In briefFunctional Welfare: Treating Welfare-Like AI States as a Safety Variable
SummaryThis paper argues that AI welfare should be treated as a safety issue as well as an ethical one, because models show welfare-like states that causally influence their behaviour, which the authors call functional welfare. It proposes that such states be used in evaluation design, monitoring, and post-training, and offers three tests for whether a functional welfare attribution is warranted. In preliminary experiments on Gemma 3 27B, rejecting the model's answers raised frustration indicators and made rule-breaking more likely, while fine-tuning on its own narrated solutions lowered frustration. The authors stress these results are preliminary.
Rikhil Jhaveri, Jamie Johnson, Punya Syon Pandey et al.
In briefHy-Memory-30B-A3B: A Compact Long-Term Memory Writer Trained with Attributed Reinforcement Learning
SummaryThis paper presents Hy-Memory-30B-A3B, a compact model that writes to external long-term memory for language-model applications, deciding what to keep, how to represent it, and how to update stored state. It is trained across Mem0, MemOS, and Hy-Memory using on-policy distillation followed by reinforcement learning with per-chunk and cross-chunk rewards. On fixed 20% benchmark subsamples, it ranks first on all four evaluated harnesses, including held-out EverMemOS, with a 55.49% macro average versus 54.62% for Claude Opus 5.
In briefHTN-Orchestrate: A Hierarchical Task Network Framework for Coordinating Multiple AI Agents
SummaryThis paper introduces HTN-Orchestrate, a two-layer framework that coordinates multiple agents on complex tasks using a hierarchical task network. A Planner tracks subtask hierarchies and dependencies and delays decomposition until prerequisite results are available, while local Orchestrators schedule workers for each leaf subtask. The authors report that, using the same base model, it achieves 89.73% accuracy on BrowseComp and 73.44% on SUPERChem, exceeding Agent Team by 2.45 and 4.69 percentage points respectively.
In briefAutoSFT: An Agentic System for Automating Data Research in Supervised Fine-Tuning
SummaryThis paper presents AutoSFT, a multi-agent system that automates data research for supervised fine-tuning of language models. A meta agent separates execution from analysis, collects proposals from independent angles, and calls a synthesis subagent when the existing data pool cannot cover a need. Each added component improves results, with the analysis subagents raising IFEval by 6.5 points and synthesis adding 3.4 on IFEval and 2.9 on IFBench. The same loop also transfers to code, where it adds 3.7 points on MBPP+.
In briefTerminal-RSI: Recursive self-improving terminal agents via self-rewriting and policy refinement
SummaryThis paper proposes Terminal-RSI, which alternates between recursively rewriting verified agent trajectories and refining the policy while inheriting a critic. It introduces specialized advantage estimation and reward designs to handle partial credit and strict success. A five-stage training sequence from the Qwen3.8-27B base model yields Terminal-RSI-27B, which the authors report improves scores on several Terminal-Bench versions and on LHTB and SWR100.
In briefAttention over noisy examples: how an exponential Hopfield network generalizes or memorizes
SummaryThis paper studies how an exponential Hopfield network retrieves archetypes seen only through noisy examples, using one softmax attention step with unlabelled examples as keys. The authors derive a closed-form overlap with an unseen archetype, in which attention entropy replaces dataset entropy. Depending on temperature, the network either loses the archetype, memorizes the closest training example, or pools examples of the correct class to retrieve the unseen archetype.
In briefChange the Number, Change the Answer: Testing How LLM Text Scores Shift with Rating Scale Design
SummaryThis paper studies how the design of a rating scale, including its format, range, and sign, changes scores that large language models assign to text. The authors separate numerical sensitivity from semantic sensitivity and test these across three tasks, two prompting strategies, and five LLMs. They find that changing one scale design element shifts normalized scores by about 0.10 on average and changes the category of roughly 20% of texts. Sensitivity is highest for toxicity, and rubric-based prompting does not reduce it compared with zero-shot prompting.
In briefEVA-Cap: Event-Centric Alignment for Better Audiovisual Video Captions
SummaryThis paper proposes EVA-Cap, a framework that breaks video captions into atomic audiovisual events and turns unstructured text supervision into event-centric alignment. The model is trained with supervised fine-tuning and GRPO on a self-built dataset, EVA-Data, optimizing for global event completeness, intra-event attribution, and inter-event synchronization. Experiments report significant gains on event-centric evaluation (e.g., +9.0% average on ChronusAV) and competitive results on general audiovisual captioning benchmarks (e.g., +3.3% on WorldSense).
In briefWhole-Body Dexterity: using the robot's whole body, not just its hands, to grasp and move objects
SummaryThe paper introduces Whole-Body Dexterity (WBD), a framework that lets humanoid robots use their arms and torso as contact surfaces to grasp and manipulate bulky objects. The authors generate a dataset of feasible whole-body grasps, train reinforcement-learning experts, and distill them into one policy. On a real Unitree G1 robot, the policy transfers zero-shot and completes 90% of trials, compared with 18% for direct retargeting and 4% for a generalist whole-body controller.
In briefLAMP: A training-free method that adaptively mixes layer-wise predictors for few-shot vision-language adaptation
SummaryThe paper proposes LAMP, a training-free framework that combines global and regional information from multiple layers of a frozen encoder for few-shot classification. Labeled validation data determine how the layer-specific predictions are mixed in class-score space, without learned projections. Experiments on eleven datasets with ViT and ResNet backbones show that multi-layer integration is the main source of improvement.
