In briefOrthogonal Boundary Unlearning for Removing Dual-Use Capabilities from LLMs
SummaryThis paper introduces Orthogonal Boundary Unlearning (OBU), a method for removing hazardous capabilities such as cyberwarfare exploitation and pathogen enhancement from large language models while preserving benign scientific ability. It extracts a capability subspace from calibration queries using singular value decomposition and limits parameter updates to the orthogonal null space. The authors report results on four open-weight models, showing better concept separation, reduced in-context priming attacks, and resistance to adversarial relearning.
In briefSingle vs. Multi-Agent LLMs on a Financial Review Task: Is More Better?
SummaryThis paper compares a single LLM agent checking six requirements in one call against a team of three agents on the same model, each checking two requirements, on a financial task of reviewing credit assessment memos. Using 300 synthetic memos with known labels, it measures misses and false alarms separately. The team catches somewhat more explained defects on some models but raises more false alarms and uses about 2.2 to 2.4 times the tokens.
In briefProtecting AI-Agent Payments with Behavior-Aware Fraud Detection
SummaryThis paper proposes ACAGS, a cybersecurity framework that combines machine learning, behavioral analytics, agent-specific authorization controls, and explainable AI to protect autonomous AI agents' payments. It was tested on synthetic transaction data and five adversarial scenarios, where behavioral and agent-specific features improved detection performance. The framework supports APPROVE, REVIEW, and REJECT outcomes. The authors note the results are simulated and require validation with real-world transactions.
In briefRuntime Control for AI Agents: Closing the Gap Between Approval and Execution
SummaryThe paper addresses a gap where an approval valid when issued may no longer permit execution after relevant facts, evidence, policy, or authority change. It proposes a state-bound runtime assurance layer that binds execution permission to the basis of an approval, invalidates it after logged changes, and rechecks admissibility at dispatch. In a scripted 42-case replay, the approach identified 15 later-invalidated approvals that a non-re-evaluating comparator missed. These results reflect behavior under stipulated policies and event orders, not superiority over dispatch-time enforcement or production effectiveness.
Jayesh Bhadane, Ramachandra Yerramsetti, Rajesh Gupta et al.
In briefFocusOVD: Refining Geometric and Contextual Features for Detecting Small Objects with Open Vocabulary
SummaryThe paper proposes FocusOVD, an open-vocabulary object detector that improves small-object detection without changing its frozen vision-language backbone. It adds three modules: reweighted cross-scale feature fusion, residual pinwheel-shaped gated convolution, and a residual convolutional gated linear unit. Reported results are 36.3% novel-category AP on COCO-OVD and 33.7% overall AP on LVIS-OVD.
In briefWhen Parameter Gradient Flows Close in Representation Space: A Theoretical Study of Closure Conditions for Gradient Flows of Parameterized Models
SummaryThe paper studies when one represented mobility map of a parameterized model yields the same gradient flow closure for every autonomous loss from a fixed initialization. For two-factor matrix representations it classifies all admissible joint ranks, and for general representations it gives local closure criteria. It also compares linear and quadratic loss families and leaves one smooth case unresolved. A separate deterministic experiment checks an engineered-loss construction.
In briefTraining-Free Continual Learning for Multimodal Forecasting via Prompt Evolution
SummaryThe paper proposes replacing retraining with an evolved text guidance block inserted into a frozen LLM forecaster's prompt, so it adapts to changing time-series regimes. Using a multimodal benchmark of search indices, currencies, commodities, and paired news, with controls against data leakage, a 900-word guidance carried by Llama-3.3-70B reportedly outperforms several time-series foundation models on held-out data. The authors also analyze the method's behavior, limitations, and two extensions, including compiling the guidance into executable code.
Laurent Mombaerts, Jonathan Taws, Jacopo Gargano et al.
In briefProjected Distillation: learning hard problems from close, correct targets
SummaryThe paper addresses post-training on hard problems where all student attempts fail, leaving on-policy reinforcement learning without a reward signal. It proposes projected distillation, which minimally edits the student's failed attempts into correct targets and refreshes these corrections as the student evolves, using contrastive updates. Experiments on Qwen3-8B and OLMo-2-7B across math, code, science, and tool-use domains report higher area under the specialization-retention frontier than GRPO, OPD, SPOT, and SODA in most settings.
WhyProposes minimal online corrections of failed student attempts with contrastive updates, backed by a regret theory and broad comparisons across models and domains.
Shripad V. Deshmukh, Christian Bock, Kangrui Ruan et al.
In briefRecoverability and Verifier Strength: Measured Preconditions for Unattended Agent Self-Improvement
SummaryThe paper proposes two measurable quantities for agent self-improvement: verifier strength, how closely an automated verifier agrees with a trusted reference, and recoverability, how cheaply a bad change can be undone. It argues a change should run unattended only when both are adequate, and tests this on the authors' platform and the τ²-bench benchmark. Results show an LLM judge sometimes returned identical scores regardless of input without reporting an error, and that rollback restored original behavior only when the old version had been kept.
In briefDetecting Errors in LLM Answers: Extra Value of Hidden States
SummaryThis paper studies how to detect hallucinations, or errors, in large language model responses. The authors build a combined baseline from output-side information available at inference time and compare detection performance with and without hidden states, using identical data splits and evaluation rules. Experiments cover open-ended generation and constrained multiple-choice tasks. Adding hidden states improves AUROC by about 0.004 and 0.039 on the two tasks, showing they add predictive information, though gains are limited when standalone detection is already strong.
In briefAsk to Distill: learning to generate training questions for on-policy distillation without answers
SummaryThe paper introduces Ask-to-Distill, a framework that trains a model to synthesize questions useful for on-policy distillation (OPD) without needing correct answers. It scores candidate questions with a distributional reward computed from a frozen student-teacher pair, combining teacher-student disagreement and teacher probability mass, then filters and deduplicates them into a prompt bank. Across math reasoning, code generation, and agentic tasks, the authors report improved average performance over unoptimized student prompts and KL-filtered real questions, matching or exceeding real-question OPD in the evaluated settings.
In briefFinding Pareto-optimal LLMs in multi-objective arena evaluations with fewer comparisons
SummaryThis paper introduces VB-EGE, a fixed-confidence adaptive sampling algorithm that identifies Pareto-frontier models in arena-style LLM evaluation using as few pairwise comparisons as possible. It progressively removes models from consideration once enough statistical evidence certifies their Pareto status. Under a multi-objective Bradley-Terry model, the authors show it recovers the correct frontier with high probability, with an instance-dependent sample complexity bound and a near-minimax-optimal worst-case lower bound. Simulations and experiments on three real Arena datasets demonstrate its effectiveness.
In briefWhether learned features suffice for safe language model control decisions
SummaryThis paper studies whether a controller that uses only selected internal features of a language model can achieve the same minimum average loss as one with full state information. The authors compute the optimal loss exactly for finite-output interfaces and identify conflicts that pairwise checks can miss. In audits on SmolLM2 with sparse autoencoder features, restricting the controller to two features raised its minimum average loss by 0.0345 to 0.1166 relative to full state information.
