In briefRouter Drift Control: Reducing Off-Policy Errors in Mixture-of-Experts Reinforcement Learning
SummaryThis paper studies how training-inference mismatch makes reinforcement learning off-policy and causes router drift in Mixture-of-Experts models. It shows that Rollout Routing Replay (R3) introduces gradient bias and proposes Router Drift Control (RDC), which uses routing disagreement as a signal to filter high-drift responses. In asynchronous training, RDC combined with Truncated Importance Sampling improves average peak validation accuracy on AIME 2024 and AIME 2025 over TIS and R3 alone.
In briefUnified review of data strategies for training large language models
SummaryThis survey reviews how training data is selected, weighted, and scheduled across the full lifecycle of large language models. It frames these methods as policies that allocate a fixed training budget over candidate examples, and organises them by decision granularity, model dependence, and temporal adaptivity. The authors note that fine-grained or state-dependent policies offer more control but add scoring, search, or rollout costs. Evidence for transfer across model scales and comparable end-to-end costs remains limited.
In briefNo single tokenizer metric reliably predicts language model performance
SummaryThe paper analyzes about 200 language models trained on nine languages, along with their tokenizers, to test whether intrinsic tokenizer metrics predict performance across many benchmarks. None of the many proposed metrics consistently predicts downstream performance across model families, tasks, and languages. The authors argue for new tokenizer evaluation metrics and introduce a linguistically grounded metric that captures some aspects existing metrics miss, though its predictive power may vary.
Antonio Lopardo, Luis Frentzen Salim, Shikhar Srivastava et al.
In briefClassical methods for modern tabular data: a lightweight latent-variable model for generating synthetic data with continuous and discrete variables
SummaryThe paper presents a lightweight algorithm for modelling tabular data with both continuous and discrete variables. It maps the data into a continuous latent space using empirical marginals and a quantile function, where a distribution such as a multivariate Gaussian can be estimated and sampled efficiently. Because discrete observations correspond to regions rather than points in latent space, the authors formulate training as a Monte Carlo expectation-maximisation (MCEM) problem that lets conventional maximum-likelihood estimators be used.
In briefImproving Multimodal RLVR with Pairwise Token Contrast and Adaptive Trajectory Confidence (a method for more precise credit assignment in reinforcement learning)
SummaryThis paper addresses a limitation of reinforcement learning with verifiable rewards (RLVR) for multimodal reasoning, where a single sequence-level advantage is shared across all tokens even though errors tend to occur at a few key steps. The authors propose PACE, which compares each token across four teacher views (correct or incorrect hints, with or without the image) to judge whether it relies on the image, and uses adaptive trajectory confidence to set how much credit is redistributed within each response. According to the paper, PACE achieves the best average accuracy on three backbones and ranks first on all evaluated multimodal reasoning benchmarks with Qwen3-VL-8B-Instruct.
In briefEdit4D: Editable 4D Scene Generation with Text Control
SummaryEdit4D is a text-to-4D scene generation framework that builds scenes from separate background and dynamic character representations, so each part can be edited independently. A user instruction is split into stage-specific prompts, and the background is reconstructed as a 3D Gaussian splatting representation while the character is animated and reconstructed as a 4D mesh bound to Gaussian primitives. The two are composed through text-guided segmentation and contact-aware optimization, followed by a global refinement step. The paper reports that this enables controllable camera motion and flexible object-level editing.
In briefFusionAudit: Auditing Robustness of Native Multimodal Models by Attack Pathway
SummaryThis paper introduces PaRMA, an auditing framework that measures risk for vision-language models along two image-level pathways: pixel perturbation and rendered-text insertion. The authors report that native models attend strongly to typographic text yet visually reject it, and that encoder-level and answer-level attacks mismatch in deep-fusion models. They build two diagnostic-motivated attacks, MANTIS and Nat-Evid, and evaluate 18 open-weight VLMs, finding that legacy single-attack rankings are unstable for native models.
WhyProposes pathway-conditioned auditing and finds see-but-not-deceived and objective mismatch effects; evaluated on 18 open-weight VLMs with reasonably solid evidence.
In briefWeight perturbation method nearing backpropagation accuracy on MNIST
SummaryThis paper studies weight perturbation (WP), a learning method that estimates update directions from random parameter perturbations and the resulting scalar loss change, without backpropagating derivatives. On fully connected MNIST networks, a shallow configuration with large-batch training reaches about 97% test accuracy, close to an early backpropagation reference of 97.5–98%. However, under a ten-epoch budget, WP reaches 90.2% versus 97.4% for backpropagation. The authors also give a finite-horizon approximation argument, but the results depend on configuration and budget and do not establish equal efficiency or an unconditional accuracy guarantee across depth.
In briefConditional Mean Flows: Fast Few-Step Density Estimation for Simulation-Based Inference
SummaryThis paper proposes conditional mean flows for approximating posteriors and likelihoods in simulation-based inference. The method combines few-step sampling with tractable density evaluation, addressing the cost of numerical integration in flow matching when repeated density evaluations are needed for MCMC. The authors report that single-step density evaluation achieves error comparable to 200-step flow matching and enables about 200-fold faster downstream MCMC sampling.
In briefData World: Building Data Analysis Training Environments from Scratch
SummaryThis paper introduces Data World, a multi-agent engine that first specifies the mechanisms behind a dataset and then synthesizes files, tasks, and rubrics for training and evaluating data analysis agents. It releases DataWorld-Bench, with 100 tasks across 50 held-out worlds and 16 domains, and 2,135 training tasks. RL-zero training on these synthesized tasks produces DataWorld-35B-A3B, which the authors report improves over its base model on several benchmarks.
In briefDigital Typhoon Dataset V3: Multi-spectral Extension and Bias Analysis in Typhoon Intensity Regression
SummaryThis paper introduces Digital Typhoon Dataset V3, a new version of a typhoon satellite image dataset spanning nearly 50 years, for benchmarking machine learning models on long-term spatio-temporal data. The update adds multispectral satellite data. Under the current setup, multi-channel inputs do not clearly outperform single-channel inputs for pressure regression. The authors also find a systematic bias that underestimates the intensity of strong typhoons, which imbalance-aware weighting strategies do not remove.
Asanobu Kitamoto, Victor Le Louarn, Noelia Falczuk et al.
In briefThe Self-Improvement Problem: Can Automated Gates Reliably Judge Real Agent Improvements?
SummaryThis paper examines whether automated gates, such as reward models or automated reviewers, can reliably decide if a proposed update to a self-improving agent is a genuine improvement. The authors model agents as programs and show that in the fully general setting, no terminating procedure can both never accept a non-improvement and never miss an available one. They identify a bounded, decidable regime with exhaustive simulation that recovers these guarantees, but show that a chain of accepted updates can still lower the true score.
WhyProves no terminating gate can be both reliable and non-conservative in general, and delineates a decidable regime where guarantees are recoverable.
Brian B. Moser, Stanislav Frolov, Tobias C. Nauen et al.