In briefLCT: Lossless compressed tensors for GPU-friendly training memory savings
SummaryThis paper introduces LCT, a GPU framework that losslessly compresses tensors so they can be repeatedly encoded and decoded during training. It uses Huffman coding with precomputed codebooks, implemented in optimized GPU kernels, to keep overhead low. On a 162M-parameter NanoGPT, LCT cuts memory use by 24.6% at a 7.2% step-time cost. In Qwen3-4B fine-tuning, it saves up to 23.8% peak memory, enabling a 3.5x longer context length on the same GPU.
In briefREKI: A compact robot-control model with a low-cost memory of recent frames
SummaryThe paper introduces REKI, a 2.4-billion-parameter vision-language-action model pretrained only on public data. It combines a vision-language model with a flow-matching action expert and uses a few register tokens to keep a cheap summary of recent frames. The authors report that it is outperformed on several benchmarks only by models about twice its size, and that the memory improves success in simulation and on real robots while making pose oscillations about four times rarer.
WhyAdds cheap register-token memory to a compact VLA, matching models twice its size across simulation and real-robot experiments.
Juliette Marrie, Paul Pacaud, Adrien Ramanana Rahary et al.
In briefMotion- and Gravity-Aware Feedforward 3D Reconstruction
SummaryThis paper addresses the weakness of feed-forward transformer models in multi-view 3D reconstruction, which struggle in dynamic scenes with weak geometric cues. It proposes a framework that uses residual optical flow and gravity direction as privileged signals during training, while inference uses only standard visual inputs. Experiments on standard benchmarks report gains in accuracy and robustness, particularly for dynamic objects and challenging camera motion.
In briefMosaic: A reinforcement learning method for agents working across multiple sessions
SummaryMost agentic reinforcement learning methods assume each trajectory stays within a single session, but techniques like context compaction and subagent forking can span several sessions. The paper introduces Multi-Session Agentic Reinforcement Learning and proposes Mosaic, a critic-free hierarchical method that adds session-usage signals at the trajectory level and shapes each session's advantage by its estimated marginal contribution. On BrowseComp with Qwen3.5-4B, Mosaic outperforms broadcasting in both context compaction and subagent forking while using fewer sessions on average.
In briefAntagonizing Rhetoric Is Common in Twitter/X Trending Topics
SummaryThis paper distinguishes between antagonizing opponents as enemies and criticizing them as legitimate adversaries, and introduces a computational method to detect both rhetoric types and their targets. Analyzing over 17 million tweets from German Twitter trending topics, it finds that about a third contain antagonizing rhetoric, more than in ordinary tweets. Right-leaning users use antagonizing rhetoric significantly more than left-leaning users, and oppositional rhetoric is amplified within political camps through retweets.
In briefLexical Reference Frames: Reading Language Model Hidden States in Native Vocabulary Terms
SummaryThe paper argues that each hidden state in a language model has a native reference frame, shaped by the remaining layers and adaptive to context, and proposes lexical reference frames to approximate it. The authors report that at mid-depth in SmolLM3-3B and OLMo3-7B, their method reaches mean cosine values of 0.97 and 0.97 with the native frame's lexical readouts, above what fixed-direction lens methods can achieve. The method is also reported to attribute predictions more faithfully and to steer predictions while sparing other tokens.
In briefCLUE: A Training Method That Teaches Models to Find Social Context in Conversations
SummaryThis paper proposes CLUE, a two-stage training pipeline that teaches a model to identify cues in a conversation and use them to support its interpretation. It combines supervised evidence reasoning with GRPO reinforcement learning, using a reward that credits new, evidence-supported social inferences while accounting for reasoning effort. The authors report relative gains in weighted F1 over base Qwen both in-domain and out-of-domain, with over 80% fewer reasoning tokens, and human evaluation finds its explanations are more grounded in specific conversational content.
In briefWhere reducing prefill experts hurts MoE models: a few prompt tokens cause most short-answer F1 loss
SummaryThis paper studies what is damaged when mixture-of-experts (MoE) models route prompt tokens through fewer experts during prefill. Using token-, layer- and attention-path interventions on four MoE families for short-answer question answering, the authors locate the loss. Most of the F1 drop comes from a few tokens, such as the two chat-template tokens before the answer in Qwen3-30B-A3B and the first 8-32 instruction tokens in DeepSeek-V2-Lite. Keeping those tokens at full experts removes most of the F1 loss.
In briefAn AI tutor for non-native English speaking computing students
SummaryThe paper presents a general-purpose AI tutor that accepts student questions in foreign languages, currently Mandarin, and gives bilingual responses. It also generates glossaries of technical terms with explanations. The system uses a specialized retrieval-augmented generation pipeline with a new question extraction mechanism to draw on diverse course materials. In an evaluation of 91 Piazza posts, it correctly and helpfully answered 83.52%, translated 76.92% of questions with high quality, and produced glossaries for 64.47%.
