fits a dataset, a search over combinatorial candidates. This paper presents RASR, which trains neural symbolic regression models on a synthetic prior closer to real noisy data, followed by reinforcement-learning post-training that rewards held-out fit and penalizes expression size. On the SRBench 2.0 black-box track, RASR achieves the best mean test R² rank among all evaluated methods.
In briefCertifying sequential machine unlearning against a counterfactual model with long-run performance bounds
SummaryThis paper studies sequential machine unlearning, where deletion requests arrive as groups of varying size. For each model release, the authors certify it against a counterfactual trajectory in which the deleted group never participated in training. Under a fixed fine-tuning budget per deletion and smooth strongly convex objectives, they show performance loss stays bounded regardless of the number of requests. Experiments on FEMNIST with multiclass logistic regression report fewer corrective steps than existing sequential methods and better results than training-time differential privacy, especially when group sizes are heterogeneous.
Hanna Benarroch (DI-ENS), Jamal Atif (CMAP), Olivier Capp\'e (DI-ENS)
In briefPolUR: Polarization-aware multimodal degradation perception for restoring underwater images
SummaryThis paper proposes PolUR, a framework that jointly uses polarization and RGB images to identify mixed underwater degradations such as color casts, haze, and low light, along with their severity. It uses degradation-aware semantic features to guide image restoration. The authors also build MixPolar, a benchmark with real-world degraded underwater images and matching degradation descriptions. Experiments report competitive results in degradation perception and restoration.
In briefDetecting Forgeries Is Not Enough: Models Still Abandon Correct Medical Answers Under Fake Evidence
SummaryThis paper studies how medical vision-language models react when a correct answer is challenged with a fabricated prescription. Testing four open models on 400 VQA-RAD yes/no questions across 19,200 multi-turn conversations, the authors found models abandoned 80.1% of correct answers when documents were not disclosed as synthetic. Some models rejected the forgery yet still changed answers later, and disclosing synthetic documents in the system prompt lowered one model's flip rate. The authors conclude that recognizing a forgery does not stop a model from adopting its answer.
Sadif Ahmed, Soumya Swagata Biswas, Rifat Shahriyar et al.
In briefAffordable Large-Scale Retrieval Evaluation with Tail-Aware Estimation
SummaryThis paper challenges the practice of enlarging image retrieval galleries to benchmark discrimination, showing that large-scale metrics can be estimated from a small random sample. It defines a per-query retrieval hardness that is independent of gallery size and estimates it with extreme value theory (EVT). Validated on ILIAS, INSTRE and Google Landmarks v2, a 5% sample converges close to full-scale mAP with a mean absolute error of about 0.005.
WhyAdapts sampled-rank machinery to visual retrieval, estimating large-scale mAP per query; validated across five backbones and three benchmarks with MAE 0.005.
In briefExact power indices for measuring each classifier's share in plurality-vote ensembles
SummaryThis paper studies how much each base classifier contributes to the final prediction of a plurality-voting ensemble, including its role in breaking ties at random. The authors adapt the Banzhaf and Shapley power indices from cooperative game theory and model plurality voting with random tie-breaking as new games. Their methods, named ENPOWER, compute these values exactly, in linear time for binary classification and in low-degree polynomial time for multiclass. The authors apply the attribution to ensemble pruning and prompt-based ensembling of vision-language models.
WhyApplies Banzhaf and Shapley power indices to plurality voting with exact linear-time algorithms, validated empirically and on vision-language ensembles.
In briefVision-language models sense depth but let image cues sway their judgments
SummaryThis paper studies whether vision-language models judge which object is closer using depth information or image-plane cues such as apparent size and vertical position. Through controlled experiments where depth and these cues conflict, plus representation analysis and interventions, the authors find that depth judgments systematically shift toward misleading cues. However, the correct depth order remains decodable from object representations, even when the model answers incorrectly.
In briefAll-electron machine-learned orbital-free density functional for periodic CHNO systems
SummaryThis paper introduces STRUCTURES-P, a machine-learned all-electron orbital-free density functional for periodic CHNO (carbon, hydrogen, nitrogen, oxygen) systems. Electrostatics are computed analytically in an atom-centered Gaussian basis, and an equivariant network learns kinetic and exchange-correlation energies from non-ground-state densities. On held-out crystals, the density error falls from 6.7% to 0.4% of the electron count, with an energy error of 0.25 mHa per atom.