In briefEfficient Pairwise Difference Learning: Comparing Encoding Strategies and Training Pair Selection
SummaryThis paper studies two design choices in Pairwise Difference Learning for Classification (PDC): how input features are encoded and how training pairs are selected. The authors compare four encoders and four pair selection strategies across nearly 200 benchmark classification datasets. The proposed triplet with absolute difference encoder achieves the highest average F1 score, and restricting training to upper-triangle pairs cuts computation time by about 40% with negligible accuracy loss. PDC configurations outperformed base models in 94% of tested cases, but no single configuration dominated across all datasets.
Bernard Georges, Mohamed Karim Belaid, Manuel Rodrigues
In briefSCMerge: Spatial-consistency-aware token pair merging to make semantic segmentation transformers faster
SummaryVision Transformers perform well on semantic segmentation, but their attention cost grows quadratically with sequence length, which limits practical use. Existing token merging methods pair tokens mainly by feature similarity and ignore spatial consistency, which can cause segmentation errors. This paper proposes SCMerge, a training-free method that pairs tokens using both feature similarity and spatial consistency, merges them in the encoder, and restores them to the original patch grid after decoding. Experiments on ADE20K, Pascal Context, and Cityscapes report a better accuracy-efficiency trade-off than baseline methods.
Junfeng Ren (1), Yi Gong (1), Yuan Zeng (2) ((1) Southern University of Science et al.
In briefReFlow: Replay-Driven Decoupling for Faster RL Post-Training of Language Models
SummaryThis paper presents ReFlow, a reinforcement learning framework that jointly co-designs system and algorithm to make post-training of large language models more efficient. It reuses recent trajectories through experience replay, amortizing the cost of rollout generation without causing excessive policy staleness. The framework combines a dual replay data cycle, an elastic resource pool, and an adaptive scheduler. In experiments on single-turn reasoning and multi-turn agentic RL tasks, ReFlow improves training throughput by up to 2.06 times while keeping model performance comparable or better.
In briefDIAL: Adjusting a Weather Model's Built-in Randomness to Make Forecast Ensembles More Reliable
SummaryThis paper introduces DIAL, a method that adjusts the built-in randomness of a frozen weather model without retraining it. DIAL learns from past forecasts whether each variable and spatial scale needs more or less spread, and makes small noise updates at each forecast step. Tested on FourCastNet 3 with ERA5 years unseen during training, DIAL improves the multiscale continuous ranked probability score (CRPS) by 8.27% over the unmodified model and gives better scale-resolved reliability than learned postprocessing and reward-based control baselines.
In briefHidden shortcut in vision transformers: attention heads reading chest X-ray annotations instead of lungs
SummaryThe paper uses mechanistic interpretability to find an attention head (L8H10) in ViT-Base 16 that focuses on corner text annotations rather than lung pathology on chest X-rays, in 62% of 200 COVID-19 images. Inpainting experiments indicate annotation content drives this behaviour, and lung attention rises 1.52 times when it is removed. The pattern also appears in NIH ChestX-ray14, DeiT-Base and ViT-Large, and full fine-tuning does not remove it. The authors report that output-level tests fail to detect this circuit, so mechanistic auditing can reveal problems that behavioural evaluation misses.
In briefAnchor3DGS: Compact Anchor-Based Feed-Forward 3D Gaussian Splatting for Scene Reconstruction
SummaryThis paper presents Anchor3DGS, a pose-free feed-forward 3D Gaussian Splatting framework for reconstructing 3D scenes from multi-view images. Instead of predicting one Gaussian per input pixel, it places a compact set of 3D anchors guided by information entropy, concentrating them in visually complex regions. The authors report state-of-the-art novel view synthesis quality on diverse benchmarks while using an order of magnitude fewer Gaussian primitives.
In briefContextLift: Removing the Spurious Link Behind Conditional Emergent Misalignment
SummaryFinetuning a language model on narrowly misaligned data can make it broadly misaligned, and mixing in generic safety data helps but fails when prompts resemble the finetuning context. The authors attribute this to a spurious correlation and propose ContextLift, which duplicates half of the safety examples with a context sentence taken from a random training prompt while keeping the aligned answer. On Qwen2.5-7B, misaligned answers under a domain cue drop from 18.9%, 41.9% and 58.2% to 2.6%, 6.1% and 12.8%.
In briefTemporal contrastive training gives human motion embeddings an ordered geometry
SummaryThe paper proposes SemTAC, an encoder trained with a temporal-neighborhood contrastive loss so that representations of human motion sequences keep their global temporal structure. On a five-step motion sequence discrimination task, it outperforms four baselines on truncated sequences, reaching a median F1 of 0.77 one step before completion versus 0.27 to 0.70 for the baselines. The authors also introduce a displacement-smoothness metric showing its representation trajectories have interpretable geometric structure.
In briefAgentPerturb: Testing how well AI agents cope with controlled disruptions to their inputs and interactions
SummaryThis paper introduces AgentPerturb, a framework that evaluates LLM agent robustness along two dimensions: perturbation targets and perturbation types. Controlled experiments on APEX-Agents and Toolathlon show that perturbations consistently degrade performance, with the main vulnerabilities varying by task environment. Trajectory analysis finds that agents rarely notice altered inputs and often pass corrupted information along without verification.
In briefTraining-free target-aware refinement of a 3D Gaussian language map
SummaryThis paper proposes a training-free method to refine an already-distilled 3D Gaussian language map. Using a text-promptable segmenter, it blends each detected target region's visual embedding with the target's text embedding via a convex combination, leaving non-target regions unchanged, and computes the smallest sufficient coefficient in closed form. It also identifies that injected text embeddings make cosine readout query-agnostic and fixes this with a centering step. Experiments on three benchmarks show improved open-vocabulary segmentation and more reliable language-goal localization, and the method works as a plug-in for other language maps, with a real-world legged robot navigation demo.
WhyTraining-free target-aware blending with closed-form coefficients and centering is a clear new idea, validated on three benchmarks and a real legged-robot demo, roughly doubling Replica mIoU.
In briefA decision-aware protocol for judging uncertainty in medical image segmentation
SummaryThe paper argues that common metrics for medical segmentation uncertainty, such as calibration error and pixel-level error detection, do not show whether uncertainty improves downstream review decisions. The authors propose an evaluation protocol that separates uncertainty estimation from deferral policy and tests their combination under fixed review budgets. Experiments on retinal vessel, skin lesion, and polyp segmentation show that estimator rankings are not universal, and that boundary and thin-structure failures are partly tied to annotation ambiguity.
WhyProposes a decision-aware protocol evaluating uncertainty with deferral under fixed budgets, validated across datasets, backbones, and estimators with robustness checks.
In briefQwen-Audio-Jev: A Unified Interface for Audio Decisions
SummaryThis paper introduces Qwen-Audio-Jev, which adds a dedicated Jev decision head to Qwen-Audio-3.1-ASR-Flash-Next so it can make audio decisions without generating text token by token. It unifies selection, binary judgment, and scoring through three decision primitives: CHOICE, NOUL, and SCORE. Evaluated on 18 benchmarks, it reports 78.40% accuracy on MMAU and 83.14% on MMSU for CHOICE, 93.76% on VoxCeleb1-O for NOUL, and a normalized mean absolute error of 0.1235 on NISQA for SCORE.