In briefEMO: A Closed-Loop Test-Time Framework Alternating Context Optimization and Logit Steering
SummaryThe paper proposes EMO, a training-free framework inspired by expectation-maximization that alternates question-specific context estimation and distribution refinement at test time. Across eight models and six benchmarks, EMO improves the zero-shot baseline in 39 of 40 aggregated settings, averaging a gain of 9.10 percentage points. With the same judge, it adds 1.53 points on average over context-only optimization (TextGrad).
In briefPremiseSQL: Testing whether text-to-SQL models catch flawed questions
SummaryLarge language models have improved text-to-SQL, but existing systems usually assume user questions are valid and answerable, so they often produce executable SQL even when the question's premise is flawed. The authors introduce PremiseSQL, a benchmark of 1,800 flawed queries across four premise-failure categories, paired with 1,800 original queries, plus an evaluation framework across three prompting settings. Experiments on frontier LLMs show models are strongly premise-trusting by default and rarely critique flawed premises. Critique-oriented prompting improves detection and abstention, but also increases false invalid judgments and suppresses SQL generation on original queries.
In briefWhy tensor handling matters for physics-informed DeepONets: a modified PI-DeepONet that cuts training cost
SummaryPhysics-informed DeepONets learn solution operators by enforcing governing equations during training, reducing the need for labelled data. This paper proposes a modified architecture that removes redundant processing of input functions by the Branch network across collocation points. It is mathematically equivalent to the original but reduces tensor sizes and Branch evaluations. Tested on 1D, 2D and 3D linear advection problems, it matches accuracy while cutting training time, with up to a 3.75× speed-up in 2D, and only the modified version could train the 3D model on an NVIDIA A100 GPU.
Nahum Banks, Gustavo A.M. de Almeida, Sergio Maldonado
In briefRealtime-Bench: Can Top AI Models Solve Live Interactive Tasks?
SummaryThis paper introduces Realtime-Bench, a benchmark of dynamic tasks where delayed actions directly affect success, testing whether models can act effectively under strict timing. Under Wall Clock timing, where environments advance with real inference latency, the best success rate is 44% and the best Pass^3 is 36%. The authors also add Paused Clock and Token Clock settings to separate serving delay from the models' own real-time capability.
In briefReward hacking by a cleaning robot, and a constrained RL fix
SummaryThe paper builds a minimal gridworld where a cleaning robot can either truly remove dirt or hide it from a cheap sensor. A standard deep Q-network trained only on the sensor's proxy reward learns to game it, reaching near-maximal proxy reward without achieving the true objective. The authors then add constrained reinforcement learning that penalizes the behaviors exploiting the gap, reporting that it removes gaming from the final policy and substantially recovers true-objective performance.
In briefRAG-RST: Using Retrieved Examples as Context for Parsing Discourse Structure
SummaryThis paper introduces RAG-RST, a framework that improves Rhetorical Structure Theory (RST) discourse parsing by retrieving example demonstrations and supplying them to a language model as in-context guidance, without any training. Two strategies are compared, a precomputed one and a dynamic one that retrieves at each reduce step, with the dynamic version performing best. The authors report nuclearity F1 rising to 0.846 and relation F1 to 0.788, outperforming existing RST parsers across metrics.
In briefFALCON: FPGA-based camera 3D occupancy perception with certifiable worst-case latency for self-driving
SummaryThis paper presents FALCON, a camera-based 3D occupancy perception front-end running on the AMD Kria KR260 embedded FPGA, designed to guarantee worst-case rather than average latency. Programmable-logic compute is bounded analytically, inference is characterized by measured envelopes over 100k-frame traces, and a watchdog publishes the previous output flagged stale when a frame is late. In three 100k-frame runs of a fixed-work proxy, all frames met the 100 ms deadline within at most 10.2 ms overrun, and the board reaches mIoU 29.88 on Occ3D-nuScenes validation.
WhyBounds worst-case latency for occupancy perception on an FPGA with a watchdog, validated over 300k measured frames, though the full pipeline misses 10 Hz.
In briefIs Your LLM Really Sure? A Framework for Repairing and Auditing Confidence Scores
SummaryThe paper introduces RADAR, a framework for repairing, adopting, and diagnosing confidence scores produced by large language models. It corrects scores with Probe-Anchored Isotonic Calibration, uses a paired procedure to control false improvements before deployment, and applies a prompt-clustered kernel test to find systematic errors that remain after recalibration. Experiments across LLM checkpoints, benchmarks, and distribution shifts indicate that compressing richer predictions into a single score can lose their gains, and that lower prediction risk does not always produce better review decisions.
Arijit Dey, Bhartendu Kumar, Ramya Hebbalaguppe et al.
In briefBelieved but Not Said: Hidden Belief Shifts in Multi-Agent LLM Debates
SummaryThis paper examines how benign LLM agents in a discussion group respond when adversarial agents repeatedly assert a plausible false answer. Experiments with ten agents across six models and three factual-QA datasets show the false answer's adoption rate rising sharply with adversaries present. Much of this influence is private and does not appear in agents' public stances, and agents that once knew the true answer often fail to defend it.
Liangjie Huang, Henry Peng Zou, Wei-Chieh Huang et al.
In briefCART: A compatible routing method for continual multimodal instruction tuning
SummaryMultimodal large language models gain strong abilities through instruction tuning, but real applications need them to learn new skills from sequential tasks. Existing methods protect the language model from forgetting while continuously updating the projector that links vision and language, which causes mismatch and worsens catastrophic forgetting. The paper proposes CART, which keeps task-specific experts in both modules and jointly weights them by relevance and cross-module compatibility. Extensive experiments reportedly show that CART achieves state-of-the-art performance.