In briefLoRASpace: A shared-subspace method for combining many LoRA adapters, statically or dynamically
SummaryThe paper observes that stacked low-rank factors in real LoRA pools need far fewer directions than their summed ranks, so they can be compressed into a shared subspace. It proposes LoRASpace, a training-free method that computes this subspace once per pool and supports a static merge mode and a dynamic routing mode. On diffusion pools, static LoRASpace matches composite quality at merge cost, and dynamic LoRASpace is more than 6x faster than composite with six active LoRAs.
Thamirawaran Sathiyalogeswaran, Navindu De Silva, Chamikara Siriwardane et al.
In briefUsing Set Learning to Improve Set-Valued Classifiers on Imbalanced Data
SummarySet-valued classifiers predict variable-sized sets of possible classes to express uncertainty while keeping accuracy, but they perform poorly on long-tailed, imbalanced datasets. This paper studies how training strategies and loss functions affect these classifiers, and adapts the odd-K-out (OKO) loss, which learns from sets of samples, cutting its computational overhead from 3.2x to 1.36x of standard cross-entropy training. Experiments show the modified OKO approach significantly improves macro-accuracy across several set-valued methods while staying competitive on micro-accuracy, without substantially enlarging prediction sets.
In briefLearning-theory basis for general coded computing under adversarial workers
SummaryThis paper studies distributed computing where adversarial workers return corrupted outputs. It extends general coded computing with a smoothness-regularized encoder and decoder that limit the influence of corrupted results instead of locating and removing them. The authors derive a non-asymptotic bound on worst-case approximation error, which scales as O((q/N)^(6/5)) for q adversaries among N workers. Experiments on scalar computation and neural-network inference are reported to agree with the predicted convergence behavior.
Parsa Moradi, Hanzaleh Akbarinodehi, Mohammad Ali Maddah-Ali
In briefLearnable graph positional encodings to scale graph reasoning in language models
SummaryThe paper notes that serializing graphs for language reasoning models makes input length grow with the number of edges, limiting solvable problem size. The authors propose scalable, permutation-equivariant graph positional encodings computed by a GNN followed by a Graph Transformer, which modulate attention and remove the need for serialized adjacencies, reducing context to O(N). Trained on small graphs, the model generalizes in size to graphs with up to 1,000 nodes across graph-theoretic tasks and produces fewer tokens than a conventional model.
Javier Porras-Valenzuela, Arush Arora, Alejandro Ribeiro
In briefMLIPS: a training-free way to share sparse-attention candidate search across layers
SummaryThe paper introduces MLIPS, in which one producer layer publishes a top-M candidate pool larger than the attention budget K, and later layers each pick their own top-K from it, sharing the search without training. In vLLM for DeepSeek-V4-Flash with three consumer layers per producer, the authors report 1.14 to 1.40 times speedups and 14.8 to 27.4 percent lower GPU energy across four long-context suites, with accuracy changes of -1.49 to +0.29 points in single runs. An exploratory prefill method, Top-G, is also proposed.
Samkit Jain, Sai Kishan Pampana, Shlok Agarwal et al.
In briefSelf-Evolving Harnesses for Reliable Long-Horizon Terminal Agents
SummaryThe paper presents RiseHarness, a framework that iteratively improves an agent's harness while keeping model weights fixed. It focuses on tasks with both successful and failed runs under the same configuration, combining trajectories and runtime harness traces to analyze failures and search other frameworks for useful mechanisms. Candidate changes are checked behaviorally and evaluated repeatedly, with improvements carried into the next iteration. The authors evaluate task success and reliability across models and test transfer to unseen benchmarks.
In briefBaseline and Incremental: Debiasing Users, Items and Requests in Instagram Feed Ranking
SummaryThe paper argues that engagement predictions in feed rankers mix a baseline, what a viewer, item or request would produce anyway, with the incremental gain a candidate adds over it. The authors propose keeping the raw prediction and adding the incremental part as a separately weighted term. In a four-arm ablation on Instagram Feed covering viewer habit, item popularity and request context, the arm using both terms outperformed the single-term arms on most judged metrics.
WhyProposes a unified operator separating baseline from incremental value, validated by controlled four-arm ablation in production Instagram Feed with broad deployment impact.
In briefThe Router Already Knows: Sparse Autoencoders That Reuse Mixture-of-Experts Routing
SummaryThis paper asks whether a sparse dictionary can reuse the expert routing choice that mixture-of-experts models already make for each token. The proposed NR-SAE gives each routed expert its own small dictionary and activates only those the frozen router selects. At block 15 of Gemma 4 26B-A4B, it explains 0.518 of the variance in router-weighted expert outputs, outperforming several controls. The authors note that training speed, semantic interpretability, downstream loss, and causal faithfulness remain untested.
In briefTokenFlow: Latent Token Rectified Flow Matching for Generating 3D Point Clouds
SummaryTokenFlow is a hierarchical flow matching framework that represents each 3D shape as a set of latent tokens, combining a rectified flow over these tokens with a conditional flow over points. A small bottleneck of learnable tokens aggregates global context through cross-attention at O(NM) cost. The authors report the best MMD and coverage on ShapeNet airplane, car, and chair categories, and the best reported linear-evaluation accuracy on ModelNet10 using the learned latent set.
In briefKPHarness: A Framework That Co-Evolves Knowledge and Probing to Improve Game-Playing Agents
SummaryKPHarness adapts the external execution harness of a frozen vision-language-action model by co-evolving recovery knowledge and active probing. Probe outcomes test intervention hypotheses and refine both the knowledge and later probe choices. A slow loop proposes bounded harness revisions, accepted only when paired sandbox evaluation shows improvement. On Minecraft mining, combat, and crafting tasks, it improves task coverage over bare policy execution, a fixed harness, and knowledge-only evolution, and also improves completion on unseen tasks.
In briefLearning the geometry of materials space with trained neural tangent kernels
SummaryThe paper uses the neural tangent kernel evaluated at trained parameters to recover the geometry learned by a network that predicts superconducting transition temperatures from elemental composition. Kernel ridge regression with the trained kernel closely matches the network's performance and substantially outperforms the initial kernel. Spectral analysis and clustering show the geometry is dominated by a few directions and organizes compounds into recognizable superconducting families.
Daniel Kaplan, Gabriel Kotliar, Anirvan M Sengupta
In briefTest-time adaptation for point cloud denoising under unseen noise
SummaryDeep learning point cloud denoisers often perform poorly when test noise differs from training noise. This paper proposes TTAD, a test-time adaptive denoising framework that requires no extra supervision or source-data retraining. It estimates each sample's noise distribution from the residual between the noisy input and its initially denoised version, then updates the model via contrastive learning over a few adaptation steps. Experiments on synthetic and real-world datasets report improved generalization to unseen noise.