In briefLearning Intrinsic Dynamics in Latent World Models: Using Action Predictability to Preserve Control-Relevant Information
SummaryThe paper argues that encouraging an encoder to keep distinctions between actions and their consequences improves latent planning under visual variation. It proposes Concat IDM, which predicts actions from consecutive latent states using action labels from offline trajectories, without changing the planner. On Push-T across 24 appearances, reported planning success rose from 9.1% to 77.2%.
In briefTRACE: Communication-efficient distributed Transformer inference on edge devices using codebook lookup
SummaryThis paper addresses the communication bottleneck in running Transformer models across several co-located edge devices. It proposes TRACE, which places a vector quantizer after each layer's attention residual and precomputes codebook-indexed key and value representations, so receivers recover them by table lookup. Experiments on ViT-Base and ViViT-B report up to 1.89% higher accuracy and 1.27x speedup over a prior VQ-based method, and up to 22,624x lower communication overhead than uncompressed sequence-parallel inference.
Muhammad Azlan Qazi, Alexandros Iosifidis, and Qi Zhang
In briefStudy of how market profit incentives shape AI safety spending
SummaryThis paper examines how firms' profit-seeking affects investment in AI safety within a supply chain of an upstream LLM provider and several competing downstream firms. The authors build a game-theoretic model and compare the market-driven safety investment level with levels that maximize market welfare and industry profit. They find downstream competition generally leads firms to invest more in safety than the industry-profit-optimal level, while the upstream provider may underinvest because it is not fully compensated for safety surplus. The paper also analyzes how different regulatory interventions affect participants' incentives, profits, and resulting model safety.
In briefTesting how well AI language models can run a full scientific inquiry loop
SummaryThis paper introduces a unified, interactive framework for evaluating large language models on scientific inquiry. The authors extend the Countdown task into a math environment where an agent manipulates variables, collects observations, and infers relationships. Across seven LLMs and multiple prompting strategies, models showed limited inherent inquiry ability and low success in finding the underlying expressions. Structured guidance inspired by educational scaffolding significantly improved performance.
Marta Kne\v{z}evi\'c, Seyed Parsa Neshaei, Jibril Frej et al.
In briefWaypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation
SummaryThis paper studies how to use the change in a teacher model's behavior before and after reinforcement learning to train a student model. The proposed Waypoint-OPD method uses an intermediate checkpoint from the same RL run to compare early and late shifts in action preferences, and suppresses the endpoint shift when the two point in opposing directions. Experiments with Qwen3-4B teachers and several student models on math and code reasoning report the highest domain-average scores in all six settings compared with other distillation methods.
In briefDecoding iOS Location Traffic to Apple: LLMs Make Unreliable Analysts but Useful Tool Builders
SummaryThe paper examines whether AI can help reverse engineer the protobuf formats that iPhones send to Apple's location services. Two frontier models reconstructed schemas with substantial, hard-to-detect errors, so direct AI use is judged unsuitable for forensic and privacy analysis. The authors instead used Claude to write Extractor, a tool that recovers schemas from Apple binaries, validated against known test cases and measurements. They then used Extractor to decode the location traffic of a jailbroken iPhone 8 running iOS 16.7.11.
In briefLoopAhead: A Speculative Decoding Framework for Faster Looped Language Models
SummaryThis paper presents LoopAhead, a speculative decoding framework for looped Transformers that repeatedly apply the same layers. A lightweight drafter uses early-loop states to predict multiple future tokens, while deeper-loop feedback decides which draft branches to continue or pause. The full-depth model makes the final token decisions. The authors report up to 7.82x decoding speedup over autoregressive decoding while preserving the target model's output distribution.
In briefLOADBench: An Open Analog Dataset and Benchmark for ML in IC Design
SummaryThis paper addresses the lack of open data and standardized evaluation for machine learning in analog IC design. It releases an open dataset of over 100 operational amplifier topologies on the SkyWater130 PDK, with structural and functional annotations, sizing constraints, and NSGA-II sizing results. The authors also define six benchmark tasks with precise protocols and release reference implementations of the metric routines under an open license.
In briefHear, Localize, and Reason: Spatially Aware Audio-Visual LLMs
SummaryThis paper notes that most audio-visual large language models treat audio as mono semantic content, which limits their ability to tell where sounds come from and how they relate to the scene. The authors introduce HLR-AVSceneQA, a benchmark that tests recognizing sounds, localizing sources, and reasoning about them. They also propose HLR-LLM, which adds a binaural spatial audio branch and is trained with a curriculum including chain-of-thought supervision. Experiments show existing models are limited in spatial grounding and relational reasoning, while HLR-LLM substantially improves spatial audio-visual understanding without hurting semantic perception.
In briefAVALON: Zero-Shot Indoor Navigation with Human-Seeded Topological Memory
SummaryThis paper presents AVALON, a topological navigation framework for indoor service robots operating in unmapped buildings. It needs only a floor plan, place clues, and an initial alignment, with humans and the robot jointly updating layered spatial memory. In experiments on a Unitree Go2, AVALON with X-NavDP completed a 12.81 m mission in 46.4 s in all 5 runs, 40.9% less robot time than DimOS-Go2, while policy-only X-NavDP timed out in all runs.
In briefBandit transfer learning from unknown adaptive data: problem-dependent bounds
SummaryThis paper studies transfer learning in multi-armed bandits where historical data comes from an unknown, possibly adaptive decision process rather than a fixed dataset. The authors derive an asymptotic problem-dependent lower bound and propose KL-UCB-Transfer, which combines online and historical data. It achieves logarithmic regret with an improved constant under a specific condition, and otherwise matches classical KL-UCB. The paper also analyzes a relaxed one-sided misspecification condition, identifying three regimes.
Adrien Prevost, Timothee Mathieu, Waris Radji et al.
In briefClaimScope: A Label-Blind Benchmark for Auditing Mathematical Claims
SummaryThe paper introduces ClaimScope, a 72-task benchmark that audits proposed mathematical claims through a verdict, scope, counterexample, and repair. Tasks cover optimizer-state transport, finite-horizon affine recurrences, and exact 2×2 linear algebra. Using Qwen3-8B, the authors compare direct generation with matched two-pass revision, where a witness checker verifies only the submitted state. Verifier feedback yields more accepted counterexamples and safe completions than checker-free feedback, but strict completion stays at 17/72 and verdict accuracy barely changes.
In briefLearning to search for editing steps that turn one sewing pattern into another
SummaryExisting models generate executable sewing patterns from images or text but do not model how a patternmaker edits one pattern into another, and pattern archives record only the results. SemDraft recovers editing programs from paired before-and-after patterns through bounded search, accepting a program only after it recompiles, its action pairs are classified as ordered or interchangeable, and an independent replay reaches the same pattern. Certified programs train a learned ranker for candidate actions. On two frozen draws of 80 cases, the learned ranker certified 63 and 64 cases within 1,000 search executions, versus at most 53 and 52 for non-learned rankers.