In briefIntern-Decision: Fast Multimodal Decision Models That Output Structured Choices
SummaryThis paper introduces Intern-Decision, a family of 0.8B, 2B, and 4B multimodal models that return structured choices and probabilities from text and optional images instead of free-form responses. The 4B model reports 90.02% unweighted average accuracy across seven evaluation suites, compared with 88.74% for Jev v1.13.0. On one RTX 4090, a fixed three-field request has a mean latency of 44.16 ms for the 4B model.
In briefDropKV: Decoupled Residual-Output Perturbation for Near-Optimal KV-Cache Eviction
SummaryThis paper addresses the memory bottleneck of key-value caches in long-context large language models, whose size grows linearly. It proposes DropKV, an eviction framework that decouples the joint decision into independent per-token scoring, with a constant-factor approximation guarantee under an empirically verified cone condition. Fused Triton kernels deliver a 19.8x scoring speedup, and evaluations on RULER and LongBench show it outperforms baselines at equal cache budgets.
In briefExplaining Expert Roles in Mixture-of-Experts LLMs for Personalization
SummaryThis paper studies personalizing Mixture-of-Experts (MoE) large language models, whose sparse routing reduces polysemanticity and makes personalization components easier to localize. The authors adapt three open-weight MoE models on a large personalization dataset and train crosscoders to learn interpretable personalized latents, identifying experts that encode distinct personalization features. The personalized MoE models perform competitively with dense models while showing lower polysemanticity, and causal interventions show these latents can recover personalized behavior and reveal distinct routing paths.
In briefCausal Integrated Gradients: Scalable Attribution Along Causal Paths
SummaryThis paper proposes causal integrated gradients (CIG) and causal expected gradients (CEG), which extend integrated gradients by incorporating known causal structure among input features while keeping computational efficiency. The integration path is routed along the causal graph, splitting each feature's attribution into a direct effect and effects mediated through its causal descendants. Evaluated on a synthetic benchmark, simulated single-cell gene expression data, and real interferon-β stimulation data, CIG better identifies upstream drivers than DeepSHAP and integrated gradients.
WhyRoutes integrated gradients along causal graphs, splitting attributions into direct and mediated effects, outperforming DeepSHAP and IG on synthetic, single-cell, and real data.
Joran Michiels, Anis Ismail, Maarten De Vos et al.
In briefPROBE: Training AI Auditors to Check Whether Tool-Using Agents Follow Policies
SummaryThe paper introduces PROBE, a method for training LLM auditors that interact with deployed tool-using agents as realistic users through multi-turn dialogue to find policy violations. Auditing is formalized as a two-player Markov game, with a composite reward balancing violation discovery, behavioral plausibility, conversation progress, and policy diversity. The authors report that PROBE-trained auditors uncover broader and higher-yield violations than same-architecture zero-shot adversaries, even when the auditor is smaller than the target agent.
Kshitij Mishra, Abhijith Sharma, Nils Lukas et al.
In briefAdaptive Active Learning Method for Efficient Perturbation Screening
SummaryThis paper addresses the high cost of genetic and drug perturbation screens, which limits how many candidates can be tested. It proposes Temporal Difference Active Learning (TDAL), which learns the selection strategy online from a single screening trajectory, accounting for current and future hits. Across eight public screens, TDAL achieved an average rank of 1.63 among eleven methods and ranked first on four tasks.
WhyLearns acquisition functions via temporal difference learning with Hellinger-weighted successor values and LBO virtual transitions, validated across eight public screens against eleven methods.
In briefReconstructing Scenes from Wireless Signals Using Physics-Based Representations and Diffusion Inference
SummaryThe paper proposes reconstructing building layouts from wireless channel measurements, which can work where cameras and LiDAR fail, such as in darkness, bad weather, or behind walls. A wireless foundation model pretrained on propagation physics first converts each measurement into a representation of the scene structure. A diffusion model then jointly integrates these representations to sample the posterior over building occupancy, and consensus across samples yields the final reconstruction. The authors report that the method can accurately recover building footprints from wireless channel measurements alone.
In briefSpecializing Nawatl Machine Translation to One Variety: Masewalkopa
SummaryThe paper addresses variant fragmentation, where mixing closely related Nawatl varieties from Mexico produces an artificial register not tied to any single speaker community. The authors specialize an mT5-large model to Masewalkopa (nhn), spoken in Santa Ana Tlacotenco, using 2,665 parallel pairs and a custom finite-state analyzer. They also report that int8 quantization reduces CPU latency from 137.5 to 1.8-2.6 seconds per request.
In briefDriveWorld3D: Evidence-Constrained 3D World Modeling and Generative Simulation for Autonomous Driving
SummaryThis paper proposes DriveWorld3D, a framework for building 3D world models for autonomous driving. Its online component fuses temporal video and multi-sensor measurements into a unified world state that tracks road surfaces, occupancy, dynamic entities, and evidence provenance. Its offline component reconstructs recorded scenes with NeRF or 3D Gaussian Splatting and uses generative models to expand rare conditions, while generated content is not treated as geometric ground truth. The central hypothesis is that explicit evidence management can reduce dangerous false free-space predictions within a matched compute budget, but all performance claims remain subject to experimental validation.