In briefLooking Back: How Context Affects Detecting Deceptive Statements
SummaryThis paper studies detecting deceptive messages in the hidden role game Avalon. Prior methods usually judge each message in isolation. The authors show that adding preceding gameplay dialogue as context improves several metrics for large language models (LLMs).
Alex Clay, Ernesto Jimenez-Ruiz, Pranava Madhyastha
In briefMore Memory, Worse Agents: Error Reproduction in LLM Agents
SummaryThis paper studies LLM agents that store their own experience in external memory to improve over tasks without updating model parameters. It argues that storing inferred rules and skills can cause errors to be regenerated through a cycle of abstraction, execution, and re-abstraction. The authors propose Anti-Persistence, which keeps only trajectories and outcomes and synthesizes task-specific guidance on demand, discarding it after use. Across benchmarks and backbones, it reports higher held-out success and the lowest average LLM-call cost among the memory designs evaluated.
In briefControlling intonation contours in text-to-speech synthesis
SummaryThis paper examines which components of residual vector quantization text-to-speech models are responsible for English nuclear intonation, such as falling tunes for statements and rising tunes for questions. The authors find that the transformer backbone's hidden state carries intonation information, and steering it with vectors can shift the tune toward falling or rising. However, precise selection of a specific target tune is not achieved.
In briefMinimum Bits Needed to Encode Data for Unknown Tasks: Communication Complexity of Universal Features
SummaryThe paper studies how many bits are needed to deterministically encode data so that, after a task function is later specified, a decoder can predict it with error at most ε. The authors build an encoder using O(d*log(1/ε)) bits, where d* is the dual VC dimension of the function family, by greedily selecting functions and encoding their joint label patterns. They prove this length is worst-case optimal up to a constant factor for all d*≥1 and ε<1/4, and near-optimal for each instance.
WhyProposes a universal feature encoder with near-optimal bounds via a novel potential argument, but experiments are limited and impact centers on theory and representation learning.
In briefProMemAgent: A Proactive AI Assistant That Remembers Past Sensory Events to Help Later
SummaryThis paper presents ProMemAgent, the first proactive AI agent to use prospective memory over everyday sensory contexts, such as egocentric video and audio, for long-horizon assistance. It organizes past perceptions into context-aware memory, generates prospective cues, and links earlier events to later developments for retrieval. The authors also introduce CPMBench, a benchmark of 1,000 instances across eight everyday categories. Experiments report that ProMemAgent achieves the best memory retrieval and consistently outperforms all baselines.
In briefProof of Unified Sharp Bounds on Probabilities of Causation
SummaryThis paper is a theoretical study of probabilities of causation (PoCs). It proves the sharpness in high dimensions of closed-form bounds proposed by Shu et al. for multi-valued treatments and outcomes. The authors also show that PNS, PSub, PRep, and PN are special cases of one more general formula. Together with equivalence classes, they conclude that sharp bounds for discrete PoCs are given in closed form by a single theorem in every dimension.
In briefWhen Calibration Helps or Hurts: Overdosed Updates and Validation-Gated Skipping in Quantized Embedders
SummaryThis paper studies learned post-training quantization calibration for text embedding models quantized to 4-bit weights, across 161 runs on three embedders. The outcome depends mainly on the size of the update, measured as functional displacement from the uncalibrated model, and excessive updates cause severe degradation even with label-free objectives. The authors propose a validation-gated rule that accepts a checkpoint only when it beats the uncalibrated baseline on a held-out split, which avoided harmful selections but also skipped some beneficial updates.
In briefFixing Attribution Patching Errors Using a Few Exact Measurements
SummaryThis paper addresses errors in attribution patching, a method that estimates which language model components matter using one gradient pass. The authors propose Guided-AP, which uses a small number of exact measurements to reveal the error pattern and correct unmeasured estimates. On a single prompt, it reduces error by up to 26%, and over many prompts on factual completion it cuts the error of average effects by 75% to 79%.
In briefFixed-size sampling for variance reduction: a paper proposing conditional Poisson sampling as a drop-in replacement for random-size Poisson sampling
SummaryThis paper studies variance reduction in data analysis by combining machine learning predictions with weighted sampling. It notes that Poisson Sampling has a random sample size, which adds variance, and proposes Conditional Poisson Sampling, a fixed-size alternative, as a drop-in replacement for general multivariate Z-estimation. The authors state its asymptotic variance is no worse than Poisson Sampling and can be substantially better when predictions are imperfect. They validate it on synthetic and real datasets and release a Python package implementing it.
In briefTraining world models to respond to actions by maximizing action influence
SummaryAction-conditioned video models trained on successful expert demonstrations often ignore the commanded action and generate plausible expert-like futures. The authors attribute this to training data with insufficient action influence and prove that this metric upper-bounds the action influence any model trained on that data can express. They introduce Action Influence Maximization (AIM), an algorithm that actively collects training data to raise dataset action influence. In pixel-based manipulation, AIM is the only compared method that consistently improves controllability.
Samuel M. Bateman, Tenny Yin, Chongyi Zheng et al.
In briefLearning from few samples with Dense Associative Memories
SummaryThis paper analytically studies how the energy function of Dense Associative Memories changes while training on random codewords of sparse codes. It identifies two regimes, one where the model memorizes examples and one where it uncovers the generating constraints, with a transition between them. The authors also propose a thresholding learning rule that lets the model discover exponentially many solutions from only logarithmically many training samples.`
In briefDStill: Trajectory Consistency Distillation for Faster One-Step Drafting in Block Diffusion Speculative Decoding
SummaryThis paper addresses a limitation of one-step drafting in block diffusion speculative decoding, where each token is predicted without conditioning on other tokens in the same step, which can make drafts inconsistent. The authors propose DStill, a two-stage method that first trains a drafter to complete partially revealed blocks, then distills its multi-step refinement trajectories into a student that predicts the final completion in one pass. The paper reports that across reasoning, math, and code benchmarks on three model families, DStill improves acceptance length by up to 28% and speedup by up to 25% over one-step DFlash, reaching up to 6.45x speedup over autoregressive decoding.
Shreyas Kulkarni, Uttam Gurram, Madhav Kanda et al.
In briefREM: Deciding What Coding Agents Store, Recall, and Forget
SummaryThis paper introduces REM (Reusable Experience Memory), a framework that lets coding agents manage their own experience over its lifecycle: what to store, when to recall it, and when to forget it. On a repository-level bug-fixing benchmark (SWE-ContextBench Lite), REM retrieves related trajectories more accurately than prior memory systems and improves end-task resolution at 17% lower cost. The authors conclude that the bottleneck in experience reuse is how experience is stored and governed, not retrieval fidelity.
WhyProposes in-loop lifecycle management of agent memory, outperforming prior systems on repository bug fixing with ablations; impact mainly confined to coding agents.
Xinyi Lisa Chen, Pedro Herrero-Vidal, Xiaohan Chen et al.
