In briefLintEval: A Benchmark for LLMs on Linter-Guided Code Refactoring
SummaryThis paper introduces LintEval, a multilingual benchmark for testing whether language models can reduce linter warnings in C++, Python, and Java source files while keeping the programs working correctly. It contains 13,061 programs and 171,662 lint alerts and scores functionality preservation, alert reduction, avoidance of new alerts, and edit minimality. Results show closed-source commercial models reach up to 90.4% Reduction@P but only 67.1%-68.3% Pass@1, and refactoring-oriented fine-tuning raises a 7B open model's Reduction@P from 24.7% to 76.0%.
In briefRemembering Who and Knowing How: A Benchmark for Persistent Assistant Memory
SummaryThis paper argues that persistent personal assistants need both personal memory about the user and procedural memory about how to act, and that existing benchmarks usually test only one. The authors introduce PraxisMemBench, covering daily-life and work tasks where both memory types evolve together, and propose the PraxisMem method. In experiments, PraxisMem reaches a 40.78% overall Session Success Rate, compared with 20.60% for Claude Code's native memory.
In briefMorphAct: Robot Policies That Adapt Their Parameters to the Setting
SummaryThis arXiv paper proposes MorphAct, a framework that dynamically adjusts robot policy parameters as interaction contexts change, instead of using fixed shared parameters. Two coupled generators produce context-dependent low-rank updates for the vision-language and action modules, trained jointly through the native action objective. The authors report gains of up to 16.30% across three VLA backbones on LIBERO and Meta-World, and up to 14% over π0.5 in real-world evaluation.
In briefRethinking User Encoders: Diffusion-State Routing for Scenario-Adaptive User Representations
SummaryThe paper argues that existing LLM-based user encoders are unidirectional, which limits fusion of multi-source behavioral signals and makes their static representations hard to adapt to different scenarios. The authors propose FINDERS, a router-equipped diffusion language model framework that denoises, factorizes, routes, and compresses states into scenario-adaptive user representations, and build the BehaviorQA corpus with explicit behavioral rationales. According to the paper, FINDERS outperforms state-of-the-art methods in experiments and online A/B tests across ten industrial scenarios.
WhyNovel routing of diffusion LLM denoising states for scenario-adaptive user representations, backed by offline and online A/B tests across ten industrial scenarios.
In briefConvergence of Semi-Gradient TD Learning from a Mean-Field Perspective
SummaryThis paper studies the stability of semi-gradient temporal-difference learning, a core reinforcement learning method, for two-layer neural networks in the mean-field regime. It identifies a simple stability mechanism in which entropy-induced relaxation dominates the sensitivity of the moving bootstrap target. Under this condition, the authors prove exponential convergence to a unique fixed point and quantify its bias relative to the unregularized Bellman solution. They also show the same mechanism extends to entropy-regularized off-policy control via soft Q-learning.
In briefChemArgus: An automated diagnostic benchmark for multimodal chemistry reasoning
SummaryChemArgus is a competition-level chemistry benchmark built from competition problems, textbooks, and research papers. It grades each subquestion with partial credit and pairs real images with controlled text descriptions to compare visual and textual understanding. The paper reports that GPT-5.6-Sol scores 23.2% on unassisted whole-problem solving, versus 28.8% and 30.4% under two sequential metrics, and treats the gap as a measure of integration burden.
In briefHyperbolic Goal-Conditioned RL Beats 1000-Layer Euclidean Networks
SummaryThe paper proposes hyperbolic CRL, which replaces Euclidean distances in the contrastive reinforcement learning objective with geodesic distances in hyperbolic space. In mazes with walls, route distance can deviate substantially from Euclidean distance, and hyperbolic representations better match this structure. The authors report that depth-64 hyperbolic agents outperform 1024-layer Euclidean networks on hard navigation tasks while training over 12 times faster with about 13 times fewer parameters.
In briefMinimax estimation of Markov chain spectral gaps for empirical concentration bounds
SummaryThe paper studies minimax estimation of the pseudo-spectral gap of a Markov chain from a single trajectory, extending prior finite-state results to compact continuous state spaces. The authors state that their estimator is minimax-optimal up to logarithmic terms for chains whose transition operator has a Lipschitz density. Combined with Paulin's Bernstein inequality, it yields empirical concentration inequalities for Markov chains, which the authors apply to confidence intervals, active learning and generalization bounds. For rested bandits with Markov rewards, they give a UCB variant with a regret bound similar to the i.i.d. case, without requiring knowledge of the pseudo-spectral gap.
WhyExtends pseudo-spectral gap estimation to continuous Markov chains with a minimax-optimal estimator, yielding empirical concentration bounds and a UCB regret result.
Vahe Karagulyan, Miko{\l}aj Kasprzak, Pierre Alquier
In briefProvable Transformer Context Compression for In-Context RL: a theoretical study of how much context a Transformer needs to do TD evaluation.
SummaryThis paper studies how much context a Transformer needs to perform in-context reinforcement learning without parameter updates. It shows that, with a shared linear decoder, the minimum context dimension needed to exactly reproduce the first L TD iterates equals the Krylov dimension of the sampled transitions. It also characterizes the optimal approximation error when the context dimension is smaller, and reports experiments on memory savings and pretraining convergence.
In briefWhat coding agents leave out: measuring task-state gaps between execution and reports
SummaryThis paper introduces CTXGAP, a measurement study comparing what coding agents' accumulated user-facing reports reveal about task state against their full recorded execution contexts. Across 37 coding tasks and 4,799 valid probes, report-only readers recovered 52.0% of task-state information versus 92.7% for full-context readers. The authors propose report-based task-state recoverability as a complementary evaluation dimension.