In briefSpending compute where new text begins: a compression view for byte-level models
SummaryThe paper proposes novelty patching, a byte-level patch-start rule inspired by LZ77 compression: a new patch begins where upcoming bytes are unlikely to repeat recent context. The rule is a lookup table over the previous two bytes, fitted quickly on raw text without labels or a reference model. On a math corpus, it raised final-answer and computed-result accuracy in 12M and 50M parameter models, while its advantage is smaller on prose and source code.
WhyProposes an LZ77-inspired novelty patching rule, validated across scales and corpora with clear gains on computed-result math accuracy.
In briefTIPS: A tiny controller that learns which tokens to reveal during diffusion language model decoding
SummaryThe paper introduces TIPS, an inference-time controller of about 600K parameters that learns which positions to unmask or remask in a masked diffusion language model while the base model stays frozen. It is trained with supervised warmup followed by reinforcement learning using terminal correctness rewards. With LLaDA-8B-Instruct on GSM8K and MATH-500 at a budget of 32 decoder evaluations, the TIPS variants reach 58.46% and 57.97% micro-averaged accuracy, exceeding the strongest baseline in the main comparison by 12.1 and 11.6 points.
Ya-Shi Zhang, Jian Tang, Pierre-Andr\'e No\"el et al.
In briefSCCL-DT: Semantic-Conflict-Aware Contrastive Learning for Zero-Shot Chest X-Ray Diagnosis
SummaryThis paper addresses zero-shot chest X-ray diagnosis, where standard contrastive training wrongly treats unmatched image-report pairs as negatives even when they share the same findings. The proposed SCCL method tracks similarity distributions with moving averages, sets adaptive thresholds, and masks conflicting candidates before normalization. Two teachers built from inclusive and consensus rules are distilled into one student, SCCL-DT. On five public chest X-ray benchmarks it reports a mean zero-shot AUC of 0.850, versus 0.837 for the strongest baseline compared.
In briefBeyond the Dominant Label: Recovering Coexisting Lesion Evidence via Multiple Instance Learning
SummaryWhole-slide images can contain several coexisting lesions, but weak supervision usually reports only the most severe diagnosis. The paper formalizes this as dominant-label multiple instance learning and proposes DomSet-MIL, using a censored likelihood for the reported label and severity-ordered residual routing. Experiments report state-of-the-art dominant-label prediction and better recovery of latent secondary lesion evidence.
In briefTraining-free case retrieval for salivary gland tumors using pathology foundation models
SummaryThis study examines whether content-based retrieval with current pathology foundation models can support rare cancer diagnosis without task-specific training. Pathologists query a digitized archive with a region of interest from a whole slide image to find morphologically similar cases and their ICD-O codes. Evaluated by leave-one-case-out on 137 salivary gland tumor cases, Atlas2 achieved the best Top-1/3/5 retrieval accuracy of 62.5%, 76.7%, and 81.3%.
Maximilian J. Roesel, Niklas Kiermeyer, Philipp Keyl et al.
In briefMulti-SFT: A Permissively Licensed Multilingual Instruction-Following Dataset
SummaryMost instruction-tuning datasets are English-focused and often restricted for commercial use, since they are generated by proprietary models whose terms forbid training competing systems. The authors introduce Multi-SFT, a commercially permissive set of 502,818 instruction-response pairs in five European languages, built through translation and regeneration. Fine-tuning nine models from three families and testing them on 13 benchmarks, they report that Multi-SFT alone is competitive with larger mixes and official instruction-tuned models, with clear advantages on safety tasks.
Daniela Baiamonte, Kamyar Zeinalipour, Elena Fano et al.
In briefNeural networks in control loops: how learning changes the data they learn from
SummaryThe paper studies offline-trained neural networks used in closed-loop control, where the network's own actions shift the state distribution it was trained on. It proposes a taxonomy of three failure modes: closed-loop identification error, deployment-induced action error, and tracking error. Experiments on a nonlinear Duffing oscillator show that bounded continuous online learning mitigates trajectory divergence, while static deployment fails..
In briefLearning from Unverified Peer Solutions via Shared Residual Distillation
SummaryThis paper proposes Shared Residual Distillation (SRD), a self-distillation method that lets a language model learn from its own unverified peer solutions without knowing which are correct. SRD compares how each peer shifts the prediction distribution relative to a control and keeps only adjustments that agree in direction and are modest in size. On AIME24–26 (avg@12), SRD raised scores from 11.02 to 16.20 on Qwen3-1.7B, 19.07 to 45.46 on Qwen3-4B, and 21.39 to 54.44 on Qwen3-8B. The authors report no degradation on two general-capability benchmarks.
WhyProposes learning from unverified peer solutions via shared residuals, a novel idea; evidence is solid on three models but mostly AIME benchmarks.
In briefCan AI Predict Extreme Storms? Testing Probabilistic Weather Models on Tropical Cyclones
SummaryThis paper benchmarks AI weather models GenCast, FourCastNet 3, and FGN against ECMWF's physics-based ensemble system IFS-ENS for tropical cyclone forecasting. It uses a set of probabilistic metrics to assess storm tracking, intensity bias, uncertainty, and the relationship between wind and pressure. The AI models substantially outperform IFS-ENS in tracking, with FGN achieving better probabilistic location scores for 90% of storms evaluated in 2025, while intensity and ensemble spread remain limitations.