Marc K. Ickler, Andreas Hermann, Fred A. Hamprecht
In briefMinimax rates for estimating semantic entropy from weak pairwise equivalence judgments
SummaryThis paper studies estimating the semantic entropy of a language model's sampled answers, where the underlying meanings are unobserved and only noisy pairwise equivalence verdicts are available through a known channel. The authors show that estimation difficulty depends on how many meanings can be rare, rather than on the global worst case, and derive the corresponding minimax rates. They also give an estimator that achieves these rates without knowing the rarity parameters, along with a consistency threshold.
WhyIntroduces a new framework where rare meaning count sets semantic entropy minimax rates, with matching lower bounds that sharpen the global rate picture.
In briefAlignPred: Compressing AI memory caches with controlled error
SummaryAlignPred is a codec that compresses key-value (KV) caches, the stored memory used by long-running AI agents, while keeping reconstruction error within a set budget. It aligns related attention heads or layers with local orthogonal transforms, then combines predictive coding with error-bounded quantization and entropy coding. The paper reports lower storage than the evaluated baselines at comparable error, decoding at least 1.6 times faster, and near-native task scores on LongBench-Pro at 128K context. It also reports 42-72% lower persistent storage in an SGLang HiCache integration, with comparable latency.
In briefNeural Dimensional Analysis: A Units-Based Framework for Scale-Consistent Transformers
SummaryThe paper introduces Neural Dimensional Analysis (NDA), a framework that applies dimensional consistency to neural networks by defining units for tensors and losses. It proposes dimensionless Unit Violation Indices for diagnosing scale mismatches and the Gauge-Canonical Transformer (GCT), which uses normalized coordinates and orthogonalized residual additions. In language-model experiments from 51M to 1B parameters, GCT trains stably across a wide learning-rate range and shows smaller validation-loss increases after int8 quantization than the baselines.
In briefdMRI-F: A foundation model for analyzing brain diffusion MRI from sparse scans
SummaryThis paper presents dMRI-F, a foundation model that improves diffusion MRI data and supports several downstream tasks using only a few input volumes. It is pre-trained to predict densely sampled multi-shell data from six volumes, learning general feature representations. The authors report that it outperforms existing methods across nine independently acquired datasets for tasks such as parcellation, registration, and cross-modal synthesis.
In briefBELT: Behavior-aligned low-rank KV cache compression with learned top-k rank allocation
SummaryThis paper proposes BELT, a low-rank compression framework for the key-value cache that large language models use during long-context inference. Its key-side objective targets attention distributions, yielding a quadratic surrogate with a closed-form low-rank solution, and a learned global rank allocator meets a user-specified budget. Evaluated on three LLMs and multiple benchmarks, BELT reportedly achieves up to 70% low-rank compression and up to 17.8x overall KV cache reduction when combined with quantization. The authors also report up to 3.3x end-to-end decoding throughput with Triton-based GPU kernels.
WhyProposes attention-distribution-targeted low-rank KV compression with learned rank allocation, validated across three LLMs and benchmarks with speedups.
In briefRevisiting negative sampling to better capture product interactions in retail basket choice modeling
SummaryThis paper proposes a shopping basket model that learns embeddings representing interactions among products, prices, and stores, to uncover sales patterns from large transaction datasets. To handle the computational cost of many items, the optimization loss uses random negative samples drawn from the actual available assortment. Experiments on synthetic data show this sampling approach outperforms methods from the literature. The model is also trained and evaluated on millions of transactions from Fortenova Group, a leading retailer in Southeast Europe, showing promise for retail applications.
Jules D\'esir, Vincent Auriau, Martin Mo\v{z}ina et al.
In briefCALM-5G: Event-Level Conformal Risk Control for Predictive Service-Assurance Alarms
SummaryThe paper presents CALM-5G, an assurance layer that sits on top of any forecasting model and issues alarms with distribution-free, finite-sample guarantees. It calibrates on service-level violation events, scores violation onset directly, and reports the alarm duty cycle alongside each certified miss rate. Evaluated on four separate corpora, it finds that online calibration restores compliance under label shift, while pooled batch calibration structurally exceeds the tolerance.