In briefMYOPIA: Limits of Tree-Based Reasoning in Vision-Language Models
SummaryThis paper introduces MYOPIA, a benchmark with four LLM-generated, human-refined decision trees for GTSRB, MNIST, CIFAR-10, and MSCOCO, plus prompts, diagnostic protocols, and an evaluation harness. Across 12 dataset–model pairs, tree-based traversal underperformed zero-shot prompting in every case. The authors analyze failure modes, such as models recalling class attributes more accurately than grounding them in images, and suggest designing trees that prioritize perceptual answerability and recovery from early errors.
Islam Mesabah, Sary Elmansoury, Gerrit Gro{\ss}mann et al.
In briefRethinking Tool Graph Memory in Multi-Turn Tool Use
SummaryLLM agents solve complex tasks by calling external tools over multiple turns, and recent work proposed tool graph memory, which stores past experience as tool transitions to guide later choices. This paper evaluates the approach on nine models and two benchmarks, finding it rarely improves significantly over baselines, and where gains appear they cost over twice the LLM calls. The memory's guidance is often wrong, and random or no guidance keeps most of the gain, suggesting the benefit comes from extra computation. Storing the same experience as textual insights performs better with almost no extra LLM calls, indicating memory works better as a soft prior than as a prescription of the next tool.
In briefIndependent data-integrity checking for LLM-driven coverage closure in hardware design
SummaryThe paper presents a small agent in which an LLM writes directed tests to close functional coverage in hardware designs. Coverage is read from the simulator's exported report, while data integrity is checked by scoreboards separate from the model. Across three RTL designs, the agent reached 100% coverage within eight iterations and a monitor found a clock-domain-crossing bug in the authors' own stimulus. The monitor is not yet applied to the LLM-generated tests, and the evidence is limited.
In briefDecoupled Safety Control: A Training-Free Algorithm for Safer Diffusion Image Generation
SummaryDiffusion models can generate unsafe images even though they produce strong text-to-image results. Existing training-free safety guidance methods use one unsafe condition for several roles at once, which can be unreliable when harmful and benign content are mixed. The paper proposes Decoupled Safety Control (DSC), a plug-in algorithm that decides the intervention from the current generation state. Experiments on standard safety benchmarks report stronger suppression of unsafe content while preserving benign-generation quality and inference efficiency.
In briefFrom Perception to Prediction and Intervention: A Benchmark for Proactive Video Assistants
SummaryThis paper introduces P³IBench, a benchmark for evaluating proactive video understanding in real-life streaming scenarios across daily perception, risk prediction, and intervention. The authors test 12 open-source and proprietary video-language models and find that current models struggle to make well-calibrated decisions about when to respond. They also train P³I-Assistant-4B on additional proactive training data, which improves several tasks, though gains are uneven across the three capabilities.
In briefSharp, stable label-free certificates that show a predictor is poorly calibrated
SummaryThis paper studies calibration for query-conditioned probabilistic predictors, which answer logically related questions one at a time when labels for checking calibration are often missing. A minimax theorem gives, for each restriction on betting stakes, a sharp label-free lower bound on the calibration residual, and the authors show the marginal bound is discontinuous in the answers while Lipschitz stakes yield a stable bound. On i.i.d. streams the paper builds fixed-horizon and anytime-valid certificates, and a preregistered study of four predictors on two data sets reports positive results for seven confirmatory certificates. A certificate only shows calibration is at least this bad, and zero means not refuted.
WhyGives a minimax label-free calibration lower bound with stability results and online certificates; novel theory, with a modest preregistered empirical study.
In briefA Quantum Machine Learning Approach to Content-Based Media Recommendation
SummaryThis paper develops a content-based TV programme recommender using variational quantum circuits built on the Quantum Natural Language Processing framework. Programme synopses are encoded into circuits and trained with a pairwise similarity task on one week of BBC iPlayer data. The quantum model performs comparably to a classical BiLSTM while using far fewer parameters, and the trained circuits were also run on Quantinuum's H2 hardware and emulator using a SWAP-test protocol.
Mina Abbaszadeh, Ouissal Moumou, Anna Pearson et al.
In briefDialFork: Measuring Whether Chatbots Follow the Right Conversation Steps, Not Just Facts
SummaryThis paper identifies procedural hallucination, where a conversational agent produces fluent, factually plausible replies while deviating from the intended dialogue process. The authors release DialFork, a benchmark of 36,278 dialogues built by injecting four hallucination types into conformant conversations, validated by 30 annotators. They propose dialog-flow-aware metrics, FlowKL and FlowScore, and report that FlowScore significantly outperforms all baselines across seven encoders at the dialogue level.
WhyIntroduces a new benchmark and flow-aware metrics for procedural hallucination, with FlowScore beating baselines across 36,278 dialogues and seven encoders.
Sergio Burdisso, Esa\'u Villatoro-Tello, Dairazalia S\'anchez-Cort\'es et al.
In briefWhen Does the Muon Optimizer Actually Win? A Budget-Matched Benchmark Against AdamW
SummaryThe paper benchmarks a reference single-device Muon optimizer against AdamW on one A100 GPU, using 17.7M- and 33.6M-parameter language models and 2.70M-parameter vision transformers. Under matched data, Muon reduces validation-loss area under the curve by 3.25% to 22.44% and lowers test loss in all 20 primary pairs. Time-normalized results favor Muon on vision (19.68% to 22.31%) and AdamW on language (3.38% to 6.96%). The authors note these findings apply to small models with reference kernels and do not establish rankings for larger models or optimized implementations.
Mushir Akhtar, Hazem N. Nounou, Majdi Mansouri et al.
In briefUncertainty quantification for a pairwise difference learning classifier
SummaryThe paper proposes pairwise difference learning (PDL) for classification, where a model takes two instances and predicts the difference in their outcomes, and predictions for a query are averaged over training examples. It introduces a suitable meta-learning technique for this setting. The authors extend existing decompositions of total uncertainty into aleatoric and epistemic components to PDL. Large-scale experiments analyze both predictive accuracy and uncertainty awareness.
Mohamed Karim Belaid, Maximilian Rabus, Eyke H\"ullermeier
In briefEviGround-CT: Matching LLM-Written Chest CT Findings to 3D Scan Regions for Verification
SummaryThis paper proposes EviGround-CT, a method for checking chest CT reports written by large language models. It extracts positive finding sentences and uses a 3D grounding model to predict an evidence mask for each sentence, linking statements to locations in the scan. On a held-out 100-case subset, sentence-grounded verification raised evidence-supported precision from 0.263 to 0.365 compared with category conditioning, and reduced the unsupported positive finding rate from 0.737 to 0.635.