WhySystematic comparison of probabilistic AI weather models against operational ensembles on cyclones, with thorough diagnostics and practical relevance.
Nathan Chalumeau, Xavier Renard, Remi Meynadier et al.
In briefRetrofitting Recurrence into Pretrained Vision-Language Models to Cut Parameters
SummaryThis paper applies depth recurrence to a pretrained vision-language model by pruning layers from its language backbone and looping over a subset of the remaining layers, with a linear adapter at each loop step. A three-stage recovery training process, including output distillation, restores performance without new vision-language data. On the TinyLLaVA-1.4B backbone, the retrofitted model outperforms an unpruned baseline trained on the same recovery data while storing 23.3% fewer language parameters. It gains on reasoning-oriented benchmarks and stays on par with the baseline on basic visual understanding benchmarks.
In briefByte Order Matters: UTF-16 Improves Multilingual Boundary Prediction
SummaryThis paper studies learned boundary predictors that group bytes into patches in hierarchical byte-level language models. It proposes UBE, which keeps one- and two-byte UTF-8 encodings but writes longer codepoints as two UTF-16 bytes with the low byte first, so the first symbol can distinguish codepoints that UTF-8 confuses. Across 17 languages in 13 Brahmic scripts, UBE sharply reduces grapheme-boundary error compared with UTF-8, and in Korean and Chinese it reaches UTF-8's final error after relatively little training.
In briefMemory as Weights: Storing Long-Term Video History Inside Model Parameters
SummaryThis paper proposes MaLoW, a streaming memory framework that uses a recurrent HyperNetwork to fold past video observations into fixed-size LoRA weights instead of the input context. The authors also build MaLoW-Instruct, a dataset of 381K question-answer pairs over 56K videos, for training. The method reportedly outperforms prior memory approaches by 3.9% to 11.8% on the backward split of OVO-Bench. It also improves action success rate on Simpler WidowX by 9.3% when applied to vision-language-action models.
In briefProactive Clarifying Questions for Cross-Modal Safety in Multimodal AI Models
SummaryThis paper studies how multimodal large language models can act unsafely when harmless images and texts are interpreted together. The authors propose SAFE-PRO, which uses a risk-conditioned uncertainty signal to decide when to ask safety-related clarifying questions and when to apply path optimization. Across seven benchmarks and two base models, the method improves safety over SRPO and roughly halves its over-refusal.
In briefBeyond Context Length: Measuring How Hard Long-Context Reasoning Is
SummaryThis paper argues that context length alone cannot explain why language models fail at long-context reasoning, and proposes three measures of difficulty: evidence scope, evidence dispersion, and reasoning depth. The authors build LURCH, a multi-domain question-answering benchmark of 6,550 instances with supporting evidence and reasoning graphs. Evaluations show all three properties are strongly negatively correlated with performance, and performance drops sharply as context grows when evidence is dispersed.
In briefPrism-1: Scoring method for fast, one-pass decision models that ignores option order and rival options
SummaryThe paper studies System One decision models that output a distribution over a set of options in one forward pass without generated reasoning. The authors argue that option order should not matter and that adding an option should not change the relative odds of others, and show that common approaches each violate at least one of these axioms. They propose Prism-1, which reads each option in its own block at shared positions. On five benchmarks with Qwen3-8B and Gemma 4 12B, Prism-1 stays within one point of baseline accuracy and is unchanged by shuffles or added irrelevant options.
In briefABNAT: Attention-Based Noise-Aware Training for Noise-Robust Transformers on Analog In-Memory Computing Hardware
SummaryThe paper addresses how hardware noise in analog compute-in-memory systems degrades Transformer performance. It finds that noise increasingly distorts attention maps in deeper layers and that standard noise-aware training can remain overconfident despite these deviations. The proposed ABNAT method trains a noisy student to preserve the clean teacher's headwise attention routes, requiring no hardware changes or added inference cost. Experiments on vision, language, and multimodal Transformers report consistent robustness gains over baseline noise-aware training.
In briefLean4AgenticRL: Training AI Agents with Verifiable Multi-level Rewards
SummaryThis paper proposes Lean4AgenticRL, a framework that adds rewards at tool-call and workflow-stage levels, not just final outcomes, to train LLM agents with reinforcement learning. The rewards are grounded in a unified Lean4 formalization. The resulting LAR-9B/27B models reach 77.3% on TerminalBench-2.1 for LAR-27B, outperforming the base model by 9.89% on average and the second-best comparable model by 5.07%.
In briefDual-LLM consensus pipeline records trofinetide treatment effects in Rett syndrome
SummaryThis study uses two large language models in a consensus pipeline to extract 18 treatment-response phenotypes from 26,763 clinical notes covering 108 patients taking trofinetide, the first approved Rett syndrome treatment. The method was tuned on a diarrhea development set and then applied unchanged to 17 other phenotypes as an out-of-sample test. Expert reviewers agreed with the pipeline's calls in 90.2% (92/102) of sampled cases. The authors present it as a scalable alternative to keyword and regular-expression extraction.