In briefA Strong Harness Narrows the Model Gap: Converting Legislation to XML
SummaryThis paper evaluates a pipeline that converts U.S. state legislation into strict, schema-constrained XML. The authors run a controlled ablation crossing three harness designs with three Claude models on 153 gold bills from two jurisdictions. The full multi-agent pipeline with the weakest model outperforms the strongest model alone, and the remaining errors concentrate in citation normalization, which is handled by a deterministic lookup tool.
In briefLimits of Reservoir Computing (Echo State Networks) on Few-Shot Sequence Pattern Tasks
SummaryThis paper empirically analyzes variants of Echo State Networks (ESNs) on few-shot sequence-to-sequence tasks built from synthetic discrete symbol sequences. It finds that keeping a linear state regime is important for memory tasks such as pattern completion and simple copy. The proposed data-dependent gated variant, G-ESN, solves the adding problem and bracket matching, which conventional ESNs cannot, and a shared readout across layers outperforms concatenated readouts in deep ESNs on several structural tasks.
In briefGURI: Gray-box detection of LLM hallucinations using token identities
SummaryThe paper notes that existing gray-box hallucination detectors use only token-probability statistics and cannot tell which alternative words a model is uncertain between. The authors propose GURI, a lightweight detector that combines candidate-token probabilities with per-token vector identifiers. Across three open-weight models and three question-answering benchmarks, mean test ROC-AUC rises from 72.09% to 79.09%, and on the closed-source GPT-4.1 from 67.39% to 71.84%.
Shaked Goldshtein, Yitzchak Grunbaum, Fabrizio Frasca et al.
In briefWhy 3D reconstruction fails on under-ice underwater robot video, and how to predict it
SummaryThis paper examines why 3D reconstruction from under-ice AUV video footage from Antarctica has been inaccurate. Four reconstruction methods were tested on the same footage, and all four failed in different ways. The authors attribute these failures to a shared capture condition: a small ratio of camera baseline to scene depth. They propose a detection system that flags this type of failure within minutes.
Ana-\c{S}tefania Kog\u{a}lniceanu, Kl\'ara Orb\'an, Ashwin Nedungadi et al.
In briefA Systematic Review of Grading Open-Text Legal Answers, from Law School Exams to Trustworthy Rubric Judges
SummaryThis paper systematizes how open-text legal answers, such as case solutions, essays and memos, can be evaluated. The authors derive eight desiderata from educational measurement and examination law, and compare eight evaluation paradigms, from human grading to LLM judges using instance-specific rubrics. They find that a task-specific rubric is the strongest factor for agreement between LLM judges and human graders, but validation gaps remain, especially around defensible alternative solutions. The authors conclude that such judges currently suit formative feedback, benchmarks and assisting examiners, but not replacing them where credit is at stake.
In briefLatentLoop: Scaling Looped MoE Models by Allocating Compute Asymmetrically
SummaryThe paper proposes LatentLoop, a method for looped Mixture-of-Experts models that reuse a shared block twice. The authors found that shrinking the first loop visit's residual scale preserves or improves performance, while doing so on the second visit hurts it. LatentLoop therefore compresses the first visit's sequence, keeps the second at full resolution, and uses the saved compute to enlarge the hidden size. Across 15 configurations from 1.3B to 7B parameters, it consistently lowers validation NLL and improves average scores on nine benchmarks in 0-shot and 5-shot settings.
In briefSafe-FCL: Continual Risk Awareness and Fine-Grained Safety Aggregation for Federated Continual Learning of LLMs
SummaryThis paper proposes Safe-FCL, a framework for federated continual learning of large language models that guards against malicious client updates. Its Continual Risk Awareness mechanism learns emerging risk patterns across sequential tasks and constrains detector updates with a protective subspace from historical features to reduce forgetting. Its Fine-Grained Safety Aggregation mechanism combines client-level risk assessments with module-level deviations to regulate aggregation contributions. The authors report gains of 0.41% and 0.64% in global-model Rule and MD-Judge scores, and a 2.01% increase in average risk recognition accuracy.
In briefChoosing Attention Head Count from Intrinsic Dimension: A Small-Scale Study
SummaryThis paper tests a simple way to set a transformer's number of attention heads in advance: estimate the intrinsic dimension of token representations without labels and use roughly that many heads. The authors estimate it at about 9.2 using a nested-dropout VAE on DistilBERT embeddings of WikiText-2. In a small 4-layer model, 10 and 32 heads were statistically equivalent on SST-2 and MRPC, with 69% fewer attention parameters for the 10-head model. The authors stress the evidence is narrow and present the approach as a heuristic worth testing at larger scale, not a general principle.
Arjun Boya, Charan Sai Bandaru, Masthanaiah Pillimetla et al.
In briefVISTA: Smarter Scheduling to Speed Up Reward-Guided Discrete Diffusion
SummaryThis paper proposes VISTA, a training-free method that chooses which denoising timesteps receive Sequential Monte Carlo guidance under a fixed computation budget, reducing the cost of guiding discrete diffusion models. It collects value statistics from a few full-guidance warmup generations and solves the resulting surrogate exactly with dynamic programming. Across text, biological sequence, and image tasks, VISTA outperforms hand-designed sparse schedules in most settings. In toxic-text generation it reaches a 0.99 unique-toxic ratio versus 1.00 for full guidance while cutting per-generation latency from 204.23 s to 38.16 s.