In briefWritability and the Pathway: a controlled study of injecting knowledge into addressed memory
SummaryThe paper studies post-hoc knowledge injection into addressed conditional memory. It runs a 2x2 controlled comparison on frozen GPT-2 and production-memory experiments on Qwen and DeepSeek models. Successful writes do not necessarily transfer to new questions or preserve untargeted behavior, and first-token rank gains do not establish complete-answer success.
In briefRethinking Classifier-Free Guidance in TI2I Flow Matching
SummaryThis paper studies text-and-image-to-image flow matching methods for controllable image editing. The authors observe that classifier-free guidance, a standard component for output quality, implicitly hurts editing during the noisy phase of flow modeling, weakening instruction following and controllability. They propose a simple correction in the early flow steps that yields a better alignment-preservation tradeoff at no extra cost. The approach is shown to be effective on both pixel-space and latent TI2I flows.
Andreas Floros, Elena Koritska, Teofilo de Campos et al.
In briefWhen Is Muon Optimal? Radial Fragility and Minimax Geometry
SummaryThis paper examines whether the curvature-based explanation of the Muon optimizer still holds when the fixed input radius is replaced by a general radial law. It introduces the Random-Radius Isotropic Curvature Model and an admissibility set for polar factors. The authors show that for atomless radial laws, exact scaled-polar optimality is possible only when the gradient is already a scaled column isometry. They also propose a minimax version and prove that nonzero global minimizers lie on the gradient's polar ray.
WhyIntroduces a random-radius curvature model and minimax characterization giving a new theoretical view of Muon, with rigorous proofs of substantial scope.
In briefFinite-time guarantees for constrained reinforcement learning with hidden states
SummaryThis paper studies reinforcement learning in constrained partially observable Markov decision processes, where an agent sees only incomplete information about the true state while respecting a cost limit. The authors model finite windows of recent history as states of a constrained MDP and propose a primal-dual algorithm. It gives the first non-asymptotic bounds on both reward suboptimality and constraint violation for the original problem. Under a filter-stability assumption, approximation error decays geometrically with window length, and experiments on partially observable gridworlds support the approach.
Valentina Zaccaria, Philip Jordan, Tingting Ni et al.
In briefReinforcement learning to refine diffusion models across language and vision
SummaryThis paper presents the first comprehensive review of reinforcement learning (RL) used to post-train diffusion models. It proposes a technical overview of RL's roles in diffusion-based language, multimodal, robotic, and scientific models, and evaluates representative methods on task-specific benchmarks. The authors identify three bottlenecks to real-world deployment: fragile reward design, the exploration-exploitation trade-off, and training instability under sparse rewards, and they outline future research directions.
In briefAgent Status Protocol (ASP): A lightweight way for AI agents to report their state
SummaryThis paper proposes version 0.1 of the Agent Status Protocol (ASP), a lightweight specification for heterogeneous AI agents to report operational state to a status server. Agents push small JSON events over HTTP POST, and observers receive aggregated state via Server-Sent Events. The protocol does not rely on self-reports alone; it infers liveness from missed heartbeats and accepts observations from supervisors and other parties. It is presented as a working draft, not an approved standard, to invite community comment and implementations.
In briefLocalizing Functional Subnetworks for Efficient Continual Learning in LLMs
SummaryThe paper addresses catastrophic forgetting when large language models learn continually. Using gradient-attribution profiling, the authors find functional subnetworks with very low mutual overlap and propose CiMEO, which routes each training example to its matching subnetwork and updates only that region's parameters and optimizer states. Across 1.4 million sequential instructions, CiMEO achieves the highest average performance on five generative benchmarks while mitigating forgetting.
In briefActivation-adjacent normalization and how wide networks must be to approximate functions
SummaryThe paper studies whether shallow networks with layer normalization (LN) or RMSNorm placed next to activations can approximate a given width-N one-hidden-layer MLP. The authors use width augmentation to locally linearize normalization on compact feature sets and derive width and error bounds for different activation settings. They prove that width N+1 for RMSNorm and N+2 for LN are worst-case bounds, and numerical experiments confirm the predicted approximation rates.
In briefCriticaLLM: Using LLMs as Autonomous Game Masters for Tabletop Role-Playing Games
SummaryThis paper presents CriticaLLM, a framework in which large language models act as a fully automated Game Master that runs entire tabletop role-playing games. It covers creating stories and managing exploration and combat phases. The authors evaluate the framework with real TTRPG players and compare the diversity of generated plots and antagonists across four LLMs of different sizes.
Ga\"etan Berlaimont, Justin Vanwichelen, Beno\^it Ronval et al.
In briefBOLT: Bound-guided learning for job-shop scheduling with transport time lags and shared vehicles
SummaryThe paper proposes BOLT, a learning method for flexible job-shop scheduling where a limited vehicle fleet moves jobs, so the schedule determines transport waits. It builds a lower bound from chain, machine-capacity, and fleet-capacity relaxations and uses the bound's per-step changes as a reward. On industry-motivated prefabricated-construction instances, the authors report gains over a transport-penalty policy and dispatching-rule pairs, with decisions taking 2.3 to 5.9 ms per event.
WhyProposes bound-derived rewards for resource-coupled transport scheduling, with comparisons against dispatching rules and a solver on industry-motivated instances; impact is mainly within flexible job-
In briefHierarchical Interpretable Multitask Audio Representation Learning: Event-Scene Relationships via the Information Bottleneck
SummaryThis paper studies joint acoustic scene classification and audio tagging, proposing a hierarchical method that predicts scene categories from intermediate event probabilities instead of a conventional multihead design. The authors ran 66 full training runs across three backbones (PaSST, PANNs, BEATs) under varying loss weightings. Results show competitive classification and synchronized dual-task learning at certain loss weights, with the intermediate feature space showing lower conditional entropy, higher mutual information, and better cluster separability.
Claudia Montero-Ram\'irez, Alba Mart\'inez-Serrano, Carmen Pel\'aez-Moreno
In briefSCRIBE: A Scalable Method for Interpretable, Context-Aware Brain Semantics
SummaryThe paper introduces SCRIBE, a framework that uses an ensemble of open-source large language models to produce interpretable, word-level semantic representations of arbitrary text. In a behavioral study with 921 participants, human semantic judgments shifted with context, and SCRIBE captured these shifts, reaching 75-81% of the estimated human noise ceiling. Using brain recordings from participants listening to natural stories, the authors show SCRIBE predicts activity in language-selective regions better than non-contextual representations and explains unique variance beyond syntactic features.
Camila Kolling, Omer Moussa, Emin \c{C}elik et al.
In briefPoison Valleys: Diagnosing Where LLM Code Repair Goes Wrong
SummaryThis paper examines why code-repair agents produce wrong patches, since final tests show failure but not which earlier decision caused it. The authors trace 1,844 model requests across two synthetic Python tasks, PersistQueue and Dijkstra, and evaluate outputs with an Oracle and diagnostic coordinates. Moving a misconception from the hypothesis to the implementation plan increased failure by 32.2% in PersistQueue and 28.9% in Dijkstra.