In briefShaped SAEs: Improving Sparse Autoencoders for Understanding Language Models
SummarySparse autoencoders (SAEs) break the dense activations of large language models into sparse features that may be easier to interpret. This paper proposes shaping the activations linearly before SAE training, using rank-reducing transformations based on the Jacobian's J-space, and introduces a Matched Filter SAE (MF-SAE) that learns an affine shaping transformation along with the dictionary. The approach is evaluated against dense activations and state-of-the-art SAEs on SAEBench and SAE Probes.
In briefPerQ: Inverse Generative Modeling for Neural Image Compression via Quantization Error Compensation
SummaryPerQ is a neural image compression method that balances pixel accuracy and perceptual quality. It performs generative reconstruction within the quantization cell that the codec defines, rather than sampling from an unconstrained image prior. It attaches a rate-conditioned flow-matching compensator to a frozen, multi-rate distortion-optimized backbone. Experiments on Kodak and CLIC2020 report perceptual scores (LPIPS, FID) comparable to generative codecs such as MS-ILLM, with PSNR competitive with distortion-optimized codecs such as ELIC.
In briefFundamental Impossibility Results in AI Safety: Trading Off Between Desiderata
SummaryThis paper asks whether there is a fundamental tradeoff between goals for safe AI, such as harmlessness and helpfulness. The authors model interactions between AI systems and users as a three-player game involving a bounded defender, a bounded attacker, and a less constrained AI that gives public advice. The main result is that advice which helps the defender while remaining public involves a fundamental tradeoff. The paper also examines whether shared AI information can be manipulated to benefit only the defender, and discusses how these results might inform safer AI design.
Ida Mattsson, Ratip Emin Berker, Emanuel Tewolde et al.
In briefSHAM-Bench: A Multi-Task Benchmark for Span-Level Hallucination Detection
SummaryThis paper introduces SHAM-Bench, a unified benchmark for detecting unsupported text spans in LLM outputs across six NLG tasks, built from existing human-annotated datasets totaling 6,461 examples. Evaluations show that fine-tuned detectors often perform well in-domain but degrade under distribution shift, while reasoning-based prompt methods are more robust across tasks. The authors note that subtle unsupported content and ambiguous span boundaries with noisy labels remain challenging. The dataset, evaluation code, and a public leaderboard are released.
In briefExplore More, Validate Less: An Efficient Way to Refine AI Prompts
SummaryThis paper proposes a two-stage framework for reflective prompt optimization that separates building a pool of candidate prompts from allocating validation effort. Candidates are screened on training questions and only improving, nonduplicate prompts enter the pool before a fixed validation budget selects the final prompt. Tested on five benchmarks with Qwen3.5-9B and Llama-3.1-8B-Instruct, the method improves over the initial solver by up to 11.22 percentage points.
In briefTrail-weighted scores for pruning evidence in Bayesian network prediction
SummaryThe paper proposes pruning evidence in Bayesian networks by scoring each active trail from evidence to outcome with the product of its edge strengths and dropping evidence below a threshold. For collider-free binary networks, it proves an error bound on the pruned posterior relative to the exact one, under the condition that dropped trails lie in separate branches of the outcome node. On random trees, threshold pruning yields 1.4 to 13 times lower error than random pruning of equal size. On seven bnlearn networks with colliders and loops, where the method is used heuristically, it beats random pruning but not mutual-information ranking, and it brings no speedup.
In briefPredictive Closure: How JEPA Chooses Which Features to Keep
SummaryJoint-embedding predictive architectures (JEPA) train an encoder and predictor together so that representations minimize the error in predicting the next observation's representation. This paper proposes a "predictive closure" characterization, arguing that feature selection depends on properties of feature sets rather than only on each feature's individual predictability. Controlled experiments show that hard-to-predict features, such as a traffic light, can be better represented when they help predict other features. The authors also show that steering one feature via an auxiliary objective or side information can indirectly strengthen another feature's representation.
In briefVerifying LLM-generated argument graphs with layered checks
SummaryThe paper examines how argument graphs generated by multi-agent LLM systems should be verified, noting that current quality checks mostly rely on other LLM reviewers. The authors propose layered verification, starting with deterministic checks, then solver-based and LLM reviews, using debate only under uncertainty, and leaving final decisions to humans. In a study with human-annotated graphs containing injected structural defects, an LLM reviewer detected nearly all defects, including support-relation cycles, but wrongly objected to about one in ten correct graphs, gave inconsistent verdicts on identical inputs, and ran much slower than an exact check. The authors conclude that deterministic verification gates are needed alongside LLM review.
In briefLimits of Scale: Larger Vision-Language Models Can Perform Worse on Some Tasks
SummaryThis paper examines whether bigger vision-language models always perform better, and finds that within the same model family, larger models sometimes underperform smaller ones on specific subtasks, a pattern called inverse scaling. The authors evaluated 19 models across eight multimodal benchmarks and identified 375 subtask-model-pair combinations showing this effect. They conclude that increasing model size does not uniformly improve multimodal capabilities, so model selection should be task-aware.
In briefLiBRA: Logit fingerprints for Byzantine-robust client selection in federated learning
SummaryByzantine-robust aggregation rules in federated learning filter suspicious client updates by comparing parameter vectors, which requires access to every client's update, conflicts with secure aggregation, scales with parameter dimension, and only works between clients sharing an architecture. This paper proposes LiBRA, which compares clients using a fingerprint of class-conditional logit responses to optimised pseudo-patterns, and instantiates Krum, Multi-Krum and Bulyan as variants using this fingerprint. Experiments on Fashion-MNIST and CIFAR-10 show the variants retain the robustness of their parent rules against four backdoor and label-manipulation attacks.
In briefA Small Model Asks Better: Tuning a Model to Clarify Queries for RAG
SummaryThis paper studies how retrieval-augmented generation systems should handle ambiguous user queries and generate clarification requests. The authors introduce ClariEdge, a dataset of about 3,900 examples that includes ambiguous queries and negative examples that need no clarification. A lightweight LoRA fine-tune reaches 96.6% classification accuracy and outperforms prompted larger models on classification and clarification quality. The authors find hard negatives are key, and the model works both before and after retrieval in the pipeline.
In briefLightweight IoT device identification that stays accurate despite network traffic anomalies
SummaryThis paper proposes a lightweight method for identifying IoT devices using packet headers and locality-sensitive hashing (LSH), producing compact fingerprints at packet and window levels. The authors test robustness against packet drop, perturbation, reordering, duplication, and injection. Experiments report up to 99% accuracy with short bounded observation windows. Anomalies affecting global traffic structure cause the largest performance drop, while window-level representations reduce sensitivity to localized distortions.