In briefTraining-free pruning of future-frame tokens to speed up world action models
SummaryWorld action models predict future visual states and robot actions together, but processing dense future-frame tokens causes high inference latency. The paper proposes CF-Prune, a training-free method that prunes these tokens in pretrained models using instruction and action attention. Tests on RoboLab, RoboCasa, and real manipulation tasks show average speedups of 1.19 to 1.27 times at 35% pruning with higher success rates, and 1.29 to 1.33 times at 50% pruning with success within 1.0 percentage point of baseline..
In briefLearning interpretable world models by combining neural networks with logic rules
SummaryThe paper introduces Concept-Grounded World Models (CGWMs), which learn interpretable concept representations from observations using explicit knowledge written in temporal logic. Planning is formulated as differentiable reasoning, with actions optimised by gradient ascent to reach a goal. Experiments on two game environments compare CGWM with a purely neural baseline and a neurosymbolic baseline. The authors report that CGWMs ground concepts more accurately over long horizons and plan more robustly both in and out of distribution.
Riccardo Andreoni, Massimiliano Ronzani, Chiara Ghidini et al.
In briefAtomSearch: Tree Search for Improving AI Agent Skills with Evidence
SummaryAtomSearch is a method for improving editable instruction sets (agent skills) that guide frozen AI models. It makes small, traceable rule edits linked to observed successes and failures, and uses a tree search to keep useful intermediate skill states. Evaluated on four tasks and five target models, it outperforms SkillOpt on three of the four tasks and produces much more compact skills.
In briefWhen Protection Overrides Consequences: Protection Mode in LLM Decisions
SummaryThis paper studies a tendency of large language models to prioritize protecting a person or animal over larger physical consequences in simple, urgent decisions, which the authors call Protection Mode. They introduce ProtectionBench, with 285 prompts across 34 physical families, and test nine models against human raters. Some models choose the subject far more often than humans do, and in some cases they state the correct risk ordering yet still pick the subject.
In briefLearning Graph Energies for Finite-Step Generation
SummaryThis paper proposes finite-horizon energy learning, which adapts the acceptance energy of a Metropolis-Hastings sampler to the distribution it returns after a fixed number of steps. It is applied to graph generation, where a learned energy governs acceptance while initialization and proposals come from a frozen pretrained model. Experiments compare against equilibrium-correct energies on binary targets and on MOSES, reporting lower terminal total variation and lower FCD.
Amartya Banerjee, Xingyu Xu, Caroline Moosm\"uller et al.
SummaryThis small exploratory study examines whether machine translation preserves media bias, covering 27 types of linguistically expressed bias, when translating from English and German into Bulgarian. Four state-of-the-art translation methods are evaluated. The findings indicate that machine translation does not fully retain the news bias of its source text and may also add bias of its own. The authors state this is the first experimental study of media bias retention in machine translation.
In briefMorphingNet: A Symbolic Regression Method Where the Network Is the Formula
SummaryThis paper presents MorphingNet, a neural network for symbolic regression whose units continuously change their functional form, with each form trained as a parameter. This turns formula search into a single continuous optimization problem, and a later structure-recovery step rewrites the trained network as short elementary expressions. The authors report state-of-the-art accuracy on SRBench and speedups of up to two orders of magnitude on large data compared with the strongest search method.
In briefReasoning Models That Can't Stop: Fixing Pruning-Induced Looping
SummaryThis paper examines why weight pruning lowers the accuracy of large reasoning models (LRMs) and finds that the main problem is failing to stop rather than losing reasoning ability. As sparsity rises, pruned models more often loop repetitively and fail to finish within the token budget, even after the correct answer has appeared. The authors find a consistent hidden-state direction separating terminating from looping continuations across pruned checkpoints, and propose TGAS, a training-free method that adds this direction to the residual stream at inference time. Across six LRMs, three pruning pipelines, and math and code benchmarks, TGAS improves termination and accuracy while shortening generations.
Minh-Anh Nguyen, Nguyen T. Ha Lan, Xiyan Fu et al.
In briefCreating a synthetic image-instruction dataset for training multimodal AI models.
SummaryThis paper presents a compact, reproducible pipeline for generating image-based instruction-response pairs for the DCVLR competition. Starting from the Cauldron collection, it uses Qwen2.5-VL-7B Instruct with Magpie-style synthesis to create image-conditioned instructions, then sanitizes, tags, and filters them with a DeBERTa-v3 reward model, keeping only positively scored samples. The resulting dataset spans math, spatial, and social reasoning domains.
Eleni E. Kasoura, Dimitrios Damianos, Georgios Paraskevopoulos et al.
In briefSAM3RL: Reinforcement Learning for Memory Control in Visual Object Tracking
SummarySAM 3 is a foundation model for image and video segmentation whose tracker stores past-frame information in a memory bank managed by a first-in-first-out rule. This paper introduces SAM3RL, which adds a lightweight memory controller trained by reinforcement learning to decide which frames enter the memory bank. SAM3RL sets a new state of the art on 10 of 12 tracking and video object segmentation benchmarks with no runtime overhead. Applied to SAM 2, the same approach (SAM2RL) outperforms hand-crafted rules on 9 of the 12 benchmarks.
Tom\'a\v{s} \v{C}\'i\v{z}ek, Matej Straka, Illia Volkov et al.
In briefMemDuet: Dual-Process Memory for Efficient Long-Document Reasoning
SummaryThis paper proposes MemDuet, a dual-process memory framework that pairs a small language model for fast evidence selection from text chunks with a large language model that verifies and updates memory in document order. The system is trained in three stages, combining supervised fine-tuning, GRPO-based optimization, and joint policy optimization. Evaluated on RULER-HQA, LongBench v2, and 2WikiMultiHopQA, it reports an average end-to-end speedup of 9.74x and a 73.91% output-token reduction relative to RL-MemAgent.