In briefQREF: Improving DeepSearch Agents by Aligning Search with Reasoning Through Query Refinement and Evidence Filtering
SummaryDeepSearch agents perform complex information-seeking tasks through repeated searching and reasoning, but their queries can be misaligned with intent, under-constrained, or redundant, which weakens evidence coverage and adds retrieval noise. The paper proposes QREF, a plug-and-play Query Refiner and Evidence Filter: the refiner rewrites queries based on current reasoning intent and keeps each search batch complementary, while the filter keeps evidence relevant to both the batch and the reasoning goal. The refiner is initialized by supervised fine-tuning and then optimized with reinforcement learning in a real-world web search environment. Across four benchmarks, QREF improves average relative Pass@1 by 33.55% on WebSailor and 11.04% on TongyiDeepResearch without changing architectures or retraining reasoning policies.}}
In briefProV-Mem: Proactive Memory for Streaming Video Question Answering
SummaryThis paper proposes ProV-Mem, a training-free memory framework built on a frozen vision-language model, enabling a system to decide when to respond, what to say, and whether to emit a response while monitoring streaming video. The proactive decision is split across three lightweight memory tiers: temporal-trigger, instructional, and episodic memory. The authors also introduce CAUC, a cost-aware extension of PAUC that penalizes redundant, overly frequent, and overly long responses. Experiments on four ProactiveVideoQA scenarios are reported.
In briefBreaking down how brain activity aligns with LLM representations using interpretable meaning features
SummaryThis paper examines how the language model LLaMA2-7B predicts brain responses during natural story listening. The authors use SCRIBE, an interpretable semantic representation with 65 neurocognitively grounded properties, and apply variance partitioning to fMRI data from seven participants listening to narrative stories. They find that variance shared between the two representations dominates, while both also retain unique predictive contributions, and concrete properties share more variance with LLaMA2-7B than abstract ones.
In briefFilterFusion: Prior-Guided Adaptive Filter Fusion for Fast Video Denoising
SummaryThis paper presents FilterFusion, a multi-frame video denoising framework that adaptively blends a compact set of interpretable denoising filters using a low-cost guidance signal. The model has only 109K parameters and reaches 34.72 dB PSNR and 0.9330 SSIM on the CRVD real-sensor-noise dataset. A compiled C implementation reportedly processes 1080p frames in 17.9 ms on one CPU thread.
In briefISW-Agent: Linking LLM Dialogue and n8n Workflows for Internal Solitary Wave Image Processing
SummaryThis paper presents ISW-Agent, a system that connects large language model interpretation of natural-language requests with n8n workflows and independent scientific models for processing remote sensing images of internal solitary waves. In tests on 120 Chinese instructions per backend, both GPT-5.6 Sol and DeepSeek reached a strict task-match rate of 94.2%. Of 50 timed scenes, 39 completed detection, extraction, and prediction with a median time of 48.04 s, while 11 stopped after empty extraction.
Hao Zhang (1), Hengrui Liu (1), Lina Sun (1) et al.
In briefWhen the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel Prefill
SummaryThis paper studies a remote KV cache tier shared across engines in a context-parallel prefill deployment of a large MoE model on 8 B300 nodes. Enabling the remote tier raised the 16k-token p99 latency from 3.96 s to 17.3 s, because remote hits were slower than recomputing the prompt. The authors propose a two-term cost model and apply sharding, per-rank partitioning, asynchronous loading, and client bug fixes. These changes cut p99 on a hit-heavy stream to 0.67 s and raised prefill throughput by 14% to 34%.
In briefGRASE: A video-based agent framework for building editable, geometry-faithful 3D scenes from real environments
SummaryThe paper presents GRASE, an agentic framework that converts multi-view video of a real environment into an editable 3D scene. It uses the captured video as a constant reference and revises its current hypothesis whenever rendered output disagrees with the capture, refining only the components responsible for discrepancies. According to the authors, the resulting scenes are faithful to the input, consistent across views, and suitable for simulation, and GRASE substantially outperforms other state-of-the-art methods in their experiments.
In briefOPRD: On-Policy Reminder Distillation for Long-Horizon Coding Agents
SummaryThis paper introduces OPRD, a training method that helps coding agents keep constraints and respond to tool feedback over long interactions. A teacher model re-evaluates the student's own rollouts with reminders, while the student runs without them at inference. On SWE-bench Verified, SWE-bench Pro, and SWE-bench Multilingual, OPRD improves the benchmark-averaged resolution rate by 11.3 points over the Qwen3.5-4B base model and reduces error events from 446 to 222.
In briefPhase Kernels Boost Dense Associative Memory Capacity
SummaryModern Hopfield Networks with linear kernels suffer overlapping attractor basins and spurious states under high memory load, limiting their theoretical exponential storage capacity. The paper constructs nonlinear phase kernels that modify the energy landscape and retrieval dynamics, and proves they preserve exponential capacity and local fixed-point convergence under stated conditions. The resulting Sigmoidal Phase Hopfield Network (S-Hop) guarantees monotone energy descent and mitigates practical capacity loss.
In briefMetaExp: Training AI Agents to Learn Reusable Lessons from Experience Across Tasks
SummaryThe paper introduces MetaExp, a framework that trains language-model agents to turn interaction feedback into reusable experience, using verifiable task families with controlled cross-task structure. It proposes Shared-Tool and Shared-Workflow task designs, and fine-tunes Qwen3.5-122B-A10B on 4,956 trajectories from 1,000 task families. With parameters frozen during downstream evolution, the authors report gains of +8.5 on SpreadsheetBench, up to +2.3 on APEX-Agents, and +14.8 on AutomationBench.
In briefFlashAttention Statistics: Low-Cost Attention Features for Hallucination Detection in Small Models
SummaryLarge language models can give fluent answers that the supplied context does not support. The paper uses a compact frozen proxy model that reads the context and a response to detect such hallucinations, computing attention-based summary statistics without materializing the full attention matrix. Statistics are obtained during the FlashAttention pass, via sparse lookups, or with a second tiled pass, reducing memory use. On three benchmarks, probes on Qwen3 proxies perform strongly, and adding attention features significantly improved F1 in half of the 12 benchmark-proxy pairs.
Andrei Volodichev, Maksim Savichev, Nikita Sviridov et al.
In briefAI for Disaster Response and Tracking Displacement
SummaryThis paper reviews how artificial intelligence is used across humanitarian tasks, such as climate forecasting, poverty prediction, and deforestation monitoring. It notes that AI also supports disaster response and conflict monitoring by assessing damage and urgent needs. The paper highlights AI's role in tracking displacement.
In briefKolmogorov-Arnold Networks (KANs): Curse of Dimensionality and Structural Bias explained
SummaryThis paper studies how the approximation ability of Kolmogorov-Arnold networks (KANs) changes as input dimension grows. For standard spline-KANs on isotropic Sobolev balls, the authors prove that reaching a dimension-independent accuracy requires a number of spline coefficients that grows at least exponentially with input dimension. The paper argues that favorable scaling depends on target structure and a compatible inductive bias, not solely on the KAN architecture, and compares results with MLPs across several experiments.