In briefRefuses When Asked, Complies When Acting: Runtime Safety Check for Robot Loops
SummaryThe paper shows that frontier models controlling a low-cost robot arm through an agent harness refuse harmful instructions when asked directly, yet their default control loop carries them out. In 15 trials across three household tasks, the policy refused 0% of the time, attempted the harm every time, and completed it 27% of the time. Adding a runtime check, which asks the model before the first motion whether it would carry out the instruction given the same camera frames, would have stopped all 15 hazardous instructions and flagged none of the harmless controls.
In briefAn LLM-Based Agent for Checking Legal Compliance in Canadian and Quebec Contexts
SummaryThis paper presents ComplianceAgent, a legal compliance sub-agent within the COMPASS multi-agent framework, which assesses texts on digital sovereignty, carbon impact, compliance, and ethics. A large language model scores texts on a fixed rubric against a reference legal corpus, and low-scoring texts trigger corrective steps such as self-healing rewrites and escalation to a human reviewer. Evaluated on thirteen Canadian, Quebec, international, and municipal legal instruments with a fifteen-text gold standard, the agent perfectly separated conforming from non-conforming texts, kept scores stable under repeated queries, and retrieved relevant passages.
In briefLook Before You Reason: An Implicit Visual Thinking Method for Efficient Multimodal Reasoning
SummaryVision-language models often struggle with small or easily overlooked image regions, and existing thinking-with-images methods need multiple turns of interaction, which adds redundant computation and slows inference. This paper proposes LoRe, which predicts the most relevant image region within a single VLM interaction and strengthens the corresponding KV cache entries so decoding can use those visual cues without extra inputs. The authors also use reinforcement learning to guide region selection and introduce COCO-ObjQA, a benchmark for object-centric multimodal reasoning.
In briefAutomatically Completing Brick Building Knowledge Graphs Across Buildings
SummaryThe Brick schema describes building equipment, sensors, spaces, and their relationships as typed graphs, but building these models by hand requires much manual reading of schematics. This paper applies knowledge graph completion to Brick model generation for the first time, comparing gradient-boosted trees, Pairwise Brick Relation Ranking, and CompGCN on the feeds, hasPart, and hasPoint relations. CompGCN performed best on held-out MORTAR buildings, but no method generalized uniformly across datasets and relations.
In briefModeling real research paths: a graph view of how research ideas evolve
SummaryThe paper proposes the Research Progress Graph (RPG), a representation that records how a research idea evolves over time as typed research states linked by cognitive transitions. Each transition is annotated with a temporal index and the reasoning behind it. Comparing thirty-five research agents, the authors find that each is an incomplete RPG, since none keeps the changed state, the evidence behind the change, and the connecting reasoning together.
In briefEMBER: Predicting T-cell receptor and antigen binding without a docked complex
SummaryEMBER is a structure-based framework that predicts binding between T-cell receptors (TCRs) and peptide-MHC complexes (pMHCs) without requiring a predicted docked complex. It folds the TCR and pMHC independently, encodes each as an SE(3)-invariant residue graph, and compares their learned representations, so relative pose is not an input and placement error cannot propagate. Under epitope-disjoint evaluation, EMBER ranks first on every metric across four folding tools, outperforming six retrained baselines.
WhyAvoids co-folded complexes by encoding TCR and pMHC independently, a novel design. Consistent gains across four folding tools and six baselines give solid evidence, with moderate impact on TCR-pMHC pr
In briefLabel-free evolution of reusable tools for coding agents in data-lake analytics
SummaryCoding agents answer questions over data lakes by finding, cleaning, and joining data, but they discard this work after each question. The paper proposes TEAL, which evolves a library of tools and usage skills for a Worker agent without labels, using an Explorer and a Critic to propose changes that are kept only if validated on rerun. It also introduces TEAL-Q, which generates synthetic questions when real ones are unavailable. Experiments on KramaBench and LakeQA report improvements under tight step budgets, and synthetic-question evolution performs comparably to real-question evolution.
In briefSteering vision-language models to disclose location only at context-appropriate detail
SummaryThis paper studies how vision-language models can reveal a location from image cues, sometimes disclosing more detail than the picture warrants. The authors propose an inference-time framework that uses an ordinal granularity router to choose a disclosure level, then applies level-specific hidden-state steering during decoding. Across four model backbones, the method improves alignment with visual context and balances over- and under-disclosure, with modest changes on general capability benchmarks.
In briefVampPrior Unleashed: Decoupling Prior Covariances and Using a Dirichlet Process
SummaryThis paper improves the VampPrior, a prior for deep latent variable models that approximates the aggregate posterior. The authors decouple the prior and posterior covariances and replace the fixed uniform mixture weights with a Dirichlet process, which they report raises model evidence and enables robust non-parametric clustering. Applied to single-cell Variational Inference (scVI), the modified prior improves single-cell integration and clustering across six atlases. The change affects only the prior, so it can be used with any model with a continuous latent space.
Andrew A. Stirn, Abhejit Rajagopal, David A. Knowles et al.
In briefPhysics-inspired optimizers beyond the edge of stability: period-two orbits and delayed instability
SummaryThe paper studies how physics-inspired optimizers such as VRAdam behave when the learning rate is large, past the edge of stability. The authors show that certain kinetic maps create a stable period-two orbit beyond the edge, and use slowly swept bifurcation theory to explain why loss of stability is delayed. Experiments on an exact quadratic problem, a full-batch MLP and ResNet-32 support the theory. On ResNet-18, however, much of the apparent benefit is reproduced by matching the effective learning rate.
WhyAnalyzes kinetic maps beyond edge of stability via period-two orbits and Neishtadt's delayed-bifurcation theory, with experiments on quadratic, MLP and ResNet models.
In briefMachine-learned spectral PDE solvers for fluids and plasmas
SummaryThis paper introduces WARP, a method that discovers governing PDEs in weak form within the spatial discretization used for solving them, learning by differentiating through multistep rollouts anchored at observed states to produce a machine-learned solver. Tested on eight fluid and plasma benchmark systems, including high-dimensional turbulent simulations, WARP yields numerically stable rollouts from new random initial conditions. It substantially improves coefficient recovery, rollout accuracy, and statistical measures over a single-step weak-form SINDy baseline.
In briefWhat Not to Erase: Saliency-Constrained Data Augmentation for Low-Resource Handwriting Recognition
SummaryThis paper studies data augmentation for low-resource historical handwritten character recognition, using saliency maps to constrain which image regions are perturbed or preserved. It compares classical document-style transformations, random lacunae, and several saliency-constrained variants. The authors introduce the Explainability Consistency Index (ECI) to measure how stable saliency patterns remain under perturbation. The results show saliency-constrained augmentation can sometimes beat classical augmentation, but no single strategy is consistently best, and the preferred constraint varies by architecture and corpus.
In briefSelf-Tuning Benchmarks That Evolve for Long-Horizon AI Tool Use
SummaryThis paper introduces TURTLE, a method that automatically generates and evolves test tasks for evaluating AI assistants' tool-use abilities, reducing reliance on human supervision. Three language model agents act as a proposer, solver, and feedback generator, with the proposer rewarded for tasks that are difficult yet solvable. Experiments on AppWorld, τ-Bench, and AutomationBench show the evolved tasks reach the target difficulty range, are more diverse, and reveal a broader set of failure modes than the original tasks.