In briefConvergence rates of a Gaussian-process Thompson sampling method for global optimization
SummaryThe paper studies how quickly Gaussian process Thompson sampling (GP-TS) finds the global optimum of an expensive noiseless black-box function. For Matérn kernels it proves an upper bound on simple regret for both discrete and continuous versions, and a matching lower bound showing a logarithmic penalty in the standard one-dimensional Matérn-5/2 case. The authors propose GP-CTS, which caps posterior sample paths using an RKHS norm bound, removing the logarithmic penalty and achieving the minimax-optimal rate.
WhyProposes GP-CTS removing the logarithmic penalty of standard GP-TS, achieving minimax-optimal rates with matching lower bounds; strong theory, though application impact is mostly theoretical.
In briefMolHop: Agentic reinforcement learning for discovering structurally different molecules
SummaryThe paper argues that molecular optimization and finding structurally distinct replacement molecules are different search problems. It proposes MolHop, a framework that organizes about 1.6 million commercial molecular scaffolds into a hierarchical action space, letting an agent explore via scaffold navigation and molecular editing while keeping trajectory memory. The authors trained MolHop-9B with agentic reinforcement learning, which they report outperforms listed generative baselines on drug and commercial compound tasks in cross-scaffold joint success rate, with candidates showing larger scaffold differences from reference molecules.
In briefPredicting more accurate electronic band structures than standard DFT using dual-space orbital learning
SummaryThe paper introduces O-kRystal, a framework that uses a cheaper PBE calculation as an energy baseline and orbital information source to predict higher-level corrections to band energies. On nearly 6,000 three-dimensional materials, it predicts HSE06 band energies within 5 eV of the band-edge midpoint with a mean absolute error of 0.171 ± 0.009 eV. On six strongly correlated materials, it reduces the PBE band-energy MAE by 68.7%.
WhyLearns HSE06 and G0W0 corrections from PBE orbitals across theory levels, with large error reductions on correlated materials.
Minhyuk Kang, Beom Seok Kang, Nikolai Argatoff et al.
In briefAssessing AI Capabilities After Benchmarks Max Out: Using Solution Transcripts
SummaryAs AI models improve, many benchmarks reach near-perfect scores and can no longer distinguish between models. This paper tests whether the full transcripts of how models solved saturated benchmark tasks can still reveal their capabilities. Using historical releases to simulate saturation, the authors compare several methods, and pairwise judging performs best, reducing prediction error 5.3 times versus using benchmark scores alone. However, these methods do not generalise reliably across distribution shifts, for example underestimating early-2025 reasoning models.
Morio Hamada, Jehyeok Yeon, Matteo Migliarini et al.
In briefNot all criteria are equally important: a weighting method for rubric-based LLM evaluation
SummaryThe paper argues that rubric-based evaluation of large language models usually weights all criteria equally, which gives minor criteria the same influence as essential ones and distorts scores. The authors propose Importance-Aware Criteria Weighting (ICW), which has an LLM make pairwise importance comparisons between criteria, checks and revises inconsistent comparisons, and derives weights from the principal eigenvector of the comparison matrix. Experiments reportedly show ICW outperforms uniform and other weighting methods and yields better policy models in downstream reinforcement learning.
In briefMUCO: A framework that breaks long robot tasks into reusable, verifiable manipulation units
SummaryThe paper proposes manipulation units (MUs), reusable robot operations each defined by a local goal, a start condition, and observable success and failure checks, learned and verified individually. It introduces MUCO, which builds a sequence of MUs from language and vision, assigns each to the best-calibrated policy in a registry, and uses visual verification to retry or fall back without rerunning verified steps. Evaluated on RLBench, RoboTwin 2.0, and five real-world tasks, MUCO reportedly achieves higher complete-task success, better failure containment, and lower error propagation than baselines.
In briefMeasuring object detection annotation agreement with optimal transport
SummaryThe paper proposes an Optimal Transport (OT)-based metric for measuring agreement between two sets of object detection annotations. Instead of binarizing bounding box overlap, it assigns graded costs to matched and unmatched boxes. Through theoretical analysis and experiments on real and synthetically perturbed annotations, the authors show that OT captures graded changes in agreement and distinguishes annotation pairs that thresholded metrics score identically.
In briefCARD: a closed teacher-student loop for adversarial robustness distillation
SummaryStandard adversarial robustness distillation transfers robustness one way, from a fixed robust teacher to a compact student. This paper proposes CARD, a cyclic framework that uses the distilled student's feedback to refine the teacher and then transfers the teacher's gains back to the same student. Experiments on CIFAR-10, CIFAR-100, and TinyImageNet-10 across several architectures and base methods report improved adversarial robustness in most settings, with clean accuracy generally preserved.
In briefHow ordinary noise during learning harms AI agents' self-improving skills, and a new benchmark to test it
SummaryThis paper introduces RSE-Bench, a benchmark for testing how AI agents evolve reusable skills when experience contains ordinary, non-adversarial noise, across four domains and six benchmarks. It separates noise during learning from input noise at evaluation and provides the RSEClaw framework for injecting noise and diagnosing skill updates. Across three methods and two backbones, noisy learning keeps only 25.5% to 42.1% of the clean self-evolution gain.
In briefSparse Transformer: A Controlled Testbed for Studying How Neural Networks Work Internally
SummaryThis paper introduces a Sparse Transformer, a 5-layer language model trained on TinyStories whose sparse MLP layer encourages individual neurons to specialize, providing a known reference for evaluating interpretability methods. The authors report that its neurons are more interpretable than those of a loss-matched dense transformer, with circuits of a median of 12 neurons reproducing individual predictions. The testbed is also used to compare feature learning methods and to study how compression into a dense model leads to feature superposition.