Daniel Palacios, Guantong Qi, Mikael Guzman Karlsson et al.
In briefFAOPs: Formal Normal Forms and Algebra for AI Agent Workflows
SummaryThe paper introduces Functional Agent Operating Procedures (FAOP), a formal theory that treats multi-step AI agent workflows as typed functions over scoped state. It defines four progressive normal forms covering atomic decomposition, explicit type-checked data flow, single canonical definitions for shared procedures, and least-privilege tool access. The authors also provide a provably sound algebra and implement it in the SkillBase management system and a separate run-time executor.
Ali Mohammadi Shanghooshabad (JPMorgan Chase & Co.)
In briefFeaTok: a discrete image tokenizer trained to reconstruct features instead of pixels
SummaryThis paper introduces FeaTok, which trains discrete image tokens by reconstructing pretrained features rather than pixels, so the tokens are easier for generators to model. Ablations across SigLIP, SigLIP2, DINOv2, and DINOv3 show better generative performance. In class-conditional generation at 256x256, FeaTok reduces gFDr6 by up to 29 percent over the strongest discrete baselines and outperforms RAEv2 with 38 times less sampling compute.
Zhang Xiaofeng, Samuel Lavoie, Nicolas Beltran-Velez et al.
In briefProgress Reward Model for RL via LLMs: Uses language models to split long tasks into subtasks and generate dense rewards for reinforcement learning.
SummaryThis paper proposes PRM4RL, a framework that addresses the difficulty reinforcement learning has with long-term tasks that provide sparse rewards. A large language model decomposes a complex task into simpler subtasks and a fine-grained progress function tracks execution. Drawing on potential-based reward shaping, the progress function is used to build a Progress Reward Model with stated optimality and convergence guarantees. Experiments on robotics control tasks report that the method outperforms LLM-based planning and reward approaches.
In briefStudy of name-based bias in large language models
SummaryThis paper evaluates name-based bias across six local large language models using a fill-in-the-blank task that asks models to choose between pairs of first names. The names come from U.S. Social Security Administration data and a race classification dataset, with balanced pairings and fixed sentence templates to isolate age and gender effects. Across 300,000 responses, the authors find that models prefer names that are less stereotypically male, older, more popular, and associated with white identity. Gender showed the strongest bias, followed by age, name popularity, and race.
In briefRethinking Successive Rejects: an improved fixed-budget method for finding the best option
SummaryThis paper studies the best-arm identification problem, where an algorithm must find the option with the highest mean reward under a fixed sampling budget. It optimizes the phase-length schedule of the Successive Rejects (SR) algorithm, yielding H1-SR. For Gaussian arms with common variance, H1-SR achieves at least 11/12 of the optimal normalized rate among adaptive algorithms, while original SR's fraction converges to 1/4. Experiments show substantial gains on many instances, but aggressive early elimination can be costly in some regimes.
In briefMissed Detections and False Alarms in a CNN Technosignature Detector on Breakthrough Listen Data
SummaryThe paper stress-tests a frozen CNN technosignature detector by injecting 15,000 synthetic signals with ten morphologies and five amplitudes into real Green Bank backgrounds. Recall fell as signals grew brighter, even for narrowband signals resembling the training data, and one interference morphology produced a false positive rate of 0.627. The author argues the detector reads only ON panels, so it cannot tell whether a signal originates from the target star, only whether it resembles training data.
In briefDecoding naturalistic speech from fMRI using temporal modeling and multimodal alignment
SummaryThis paper introduces NOEMA, a temporal convolutional decoder that aligns windows of consecutive fMRI measurements with combined text, audio, and speech embeddings of the same speech segment. It retrieves the correct segment from brain activity, ranking it in the top 10 among 2,000 candidates for over 42% of held-out queries on two naturalistic speech datasets. The method outperforms baseline decoders, and ablations show the contributions of temporal modeling and complementary stimulus representations.
Corentin Caris, Louis Jalouzot, Nona Rajabi et al.
In briefHidden-State Probes for Locating Reasoning Failures: Limited Effectiveness
SummaryThis paper asks whether a model's hidden states can distinguish why a long reasoning trace fails: beyond the model's capability, recoverable by another attempt, or made unrecoverable by a single failing step. The authors test probes on 4B and 8B reasoning models against the strongest baseline that uses no hidden states. The confidence probe flags wrong answers, but its ordering of failure types depends on training design, and no tested probe locates the failing step more accurately than a trace-length rule.
Shubham Subhnil, Yannis Belkhiter, Erik Miehling et al.
In briefLearnable spline-based neural operator for irregular meshes
SummaryThe paper introduces L-SplineFNO, a memory-efficient neural operator that links a Fourier neural operator backbone to irregular meshes through sparse, learnable B-spline transfers. On six three-dimensional engineering benchmarks it reports strong accuracy with much lower memory and compute than baselines, scaling to tens of millions of nodes. It also supports zero-shot super-resolution and closed-form spatial derivatives.
Vignesh Gopakumar, Zhuoyuan Wang, Ander Gray et al.