WhyProves exponential dimensional lower bound for spline-KANs on Sobolev balls, with controlled experiments locating favorable scaling in target structure.
In briefTool-augmented LLMs for reliable medical calculation from patient notes
SummaryThis paper proposes a framework that pairs large language models with executable tools to perform medical calculations from free-form patient notes. The system builds calculator tools from retrieved specifications and uses a constrained extractor to map notes to calculator inputs, so the tools perform the arithmetic deterministically. On the MedCalc-Bench dataset it reports 86.73% final-answer accuracy, outperforming one-shot prompting and supervised fine-tuning baselines.
In briefUsing a different prompt at inference recovers rare-class recall in imbalanced VLM finetuning
SummaryWhen vision-language models are finetuned on class-imbalanced data, recall for rare classes can drop well below zero-shot levels. The authors find that reusing the descriptive finetuning prompt at inference can be worse than switching to another descriptive prompt, and call this strategy cross-prompt inference. Across six datasets and diverse VLMs, it consistently recovers rare-class recall, with varying effects on frequent-class recall.
Emanuel Ben Baruch, Sivan Biham, Alon Shoshan et al.
In briefCan LLMs Tell Stories Out of Order? A Nonlinear Narrative Benchmark
SummaryThis paper introduces a benchmark of 83 published works for testing whether language models can tell nonlinear stories. Each work is annotated with its events in both story and telling order, plus a story graph of causal dependencies, character knowledge, and narrative questions. Models tend to tell events more chronologically than published works do, and when asked for a specific structure they match the human range about half the time. In a study with 71 readers, both readers and an automated judge preferred the chronological stories models write by default over nonlinear versions written on request.</summary>
Syed Rifat Raiyan, Md. Tariquzzaman, Md Farhan Ishmam et al.
In briefDuET: A dual-encoder model that turns piano performances into sheet music by tracking notes and beats together
SummaryDuET is a Dual-Encoder Transformer that jointly learns note-level performance MIDI-to-score transcription and frame-level beat and downbeat tracking. It adds TimeRoPE and Time-Pitch Rotary Positional Encoding (TPRoPE) to encode shared timing and relative pitch and time relationships. On the ASAP and Gangqinpu datasets, it reduces MUSTER mean error from 6.56% to 3.64% on Gangqinpu and from 11.30% to 11.22% on ASAP compared with the strongest baseline on each.
In briefRefining Ontologies with Embeddings and LLMs: A CASTON Case Study
SummaryThis paper addresses subsumption errors in class hierarchies of CASTON, a casting-domain ontology, using a lightweight, scalable workflow. A fine-tuned ontology embedding model extracts a limited set of candidate axioms and possible replacements, and a large language model then proposes repairs. In a small expert evaluation, the default configuration (CASTON-fine-tuned OnT embeddings with Qwen3-8B) produced six high-confidence replacements, all accepted by both domain experts.
In briefSpecAgent: A Method Where a Strong Model Reviews a Weak Model's Tool-Use Drafts
SummaryThis paper introduces SpecAgent, a protocol in which a small model drafts multi-step tool-calling sequences and a strong model verifies them in one parallel pass, keeping the longest valid prefix and repairing the first rejected step. The authors report that it nearly matches the strong model's success rate on tau²-Bench while roughly halving memory traffic and wall time. They also report reduced quality gaps and lower token generation by the strong model in other benchmarks and an edge-cloud setup.
WhyProposes strong-model verification of weak-model drafts with parallel prefill, tested on several tool-use benchmarks and an edge-cloud setup, showing quality retained with reduced latency and memory t
In briefCan image generation training help multimodal AI models answer text questions better?
SummaryThis paper studies whether training unified multimodal models to generate images can expand the knowledge they can access through text-only question answering. The authors introduce Knowledge Transfer Probing (KTP), a controlled protocol, and test seven knowledge categories across three models. They find transfer from visual supervision to language is direct but uneven, and that textual and interleaved image-text alignment can improve knowledge access. The authors suggest visual generation can be a knowledge source beyond textual supervision.
In briefAdaptive-horizon graph neural operator for fast riverbed change prediction
SummaryThis paper proposes an Adaptive-Horizon Graph Neural Operator (AH-GNO) that predicts future bed elevation changes directly on unstructured meshes from recent flow and topography states. It is tested against Telemac2D/GAIA reference solutions on the Bump and Yen bend benchmarks and a field-scale Waiho River case in New Zealand. The authors report that it reproduces key patterns such as bedform migration and bend erosion and deposition. They report bed-update workflow times falling from 12 to 1204 s with GAIA to 0.08 to 3.34 s with AH-GNO, depending on the deployment.
In briefFinding and fine-tuning the internal circuits that handle low-resource multilingual translation
SummaryThis paper studies the internal translation circuits of multilingual decoder-only language models for Hindi-to-Urdu and Hindi-to-Sindhi translation. It proposes TITAN, a method combining causal intervention, sparse feature decomposition, and targeted preference optimization, which updates fewer than 5% of model parameters. Reported results show BLEU gains of 33.75% and 45.86% over the strongest multilingual baselines, with human evaluation also showing improved adequacy and fluency.
Atai Waris Khan, Abdullah Mazhar, Raghu Kumar et al.
In briefCarryOn: A Dual-Loop Framework for Personalized, Reliable Robot Assistance
SummaryThe paper presents CarryOn, a dual-loop harness that augments a frozen vision-language-action policy for robot assistance. A planning loop keeps persistent user and scene context to enable personalized subtask selection, while an execution loop supervises actions to block predicted hazardous contact and invokes reusable recovery skills after failures. On a real robot and in simulation, the safety violation rate dropped from 52.6% to 6.1%, and recovery success rose from 13% to 80% when refined skills were reused.
In briefFull-duplex dialogue model ELLSA 2 that generates speech, gestures and expressions together
SummaryThis paper presents ELLSA 2, an end-to-end full-duplex dialogue model that jointly generates speech and streaming gestures and facial expressions. It can produce non-verbal motion even when silent, and uses contextual semantic calls with template guidance to produce meaningful gestures. Experiments report that it outperforms cascaded baselines on dialog-aware gesture and expression metrics and is competitive with open-source full-duplex speech models.
In briefRegistration method that preserves PET quantitative measures for longitudinal whole-body PSMA scans
SummaryThis paper adapts the diffeomorphic coarse-to-fine LapIRN architecture to register longitudinal whole-body PSMA PET/CT scans for the Learn2Reg 2026 PSMAReg challenge. The approach combines affine pre-registration, multi-modal deformable registration, and test-time instance optimization. It also adds losses that preserve metabolic tumor volume (MTV) and total lesion glycolysis (TLG), along with rigidity regularization for skeletal structures. On the authors' evaluation protocol, the submitted configuration achieved 81.0% DSC, 6.17 mm HD95, and 3.3% errors for both MTV and TLG.