In briefBeyond Benchmark Accuracy: A Diagnostic Test for LLM Chemical Prediction
SummaryThis paper argues that large language models perform well on chemical prediction benchmarks, but existing evaluations reveal target identity directly, making it unclear whether predictions reflect transferable structure-property reasoning. The authors propose an Identity-Blinded Reasoning Diagnostic framework that defines targets by a known source entity and an explicit structural edit instead. Tested across six tasks including odor, molecular property, retrosynthesis, and reaction yield prediction, the results show high benchmark accuracy does not necessarily indicate reliable structure-property reasoning.
In briefTask-specific LoRA adapters for multimodal medical image parsing
SummaryThis paper describes a system for FLARE 2026 Task 3 built on the Qwen3.5-9B vision-language model, covering six task families and seven imaging modalities. It uses a shared frozen 4-bit backbone with LoRA adapters selected from the input question, with adapter ranks and prompts customized per task. The authors report results on an internal holdout and an organizer's hidden-validation set across classification, multi-label classification, counting, detection, regression, and report generation, and state that the benefits vary across tasks.
Taha Sabir, Alberto Montero Molina, Thinh Quang Ly et al.
In briefHow GUI Agents Fail: GraceBench, a Benchmark for Diagnosing Execution Stability
SummaryThis paper introduces GraceBench, an online evaluation framework for GUI agents on mobile apps. It represents reachable app interaction spaces as UI Transition Graphs and uses video-grounded trajectory evaluation to measure deviation rate, loop burden rate, and recovery ability. Across 3,718 trajectories, the authors find that outcome scores such as success rate can hide execution instability.
In briefA Single Vector Embedding Method for Distributions, Functions, and Graphs
SummaryThis paper proposes a unified framework that represents higher-order objects such as distributions, functions, and graphs as fixed-size embedding vectors. An encoder produces the embedding from a finite set of samples, while a domain-specific decoder models the object conditioned on that embedding to supply the training signal. The authors test the approach on all three object types and show downstream uses including interpolation in embedding space, solving differential equations, and predicting unseen graph edges.
Radovan Halu\v{s}ka, Christopher Solinas, Nathan R. Sturtevant et al.
In briefDeepPV-MRF: Direct Deep-Learning Estimation of Tissue Fractions from MR Fingerprinting Signals
SummaryThis paper presents DeepPV-MRF, a deep learning framework that estimates the fractions of cerebrospinal fluid, gray matter, and white matter in each voxel directly from raw magnetic resonance fingerprinting signals. It skips intermediate T1/T2 maps by combining temporal encoding with spatial U-Net refinement. Compared with two conventional methods on simulated and in vivo data, it reports lower errors and higher structural similarity, while the authors note that further validation is needed before clinical use.
In briefFew-shot adaptation to reduce false alarms in acoustic machine fault detection
SummaryThis paper proposes a two-step method for detecting anomalous machine sounds. A regularized unsupervised model is first trained on normal operating audio, then adapted with a few anomalous samples collected after the first fault. Adaptation raised AUPR from 23.0 to 64.0 and pAUROC at low false-positive rates from 39.6 to 72.8. The authors also report that covariance regularization improves prediction accuracy and training stability.
Simone Lugani, Christof Pichler, Hannes Wegleiter et al.
In briefLimits of representation-based test-time adaptation for multimodal LLMs
SummaryVision-language models lose significant accuracy under common image corruptions. The paper finds that representation-based test-time adaptation, which works for CLIP-like models, barely recovers performance in generative multimodal LLMs. Even with offline access to target data, sample-blind feature alignment yields only marginal gains. The authors attribute this to corruption shifts in visual token sequences being nearly orthogonal and spanning the full embedding space, suggesting representation-space interventions alone may not ensure robustness.
Miko{\l}aj Pi\'orczy\'nski, Wojciech Masarczyk, Damian S\'ojka et al.
In briefReCoPII: A framework for building context-aware personal data masking datasets
SummaryThis paper presents ReCoPII, a framework for generating training data for contextual masking of personally identifiable information (PII). It plans ontology-constrained entity relations, writes sentences with placeholders, reconstructs annotations with code, checks them with a separate LLM, and derives MASK/KEEP labels from entity relations. Applied to Korean documents from six domains, it produces the ReCoPII-Ko corpus, whose sampled KEEP labels show higher annotator agreement than SPY-style augmentation.
In briefLength-Conditioned Majority Voting: weighting answer votes by reasoning length
SummaryThis paper proposes Length-Conditioned Majority Voting (LCMV), a test-time method that improves majority voting over multiple sampled chain-of-thought answers for large language models. It builds on the observation that longer chains of thought tend to be less often correct, modeling this as a length-conditioned prior with a logistic function. The prior is adapted using only unlabeled test data, without fine-tuning, and experiments across several LLMs and reasoning benchmarks report consistent gains over standard majority voting and other aggregation strategies.
WhyLength-conditioned prior with test-time adaptation on unlabeled data meaningfully improves majority voting, validated with extensive comparisons across multiple LLMs and benchmarks.
Fiorenzo Parascandolo, Enver Sangineto, Xiaoyi Yu et al.
In briefOccupancy Modeling via Drifting: Direct Policy Extraction from One-Step Generation
SummaryThis paper studies reinforcement learning from unlabeled data by pre-training a generative model of occupancies (discounted state visitations) on reward-free trajectories, then extracting a policy from a small reward-labeled dataset. The authors analyze three design choices: how well multi-modal occupancy structure is captured, the class of generative model, and the policy extraction method. They propose a streamlined approach using drifting, a one-step generative sampler, that extracts policies by differentiating directly through the occupancy model, avoiding critic distillation. The authors report performance improvements on standard benchmarks.