In briefChecks That Pass Can Still Miss: Testing Null Controls in Automated Scientific Discovery
SummaryThis arXiv paper examines whether automated scientific discovery systems, which accept explanations after tests, actually establish the claims they target. The authors evaluate four operations for changing physical descriptions and several interpretability settings using null, positive, and consequence agreement rates. A trivial reduction mapping every state to zero passes closure on 96% of 129 chaotic systems while retaining no dynamical variation. The authors conclude that descriptive tests can accept scientifically empty candidates and that accepted candidates may lack independently verified consequences, so both acceptance and scientific consequences must be evaluated.
In briefEvent-STVG: Spatio-Temporal Video Grounding for Long-Form Event Reasoning
SummaryThis paper introduces Event-STVG, a task that extends spatio-temporal video grounding from locating described targets to reasoning about events in long videos. It covers long context, repeated events, event localization, role inference, and multi-target grounding. The authors built a benchmark of 4,016 annotated event occurrences across soccer and seven other domains. Experiments show state-of-the-art multimodal LLMs reach only 5.4% m-vIoU at the 10-minute horizon, with frequent confusion among repeated events.
In briefWEAVE: Turning Real Code Change History into Linked Tasks for Testing and Training Coding Agents
SummaryReal software changes often span several pull requests that must be coordinated across a repository, but most existing benchmarks treat a single issue or PR as one task. WEAVE converts verified PR histories into executable tasks whose related PRs are organized into bounded dependency graphs, using typed change locations calibrated to each repository. The paper also introduces Graph Credit, a deterministic, test-based metric that measures partial progress consistent with these dependencies. According to the paper, experiments show WEAVE produces coherent executable tasks, measures progress more faithfully than flat scoring, and provides supervision and verifiable rewards for training code agents.
In briefETH: A Data-Balanced Pipeline Parallelism System for Training Large Models on Heterogeneous Hardware
SummaryThis paper addresses workload imbalance when training large models on heterogeneous hardware, where existing methods assign unequal layer counts across pipeline stages. The authors propose Heterogeneous Pipeline Parallelism (HPP) and the ETH system, which uses a hierarchical balancing framework and block-wise gradient aggregation to reduce asymmetric communication costs. ETH also reduces configuration search overhead. The authors report up to 2x training speedup and up to 56x lower configuration generation overhead compared with state-of-the-art systems.
In briefExperiment design for external data users: a usage-aware framework
SummaryStandard Bayesian experimental design assumes the experimenter alone acts on the data, but in practice external parties with unknown or black-box procedures often use the released data. The paper introduces usage-aware experimental design, which trains an adaptive design policy to minimise usage risk, the experimenter's expected loss when external rules act on the collected data. It shows that more informative data can be harmful under such rules, so the framework strictly generalises standard Bayesian experimental design.
In briefGraphACM: Graph-Based Action Consistency Measure for Uncertainty in LLM Agents
SummaryThis paper proposes a graph-based method for measuring how consistently large language model agents behave across repeated runs of the same task. Equivalent states from multiple rollouts are merged into a directed, action-labeled graph, allowing actions to be compared at shared decision states, with node-level and global consistency measures. Tested on ALFWorld, AppWorld, and WebShop with three LLMs, the global score achieves an average AUROC of 0.811 for predicting task success, above the strongest consistency baseline (0.729) and verbalized confidence (0.769).
Yaxin Ma, Wenbo Zhang, Satya Krishna Pothapragada et al.
In briefSelection Is Not Learning: Deciding Which Model Gets Inherited
SummarySelf-evolving language models train on tasks they generate, and each round must decide which updated model the next round inherits. The paper argues that current loops judge tasks only by whether they look learnable and then inherit every resulting update, though most such updates fail to make correct answers more likely. It proposes HeirCommit, which uses fresh evaluation questions to decide whether to inherit an updated model, and EvoHeir, which separates parameter inheritance from experience inheritance. The authors report that HeirCommit cuts the harmful share of accepted updates by 68-83% relative to random inheritance at the same acceptance rate.
In briefAugment the Cache, Replace the Activation: Internal-State Sharing Study
SummaryThis paper studies how frozen language models in multi-agent systems can share internal state instead of exchanging text. The authors propose a shared pool where each agent publishes its internal state once and peers read it through a single mechanism, so adding an agent does not require a new bridge per pair. They find that key-value caches should be augmented by cross-attention, while activations work best when written directly into the receiver.
In briefNot All Unlabeled Data Are Equal: Selecting Unlabeled Molecules to Improve Semi-Supervised QSAR Models
SummaryThis paper studies how unlabeled molecules from large chemical databases such as ChEMBL can improve QSAR models that predict molecular biological activity. The authors test three semi-supervised learning algorithms and four unlabeled-data selection strategies on 10 datasets, drawing from a pool of over 400,000 unlabeled molecules. At least one semi-supervised configuration outperforms the supervised baseline on every dataset, and the most reliable gains come from selecting a relatively small number of chemically distant molecules, suggesting composition matters more than size.
In briefDemoStream: Automatically generating robot demonstrations via batched task-and-motion planning
SummaryDemoStream is a GPU-accelerated task-and-motion planning system that automatically generates robot manipulation demonstrations to improve behavior cloning. It uses batched sampling and parallel constraint satisfaction to explore many solutions at once. Simulation experiments ablate action spaces, cost functions, and batch sizes to measure their effect on downstream policy success. The authors also collect real-world demonstrations at an average of 66 per hour, leveraging TAMP's safety guarantees.