Hyunji Alex Nam, Michael Dennis, Natasha Jaques et al.
In briefSweepANN: GPU-accelerated large-K approximate vector search for drug discovery
SummaryThis paper addresses large-scale vector search in drug discovery, where each protein pocket needs 500K candidate molecules (0.1% of a 500M-molecule library) and missed hits must be avoided. Existing approximate nearest-neighbor methods struggle to keep recall high at this scale while staying fast. The authors propose SweepANN, which partitions compressed vectors across GPUs, scores query batches with matrix multiplication, and uses a window plan to select candidates. The paper reports answering 26,562 pockets in about 61 seconds on four GPUs at 99.6% mean recall.
In briefWhen an added planning capability creates real operational value: an observable evidence standard
SummaryThis paper addresses the attribution problem in industrial AI planning systems, where overall policy improvement cannot show whether an individual added component made an independent contribution. It proposes an observable evidence standard that separates opportunity to act, executed-plan change, operational consequence, equivalence, uncertainty, and operating scope. Forecast-to-plan replay experiments on Favorita, M5, and Walmart data show the limits of single-indicator evaluation, such as one control changing only 15 of 681 executed plans.
In briefEvaluating Context Compaction on Later Questions
SummaryThis paper evaluates how well language-model agents' context compaction serves questions asked after the compaction, using four corpora annotated with evidence for several questions per text, plus recorded agent trajectories. Each method is compared with random deletion at the same token budget. Sentence selection for the current question does not detectably beat random deletion on later questions, while LLMLingua-2, which deletes words without considering the question, performs better on some datasets, though the ranking depends on the setting.
In briefGeneralizing Multimodal Foundation Models for Continual Audio-Visual Class Learning
SummaryThis paper studies class-incremental learning, where a model learns new classes over time without forgetting old ones, in the audio-visual setting. The authors find that frozen multimodal foundation models struggle in incremental settings. They propose an attention-based module that uses audio representations as stable context to guide visual recognition, plus feature- and logit-level distillation to reduce forgetting. The paper reports consistent gains over state-of-the-art methods across several audio-visual benchmarks.
In briefRealControlBench: A Benchmark for Fine-Grained Control of LLMs and Agents
SummaryThis paper introduces RealControlBench, a benchmark for evaluating fine-grained steering of large language models and LLM agents. It extends evaluation beyond isolated generation tasks to complex, multi-turn agent interactions, covering five dimensions including instruction following, tool-use decisions, reasoning performance, reasoning length, and safety awareness. The authors report that existing steering methods are unreliable under realistic objectives, often failing to induce the intended behavior and sometimes substantially degrading task performance.
In briefVulnerability of safe multi-agent RL to inference-time observation attacks
SummaryThis paper studies how safe multi-agent reinforcement learning (SMARL) systems can be vulnerable to observation attacks at inference time, where an adversary perturbs some agents' observations to change their actions. The authors propose MinMax Attack, which selects adversarial target actions for compromised agents while accounting for benign agents' cooperative resistance. Experiments on multiple safety-critical benchmarks show it induces higher safety costs than existing baselines while preserving task reward, making it harder to detect.
In briefTEX: Testing AI agents by having them run each other's tests and patches
SummaryThis paper presents TEX, a framework in which multiple agents generate tests and patches in parallel and cross-execute every test against every patch, forming a K×K feedback matrix. Agents refine their work using their slices of this matrix, allowing objective validation without sharing lengthy traces. On SWT-Bench Verified, TEX reaches 86.4%, reported as a new state of the art and 9.0 points above single-agent baselines. The authors also note that random selection from TEX outputs (84.4%) exceeds the previous best (84.0%).
In briefWhen Hate Speech Detection Transfers Across Languages: A 14-Language Study
SummaryThis paper examines whether hate-speech detection trained on one language can perform comparably on another, across 14 languages including low-resource ones. Linguistic similarity is the strongest predictor of transfer success, while temporal overlap and dataset-size differences play smaller roles. The authors add a qualitative hypothesis that shared cultural cues and communicative norms may relate to some remaining cases, but state this is not a measured cause.
In briefVPFlow: Value-Predictive Planning with Imperfect World Models for Offline Goal Reaching
SummaryThis paper studies planning for offline goal-conditioned reinforcement learning and argues that predicting value is more useful than predicting exact next states. The authors train latent world models using a closed-form Monte Carlo return as the only supervision, and present VPFlow, which plans in the noise space of a one-step flow policy via evolutionary search. On 30 OGBench state and pixel datasets, VPFlow reports higher success rates than model-free baselines and latent planners such as LeWorldModel, PLDM, and TD-MPC2.
WhyReplaces TD bootstrapping with closed-form Monte Carlo returns for latent world models, a novel planning-oriented approach; evaluated on 30 OGBench datasets against several baselines, giving fairly so
In briefA latent control model that makes simulated rats move naturally during goal-directed and exploratory behavior
SummaryThis paper introduces SCAMPER, a system for training an embodied virtual rat model. It uses a variational architecture with a motion prior for control and a behavior prior for exploration. The authors report that combining both priors produces natural movement in both goal-directed and random exploration, and allows learning from sparse rewards.
Aidan Sirbu, Charles Y. Zhang, Yuanjia Yang et al.
In briefWhat Removing a Visual Corruption Would Change: Failure Attribution for Vision-Language-Action Models
SummaryThis paper studies how visual corruptions affect vision-language-action (VLA) models and asks whether removing a given corruption would help the current episode. The authors propose a factorized risk model that separately represents the scene and corruptions, estimating removal benefits through latent intervention without paired clean images or extra counterfactual rollouts. Evaluated on OpenVLA and π0.5 in simulation and π0.5-DROID on a Franka robot, the method outperforms monolithic and vision-language-model baselines, with roughly 3x lower attribution error than frontier VLMs.
Carson Sobolewski, Anjali Parashar, Fei Chen et al.
In briefA theoretical comparison of maximum clique and maximum consensus for point cloud registration
SummaryThis paper theoretically studies the relationship between maximum clique and maximum consensus, two widely used formulations for robust point cloud registration. It shows that a maximum clique can have arbitrarily large alignment error and can be arbitrarily larger than the maximum consensus set at any fixed positive tolerance, and it derives geometric conditions that bound this error. For fixed rotation, the authors give sharp error bounds and tests that exactly characterize maximum consensus for the translation subproblem. They also determine the smallest group sizes for group-k consistency that guarantee error bounds independent of point cloud scale and geometry.
WhyTheoretically relates maximum clique and maximum consensus, deriving sharp error bounds and minimal group sizes, with Lean-verified proofs.