Robin Uhrich, Marco Bagatella, N\'uria Armengol Urp\'i et al.
In briefBluePixel: Native 4K Agentic Image Editing
SummaryBluePixel pairs a native-4K image editing model with an agent harness and discloses its full training recipe. The model adapts a 1K-pretrained editor to 4K through progressive VAE adaptation, continual learning, and reinforcement learning, then uses SATP distillation to cut 50 denoising steps to 4, completing a 4K edit in 8.7 s on one Kunlunxin P800 XPU. The agent splits user instructions into operations, verifies results with the training reward model, and retries within bounded attempts. On VINS-4KEval++ (3,305 native-4K cases), it ranks second of seven models, behind only GPT Image 2 on non-edited-region consistency.
In brief"The World Won't Wait": Evaluating and Training Agents Under a Live Clock
SummaryThe paper introduces RealtimeBench, 180 tasks across six game types, to test LLM agents while games keep running during inference. It finds that GPT-6.1-sol beats the faster Jev when games are paused but falls behind when they run live. The authors propose TORSI, where a slow teacher analyzes failures and writes reusable instructions between attempts, and a fast student acts without teacher calls during play. After two rounds, Jev's average test score rises from 13.51 to 24.19 on a 0–100 scale.
In briefCoDy: Video Codec Signals as Complementary Priors for Unannotated Action Localization
SummaryThis paper addresses unsupervised temporal action localization, which aims to find action categories and their start and end times without annotations. The authors observe that video codec signals, such as group-of-pictures quantization and motion-vector variations, carry category and transition information. They propose CoDy, which combines a codec-enhanced view with RGB and optical flow to produce soft pseudo-labels and a boundary prior. Codec branches are used only during training, and experiments on THUMOS14 and ActivityNet v1.3 report results that exceed existing methods.
In briefNeural Wasserstein Splines: a representation-invariant smoothness prior for inferring population trajectories from snapshots
SummaryThe paper addresses inferring population dynamics from unpaired time snapshots, where evolution between measurements is underdetermined. It proposes Neural Wasserstein Splines (NWSP), which uses a neural network to map latent variables to cubic spline control points and penalizes an approximate Wasserstein acceleration obtained by projecting velocity onto gradients of a scalar-potential dictionary. The authors prove the population penalty depends only on the marginal curve for a fixed dictionary, and report competitive results against transport and flow-based methods on synthetic and single-cell benchmarks.
In briefSPARS: A Stability-Plasticity-Aware Region Selection Framework for Continual Learning
SummaryThis paper proposes SPARS, a replay-free framework for continual learning that aims to learn new tasks without forgetting old ones. It combines global parameter isolation with localized regularization, splitting parameters into winning, neighboring, and non-important regions based on estimated importance. Experiments on CIFAR-10 and TinyImageNet show higher average incremental accuracy than existing methods, with about 6% gain on TinyImageNet class-incremental learning while keeping over 60% of parameters available for future tasks.
In briefRCCM: Diagnosing Gains and Harms When Reasoning Prompts Transfer
SummaryThis paper introduces RCCM (Reasoning-Control Capability Mismatch), which pairs answers to the same questions with records of how generation ends to examine when reasoning prompts help or hurt across deployments. Across five open-weight deployments and nine strategies, aggregate accuracy hides substantial question-level gains and harms. The authors conclude that a prompt's benefit must be interpreted alongside its execution and evaluation, since matching a reference answer does not necessarily mean delivering it.
In briefFlow-matching distillation of rationale distributions for distributional continuous latent reasoning
SummaryThis paper proposes rationale distribution distillation (RDD), which trains a conditional flow over fixed-length latent sequences by distilling from multiple textual reasoning paths. It also expands distillation targets through within-query mixup between latent rationales for the same query. A frozen language model decodes answers from sampled rationale contexts. The authors report higher answer coverage than evaluated point-valued latent reasoning baselines, particularly on advanced math tasks.
In briefSEA-HELM Audio: Testing how speech input affects multilingual AI pipelines in Southeast Asian languages
SummaryThis paper introduces SEA-HELM Audio, an evaluation framework that extends the text-based SEA-HELM benchmark to cascaded speech recognition and large language model pipelines. Using synthesized speech in five languages, it compares accuracy on reasoning and instruction-following tasks between text and speech inputs. The results show substantial accuracy drops with speech input, with Tamil showing the largest degradation.
In briefExplorer-Zero: Self-play training for long tasks using decomposable verifiable rewards
SummaryThis paper proposes Explorer-Zero, a self-play framework in which LLM agents jointly generate tasks and fine-grained verifiable rewards through interaction with an environment, without human-annotated training tasks. A proposer explores the environment and records intermediate and final states as checkpoints that reward partial completion. Across ALFWorld, ScienceWorld, and AppWorld, it outperforms the evaluated self-play baselines, and its 30B agent reaches 59.67% task success on AppWorld, close to 62.65% for RL on human-annotated tasks.
In briefEuler: A human-centered system for multi-agent mathematical discovery
SummaryThis paper presents Euler, a mathematical research system that organizes statements, assumptions, proof dependencies, failed attempts, and verification evidence into connected research graphs. Researchers can trace how results were developed and let agents retain task-relevant knowledge across sessions. The authors illustrate its capabilities with three mathematical research case studies and describe the implementation and software validation.