Fryderyk K\"ogl, Anna Reithmeir, Daniel M. Lang et al.
In briefX-Fly: Learning to Reach and Stop in Chunked Drone Vision-Language Navigation
SummaryThis paper studies drone navigation guided by language and vision, where executing full predicted action chunks reduces model queries but leaves the policy unsupervised in off-expert states and at stopping. The authors propose X-Fly, a staged framework combining Recovery DAgger and Flow-GRPO to jointly refine a chunked flow controller and a STOP policy. On OpenFly, X-Fly reports 74.18% success on test-seen and 51.73% on test-unseen.
In briefTraining LLMs to Pick Features and Read Risk for ER Outcome Prediction
SummaryThis paper proposes TIR-EC, a reinforcement learning framework that lets an LLM call EHR feature groups as tools, with rewards from a gradient-boosted tree teacher. A reverse curriculum gradually removes task skills from the prompt so they are absorbed into the model weights. On three clinical prediction tasks, the 4B-parameter agent achieves the highest macro F1 among compared models, beats conventional ML baselines, and is competitive with much larger frontier LLMs.
In briefV-SHARE: Convergence proofs for advantage-based reinforcement learning
SummaryThis paper introduces V-SHARE, an advantage-based value-learning algorithm that shares value across actions and adds an importance-weighted Bellman-residual correction. The authors give the first complete theoretical analysis of this class of algorithms, covering asymptotic convergence and non-asymptotic finite-sample guarantees. Experiments on discrete MDPs and deep reinforcement-learning benchmarks support the results.
In briefSITAR: A Situational Distortion Metric for Evaluating Speech Recognition
SummaryThis paper proposes SITAR, a metric for automatic speech recognition that measures how transcription errors distort the situation conveyed by speech, beyond lexical or semantic differences. It combines situational severity and cumulative extent into one score through a bounded recursive formulation. The score is learned by a lightweight BERT-family regressor that rates reference-hypothesis pairs in a single forward pass. The authors also describe a validation protocol covering extrinsic validity and intrinsic scorer reliability.
In briefAssigning Each Reward to Its Relevant Tokens for Multi-Reward RL
SummaryThis paper addresses training language models with several verifiable rewards at once. It proposes MRTPO, which assigns each reward's advantage only to the tokens its verifier checks, without auxiliary models or token-level annotations. The authors report faster convergence and better performance than compared methods on text-to-SQL, multi-constraint instruction following, and tool calling.
In briefRLDF: A Dataset and Benchmark for Detecting Deepfakes in Local Facial Regions
SummaryExisting deepfake detectors are trained mainly on fully swapped or fully synthesised faces. This paper introduces RLDF, the first dataset for region-level deepfakes, containing over 2 million partial face-swap images across seven facial regions. The authors benchmark 12 state-of-the-art detectors and find that performance drops by 10.5% AUC and 10% accuracy under region-level evaluation. They also propose a mask-guided cross-attention model as a baseline.
In briefHyCo: Low-cost mobile-cloud speculative decoding using hybrid drafting
SummaryHyCo runs the server model's built-in multi-token prediction head on a phone's neural processor to build a deep draft chain, and adds alternative branches with one small language model pass. Tested on a Galaxy S26 Ultra over LTE/5G, it cut server GPU time per output token by 38.6% on Spec-Bench and 29.2% on 32K-token summarization versus server-side MTP self-speculation. The paper also reports a relaxed acceptance rule and mixed quality results on translation.
In briefPolicy-aligned verifier that checks how far to execute VLA action chunks before running them
SummaryVision-language-action policies generate action sequences in chunks, but choosing how much of each chunk to execute affects task success. The paper introduces a policy-aligned predictive verifier (PAV) that evaluates candidate action prefixes in the VLA's internal representation space before execution. It learns from terminal task outcomes using trajectory-level survival supervision and a localization prior. Experiments with pi_0 and pi_0.5 on LIBERO, RoboTwin 2.0, and real-world manipulation report improved success-query trade-offs.
In briefManiFusion: Faster Diffusion Image Generation by Processing Multiple Samples Together
SummaryDiffusion models produce high-quality images but are slow because they denoise iteratively over many timesteps. Existing methods mostly cut per-sample latency, whose gains are becoming saturated. ManiFusion instead fuses multiple samples before the denoising backbone and separates them afterward, with Trajectory-Fusion Scheduling adjusting fusion factors across sampling. The authors report that it works across U-Net and transformer backbones in pixel and latent spaces, composes with existing acceleration methods, and reaches higher throughput with minimal quality loss.
In briefLocalized Context Distillation (LoCD): Learning From Interaction Experience
SummaryThis paper proposes Localized Context Distillation (LoCD), a method that helps large language models keep improving from experience gathered during deployment. LoCD finds the critical spans of a student's trajectory that need revision, and a teacher regenerates only those spans from the student's own prefix for the student to learn from. According to the paper, LoCD outperforms strong baselines on math reasoning, code generation, and agent tasks, with more stable training and smaller, sparser parameter updates.
In briefGPU-Accelerated Robust MA-RIS Pre-Optimization for Joint Communication and Target Illumination
SummaryThis paper studies movable antennas (MAs) and reconfigurable intelligent surfaces (RISs) for integrated sensing and communication, where antenna positions and RIS phases are chosen offline on a slow timescale. It yields a reusable Pareto set balancing expected worst-user rate against worst-target directional transmit illumination. The proposed RAMP method, accelerated on GPUs, preserves regime-specific populations and builds a worst-regime robust archive. Numerical results show competitive robust Pareto quality against the evolutionary baselines considered, without online physical re-optimization.
In briefDo Agents Need a New Kind of Paper? SPRA: A Structured Paper Format for AI Agents
SummaryThe paper introduces SPRA, a structured representation that organizes research problems, methods, experiments, measures, and findings in papers as linked objects that agents can query. The authors built SPRA over more than one million AI papers and evaluated it with retrieval-augmented question answering across seven subfields. Reported results show better scores than passage retrieval with far fewer input and output tokens.
In briefLocating Objects, Relations, and Moments in Video with Query-Guided Grounding
SummaryThis thesis studies query-guided grounding in video at three levels: where objects are (object-level), how they interact (relationship-level), and when moments occur (moment-level). It proposes QDETRv and TOViT for object localization, a few-shot method for referring relationships, and the Moment Alignment Transformer (MATR) for video and text-based moment retrieval. The work aims to advance video understanding across spatial, relational, and temporal grounding.
In briefCGDT-CESN: A Clustered Echo State Network for Multivariate Time-Series Forecasting
SummaryThis paper proposes CGDT-CESN, a modular echo state network in which each cluster specializes in one variable of a multivariate time series. Connections between clusters are set using correlations observed among the variables, and their weights are trained to capture time-varying cross-variable dependencies. On Twitter information cascade and solar wind datasets, it outperforms a conventional echo state network, with gains maintained across multiple prediction horizons and with limited data.
Hariharan Srinivasan, Masoom Bhargava, and Sayan Gupta