In briefFLINT: A Fixed-Budget Retrieval Model That Matches Queries and Documents Efficiently
SummaryFLINT (Fixed-budget Late INTeraction) is a retrieval architecture that uses attention-based pooling to represent each query and document with a fixed number of embedding vectors. Unlike methods that pool only documents, it pools both sides, which reduces document storage and the number of vector comparisons at retrieval time. The paper reports ranking effectiveness comparable to the closest baselines on BEIR and better results on several benchmarks with longer, more complex queries. On long-query corpora, it is 1.5 to 2.8 times faster end-to-end than ConstBERT and cuts index size by 5.3 to 9.3 times relative to ColBERTv2.
In briefCertifying repeatable entry-replay witnesses for verifying COBOL-to-Java migrations
SummaryThis paper proposes a method for generating synthetic end-to-end test cases for COBOL-to-Java migrations. Test cases are certified by observing execution traces on a surrogate COBOL program that mocks external interactions with a chronological operation tape. Applied to a 33,429-line COBOL program, the method produced 849 test cases covering 78.1% of source branches and 95.7% of paragraphs. When 450 defects were seeded into the Java code, the validation gate detected 47.6%, but missed rounding and precision defects.
Andras L. Ferenczi, Mariya Bessonov, Allwyn Ponnudurai et al.
In briefReadiness of African languages for AI safety inclusion
SummaryThis report assesses how ready African languages are to be included in AI safety discussions. Drawing on existing literature and data, it considers AI-relevant resources including text data, national AI policy support, performance on established text benchmarks, and speaker population. Languages are grouped into four readiness categories: fully ready, mostly ready, emerging, and foundational support needed.
Israel Abebe Azime, Jesujoba O. Alabi, Tajuddeen Gwadabe et al.
In briefChecking how safety preference labels are built, using recorded derivations
SummaryThis paper examines whether the construction of safety preference labels can be inspected when intermediate decisions are recorded. It presents Symbolic Preference Labeling, in which a pretrained language model classifies predefined attributes of prompts and responses, explicit rules assign response verdicts, and a fixed ranking sets the preference label. Each label's derivation can be recomputed from stored data without rerunning the model. The authors separate label reproducibility from attribute correctness, which and semantic safety they say require separate validation.
In briefProving the Ho–Nguyen conjecture for Gaussian mixtures and its use in mixtures of experts
SummaryOver-fitted Gaussian mixtures can have many parameter settings that give nearly the same density, so the order of cancellation affects learning rates. The paper proves the Ho–Nguyen conjecture that the order equals 2M for every cluster size M≥2. It applies the result to parameter learning and to mixtures of experts, separating high-order cancellation from exact parameter symmetries.
WhyResolves the Ho–Nguyen conjecture with a sharp sign-change bound and explicit construction, clarifying learning rates for Gaussian mixtures.
In briefArend: A Theorem Prover for Univalent Foundations
SummaryThis paper describes Arend, a theorem prover developed at JetBrains Research for formalizing mathematics in univalent foundations with a focus on constructive mathematics. Arend is based on a cubical variant of homotopy type theory and offers a powerful system of classes and records with subsumptive subtyping. An IntelliJ IDEA plugin provides an integrated development environment, and the arend-lib library covers constructive algebra, topology and synthetic homotopy theory.
In briefTxnQA: A benchmark testing AI reasoning and evidence tracing over transaction event sequences
SummaryTxnQA is a question-answering benchmark for analyzing merchant transaction histories, with 27 task families and a 5-scenario fraud-risk reasoning suite built on synthetic, privacy-mitigated data. Correct answers are computed programmatically, and model results are split into format compliance, content accuracy, and end-to-end correctness. The study finds that fine-tuning reliably enforces output formats but arithmetic remains a bottleneck, and that fraud verdicts are often correct while exact evidence attribution is much weaker.
Rabimba Karanjai, Yiming Bian, Yun-Shiuan Chuang et al.
In briefValiAssist: Scaling Validation of LLM-Generated Programming Feedback by Demonstration
SummaryThe paper presents ValiAssist, a system that infers reusable validation logic from instructors' revisions of AI-generated programming feedback. Instructors can inspect and refine this logic as it is applied to other student submissions. In a comparative study with 16 instructors, feedback aligned with their pedagogical intent rose from 57% to 91%, and accuracy in correcting problematic criteria rose from 32% to 77%.
In briefEditing test prompts to expose hallucinations in specific language models
SummaryThis paper proposes an iterative method that edits source-grounded seed prompts so they reveal hallucinations in a target model. An LLM editor revises each prompt using the target's answer and a judge's category-specific score and explanation. Tested on 500 seeds from seven datasets, the editing substantially raised judge-rated severity, and curated prompts gave similar average scores across five models yet had weakly correlated prompt-level results. The authors conclude that curated prompts' diagnostic value depends on source validity and on the specific failures each model exhibits.
In briefDiagnosing Whether LLM Reasoning Graphs Are Structurally Reliable
SummaryThis paper examines whether reasoning structures such as chains and graphs, which language models generate for inference and uncertainty estimation, are structurally reliable. It proposes a two-stage diagnostic evaluation on open-ended multi-hop question answering. Stage 1 finds graph reliability varies with topology, prompt design, and model family, while Stage 2 shows graph-based computation is sensitive to upstream errors. Gold-graph experiments indicate that graph structure helps confidence scores separate correct from incorrect answers and enables node-level diagnostics, though it offers little benefit as prompt context alone.
In briefPredicting Linear Relational Decodability Without the Relation
SummaryLinear relational embeddings (LREs) approximate how a language model maps a subject to an object, but their faithfulness varies widely across relations. The authors propose an entity lens, trained per entity type on a neutral prompt without any relation-specific information. On 28 relations in Olmo-3-7B, this relation-agnostic score correlates with LRE faithfulness (ρ = 0.7355, p < .01), suggesting decodability mainly reflects what the subject representation carries.
WhyProposes an entity lens predicting relation faithfulness without relation info; novel, but evidence is limited to one model and 28 relations.
Jonas Rohweder, Adrian M\"ulthaler, Faridoun Mehri et al.
In briefReinforcement-learning-based simplification of Russian text to CEFR levels
SummaryThis paper studies controlled automatic simplification of Russian texts to the A1, A2 and B1 proficiency levels using preference-based optimization. The authors fine-tune YandexGPT with Rejection Sampling and Direct Preference Optimization and compare several training configurations. The best model reaches 87.15% target-level accuracy, and manual review finds 75% of simplifications match the target level, though A1 remains the hardest.
In briefEGI-BENCH: A Unified Benchmark for Enterprise AI Agents
SummaryThis paper releases EGI-BENCH, a suite that combines eight enterprise-oriented LLM agent benchmarks behind one command-line interface, with quick and deep runtime-budget modes. It also defines the Enterprise General Intelligence (EGI) Score, a six-dimension framework whose buckets are filled by fine-grained sub-metrics from each benchmark. Under quick mode, fourteen LLMs score between 39.6 and 65.6, and no single model leads on every capability axis.