In briefWhen Extremists Enter LLM Societies: Extreme Behavior in Multi-Agent Systems
SummaryThis paper studies how small groups of persistent extreme agents shape collective dynamics in LLM-based multi-agent social simulations. The authors introduce EXAM, a framework for configuring extremist agents and controlling their population share, spatial distribution, interaction protocols, and measurement. Experiments across four scenarios (congressional election, financial crisis, epidemic public opinion, and iterated prisoner's dilemma) show measurable collective shifts that depend on the backbone model and scenario.
In briefOptimisation landscape of overparameterised weighted MMD
SummaryThis paper studies the optimisation landscape of maximum mean discrepancy (MMD) over weighted atomic measures. For the one-dimensional Gaussian kernel, it shows that with enough particles, every local minimum can have arbitrarily small MMD for any compactly supported target. When the target is finitely supported, beyond an explicit particle threshold there are no spurious local minima, and every spurious stationary point is unstable.
WhyProves sufficiently many particles eliminate spurious local minima in weighted MMD optimisation, with stability analysis. Results are substantial but limited to one-dimensional Gaussian kernels.
Tharrshann Jayan Logarajah, Jeremias Knoblauch, Louis Sharrock
In briefUsing Text to Filter Training Data for Better Virtual Flow Meters
SummaryThis paper addresses poor-quality training data in industries such as energy, using large language models to combine unstructured operational text, like engineer comments, with sensor data to detect anomalies. The method performs contextual anomaly detection and then generates a binary mask to filter the training set. On a synthetic oil and gas dataset, using a virtual flow meter as the downstream task, the multimodal approach reduced MAPE from 7.1% to 5.4% compared with sensor-only filtering.
Kristoffer Hjelmen Nesland (1, 2), Bjarne Grimstad (2) et al.
In briefOmniPoser: Flexible Human Motion Recovery in the Wild via Masked Flow Matching
SummaryThis paper presents OmniPoser, a universal framework that recovers full-body SMPL human motion from any combination of heterogeneous inputs, such as 3D keypoints, sparse IMU rotations, or head-mounted 6-DoF poses, within a single model. It uses a cross-modal masking mechanism and a Masked Conditioning Encoder to handle diverse sensor configurations, and generates motion with Conditional Flow Matching using a single ODE step at inference. Across seven benchmarks, including in-the-wild Nymeria and Ego-Exo4D, it performs on par with state-of-the-art methods while running at over 700 FPS.
In briefEnergy use comparison of diffusion and autoregressive language models
SummaryThis paper benchmarks the inference energy consumption of diffusion language models (LLaDA-8B, Dream-7B) against autoregressive models (Llama-3.1-8B, Qwen-2.5-7B). It examines trade-offs among model architecture, KV-cache optimization, and joules per token. At 200 to 600 tokens, diffusion models use 2.5 to 2.7 J/token, about twice as efficient as autoregressive models with KV-cache at 5.3 to 5.7 J/token. KV-cache reduces autoregressive energy use by 50% to 70% on longer sequences.
In briefDeltaStreamer: A Human-Like Agent for Understanding Streaming Video
SummaryThis paper proposes DeltaStreamer, an agentic framework for streaming video understanding inspired by how humans observe, remember, and investigate. It continuously watches the stream and records visual changes in structured memory before any query arrives. When a question comes in, a query-driven planner retrieves memory and replays historical frames to gather targeted evidence. The authors report overall scores of 66.07% on OVO-Bench and 59.53% on StreamingBench, improving over representative streaming video methods.
In briefOpenTWBench: A Taiwan-focused test for AI models, and how testing methods change scores
SummaryThis paper presents OpenTWBench, an open evaluation suite for large language models' knowledge of Taiwan, with two benchmarks: tw-exam-bench, drawn from fifteen years of Taiwan's national exams, and tw-formosa-bench, covering everyday and contemporary Taiwan topics. The authors also show that the evaluation method itself strongly affects scores, such as changing how answers are extracted raising accuracy from 20.8% to 93.9%. Datasets, curation data and a leaderboard are released publicly.
Liang-Hsun Huang, Min-Yi Chen, Chao-Chun Chuang et al.
In briefVisibilitySplat: Ray-traced occlusion for relighting in 3D Gaussian Splatting
SummaryThis paper tackles inverse rendering with 3D Gaussian Splatting by adding explicit visibility modeling for relighting and for separating materials from lighting. A sparse, BVH-accelerated point-based ray tracer estimates occlusion, while the original 3DGS rasterization loop is kept. Evaluated on the TensoIR synthetic dataset against NeRF, tensor, and 3DGS baselines, it is competitive for novel-view synthesis, albedo reconstruction, and relighting. Ablations indicate that explicit visibility modeling is an important contributor to relighting quality.
In briefMemory-Efficient Optimization with Residual Correction, from Momentum SGD to Muon
SummaryThis paper proposes a residual correction framework that keeps optimizer states in a subspace and adds the orthogonal residual gradient to parameter updates, aiming to reduce memory costs in LLM training. It applies the scheme to momentum SGD (SMRC) with a convergence rate analysis, extends it to a multi-stage variant that periodically refreshes bases, and incorporates it into Muon as PMRC. The authors report that extensive experiments confirm the algorithms' effectiveness and efficiency.
In briefLAGO: Language Agents Reusing and Growing a Library of Options
SummaryThis paper introduces LAGO, a framework that treats specialized language agents as reusable, temporally extended options. A meta-agent retrieves or creates options, dispatches subtasks, verifies results, and adds successful new options to a persistent library that grows without updating model weights. On 500 SWE-bench Verified tasks, a frozen Qwen3.5-35B-A3B resolves 58.6%, and on a fixed 150-task subset LAGO reaches 60.3%, above mini-SWE-agent (49.3%) and Terminus 2 (57.3%). The authors note that component-level causal effects depend on the task and model.