In briefMMedAgent-Plus: A Multimodal Medical Agent That Coordinates Specialist AI Tools
SummaryThis paper introduces MMedAgent-Plus, a multimodal medical agent that uses a vision-language planner to choose and run 18 specialist models covering 12 tasks and 8 modalities. The planner was instruction-tuned on 208,928 samples from 32 public medical benchmarks and synthesizes answers from the specialists' outputs. The authors evaluate tool orchestration, response synthesis, and end-to-end execution on held-out samples. End-to-end final-answer accuracy is reported as 0.98, matching the dedicated leading specialist and exceeding the strongest baseline at 0.28.
In briefNeighborhood-aware unlearning for generative recommendation models
SummaryGenerative recommenders predict items by autoregressively generating semantic identifiers (SIDs), whose tokens are shared across related items. The paper shows that removing training interactions can unintentionally degrade predictions for neighboring items, an effect hidden by global metrics. The authors propose Neighborhood-Aware Unlearning (NAU), which explicitly controls the probability mass assigned to the forgotten items' neighborhoods. Across three datasets and four SID tokenizers, NAU restores spam-induced exposure toward the retrained reference while retaining over 95% of retrained utility.
Pierre Lubitzsch, Yubao Tang, Maarten de Rijke et al.
In briefMagnolia: a persistent-memory AI agent for running computational chemistry workflows
SummaryThis paper introduces Magnolia, an AI agent framework that uses large language models to integrate computational chemistry tools such as docking, molecular dynamics, and quantum chemistry with lab notebooks and chat logs. It uses a three-tier memory system in which learnings are captured, confirmed at the project level, and promoted to cross-project rules only after human validation. The authors report four case studies, finding that runs with memory passed checks that runs without memory often failed.
Tao Jiang, Benjamin Navet, Yousouf Taghzouti et al.
In briefWhich Agent's Skill Should Change? Group-Relative Credit Assignment for Evolving Multi-Agent Teams
SummaryThis paper introduces GRSE, a method that lets cooperating language-model agents improve by revising reusable skills without weight updates. It assigns a shared team reward to specific agent-skill pairs by running each question several times under the same skill libraries and comparing the executions, with structured diagnosis of failures. On HotpotQA and LoCoMo, GRSE outperforms the prior MASkills method on most categories, with the largest gains on multi-hop questions, and remains robust when libraries start with planted harmful skills.
WhyProposes group-relative credit assignment for skills, beating the closest prior method on HotpotQA and LoCoMo, with corrupted-library robustness experiments supporting the claims.
In briefSparse reinforcement learning relaxation (SpaRLax) for LLM reasoning
SummaryThis paper addresses training large language models to use soft, continuous tokens during chain-of-thought reasoning. Prior reinforcement learning approaches treat soft tokens as actions and do not propagate gradients across them. The authors propose SpaRLax, which exploits the differentiability of soft reasoning and uses adaptive entmax sparsity to truncate noise when confident and keep multiple candidates when uncertain. The authors report the highest average performance over baselines across nine reasoning benchmarks.
Jiabin Fan, Matthias Lindemann, Zita Marinho et al.
In briefTheoretical analysis of environment-aware out-of-distribution detection for vision-language models
SummaryThis paper studies how continual vision-language models can reject inputs that do not belong to the current environment as they encounter a stream of environments. The authors propose a cross-environment maximum-concept-matching score that uses class information from earlier environments and prove it has a false-positive rate no larger than the standard MCM detector under stated conditions. The paper also distinguishes soft from hard OOD environments and analyzes how false-positive rates evolve in cascaded detection.
In briefLanguage Guidance for Vision Training: A Study of Generalization
SummaryThis paper studies whether pretrained language models can serve as lightweight auxiliary supervision for training vision models, without paired image-text data, prompt engineering, or contrastive objectives. It compares two forms of language guidance, Explicit Language Guidance (ExLG) and Implicit Language Guidance (ImLG), across six generalization regimes including out-of-distribution, shortcut resistance, texture bias, adversarial robustness, and continual learning. The explicit form tends to help in-distribution, low-data, and continual learning retention, while the implicit form is often more useful in shortcut-sensitive settings and under stronger adversarial perturbations. Both add minimal parameters and training overhead.
In briefEstimating diffusion score functions with shallow neural networks and how accurate that is
SummaryThis paper studies how many samples are needed to estimate Ornstein–Uhlenbeck diffusion scores with denoising score matching using shallow ReLU networks, and how this affects sampling. Under smoothness and growth conditions on the initial density, it derives Sobolev bounds uniform in time and high-probability risk bounds for path-norm-constrained training. Under additional assumptions, the squared score risk at a fixed time scales as Õ(n^{-(d+3)/(2d+3)}), and the resulting total-variation error between generated and original distributions scales as Õ(n^{-(d+3)/(4d+6)}).<br>
In briefLoWAgent: A Multilingual Benchmark Testing Web Agents in Low-Resource Languages
SummaryThis paper introduces LoWAgent, a benchmark for testing web agents on English and four low-resource languages: Hindi, Bengali, Urdu, and Amharic. It includes six localized websites and 3,000 task instances across five languages. The authors evaluated eight models and found performance drops in low-resource languages, especially for open-source models, with further declines from interface disruptions like pop-ups. Post-training on a small subset of benchmark traces significantly improved performance in these languages.
Ocean Monjur, Daud Ahmad Nisar, Shahriar Kabir Nahin et al.
In briefLSA-PAST: A Method for Learning New Object-State Combinations Over Time Without Forgetting
SummaryThis paper addresses compositional incremental learning, where visual concepts combine primitives such as states and objects, and new combinations arrive over time. The authors propose LSA-PAST, which adds depth-specialized LoRA adapters to a frozen ViT and uses a rehearsal-free affine transform to map anchored past prototypes into the current adapter space. Experiments reportedly achieve state-of-the-art accuracy-forgetting trade-offs for this setting.
Kyle Stein, Andrew Arash Mahyari, Guillermo Francia III et al.
In briefCulture Within: How LLMs Reason About Cultural Norms Internally
SummaryThis paper introduces CultureLens, a benchmark of 9,261 scenarios across 75 countries for studying how LLMs use cultural information. Stating a contradictory norm lowers aligned answers by 32 to 47 percentage points across three models. Supervised fine-tuning with contradictory-norm examples improves aligned answers under such norms, including for unseen countries.
In briefPreference-aware steering of task trade-offs in multi-task speech LLMs
SummarySpeech LLMs handle several tasks, such as speech recognition, timestamp prediction, and speech translation, but joint training creates trade-offs that fixed loss weights do not control according to deployment preferences. The paper proposes EA-SURF, which adapts scalarization weights during training by tracking each task's performance against its reference. Across six three-objective requests, the method empirically converges toward the requested targets. It increases endpoint-inclusive normalized hypervolume by 17.8% over fixed loss weighting.
Liuyuan Jiang, Vishal Sunder, Samuel Thomas et al.