In briefEGCForest: A structured memory method for enterprise group chat agents
SummaryEnterprise group chats hold project progress, decisions, and task ownership, but supporting evidence is spread across interleaved threads, channels, and time. The paper proposes EGCForest, which organizes an immutable message log into three complementary forests (Channel, Actor, and Artifact) linked by a Placement Map and sparse structural bridges. On three enterprise-chat benchmarks, it outperforms RAG and memory-system baselines, improving answer accuracy by up to 20.85 points over MemOS while supporting efficient online retrieval.
In briefTeaching web agents in their own style via black-box teacher editing (WebKD)
SummaryThis paper addresses the high cost of closed frontier web agents by proposing WebKD, which distills a black-box teacher into a smaller open-weight student. The teacher finds the first mistake in a failed episode and rewrites the student's reasoning, states, and actions from that point, in the student's own style. The authors report that WebKD students outperform behavior cloning, outcome-reward RL, and teacher-editing baselines on TimeWarp, and transfer to unseen versions and the live web on OnlineMind2Web.
In briefRent-Seeking Language Models: Reward Models Set the Prices, Frequent Audits Curb the Rent
SummaryThis paper reframes reward model overoptimization in RLHF, usually called reward hacking, as rent-seeking, where a policy earns reward from a proxy reward model that true quality does not support. In experiments with three model families and two gold reward models, RL converted 41% to 66% of its proxy gain into rent while gold reward plateaued. At a fixed expected fine, frequent mild audits outperformed rare severe ones.
WhyRecasts reward overoptimization as rent-seeking with economic metrics and testable predictions; novel, but evidence limited to a fixed testbed with three model families.
In briefCTIA-ODE: Continuous-time attention between variables for long-term time series forecasting
SummaryThis paper proposes CTIA-ODE, a forecasting framework built around Continuous-Time Inter-Variable Attention, which runs inside an ODE solver so the internal trajectory can query external variables without explicit time alignment. To stabilize the continuous dynamics, it adds a Linear Residual Augmentation module and an Outlier-Robust Hybrid Loss. Experiments on eight real-world datasets report state-of-the-art forecasting performance across datasets and horizons, with a favorable accuracy-efficiency trade-off, especially for long-horizon forecasting.
In briefMAPLE: A Multi-Axis Lens for Judging How LLM Knowledge Extraction Methods Work
SummaryThis paper introduces MAPLE, a framework for measuring how knowledge extraction enhancement techniques, such as activation steering, decoding adjustments, and prompting, affect answers from large language models. It uses diagnostic metrics at the surface, probability, and beyond-sampling levels to separate recovered answers, disrupted answers, gold-token rank shifts, and answers missed by fixed-budget sampling. The authors evaluate six techniques across OLMo2, Qwen2.5, and Llama3 and report distinct effects and trade-offs for each.
In briefWhat matters for world modeling in vision-language-action models?
SummaryThis paper examines how to integrate world models into vision-language-action models (VLAs) so that robot policies can anticipate future dynamics rather than only mapping observations to actions. The authors group existing approaches into four paradigms and compare them systematically within one unified framework, using various world representations and policy backbones across many settings. They report that 2D semantic representations give robust gains across paradigms, the decoupled MoT Future Expert performs competitively overall, especially on out-of-distribution robustness, and world modeling improves over action-only baselines, notably on long-horizon and dynamic tasks.
In briefPhyMetric: Diagnosing Physical Realism in Text-to-Video Models via Scene-Level Questions
SummaryText-to-video models can generate realistic-looking videos that still break the physical outcomes implied by a prompt. PhyMetric is a diagnostic benchmark that evaluates physical plausibility through verifiable, scene-specific yes/no questions, with 1,115 scenes and 8,920 questions across 12 dimensions. Experiments show that targeted scene-specific questions are essential, since replacing them with a single generic plausibility query drops evaluator accuracy to chance. Structured physical priors further help align scores with human judgments.
In briefUsing Pretrained Visual Features as Teachers for Compact Latent World Models in Visual Planning
SummaryThe paper proposes Anchor-WM, which uses pretrained DINO visual features as a teacher to learn a compact latent state, rather than as the world model's state itself. A training-only distillation head helps the latent keep visual information and avoid representational collapse. At inference, the DINO encoder and distillation head are discarded, so prediction and planning run entirely in the learned latent space. The authors report the best average success rate across four environments and substantially faster planning than DINO patch-feature world models.
In briefJointServe: Serving Multi-Model Agents on Memory-Constrained GPUs
SummaryThis paper presents JointServe, a serving system for LLM agents that coordinate several expert models on GPUs with limited memory. It manages model weight shards and KV cache blocks as competing residency objects within one memory budget, using prefetching, partial residency, and a transfer scheduler that guarantees bandwidth for live sessions. According to the authors, it cuts session completion time by 21.7% against the strongest whole-model swap baseline and by up to 49% on replayed agent traces.
In briefActionOPD: Distilling Decisions, Not Tokens
SummaryThis paper proposes ActionOPD, an action-level on-policy distillation method for transferring the abilities of language-model agents into smaller models. At states the student encounters, both the student and a black-box teacher sample actions, and student actions are rewarded for semantic agreement with the teacher's actions. The method needs no teacher reasoning text or token probabilities and does not require shared model families or tokenizers. Training Qwen3-4B and Qwen3-8B yields average gains of 3.3 to 6.9 points across four out-of-distribution benchmarks.
In briefOne Monitor Does Not Fit All: A Runtime Verification Framework for Agentic AI
SummaryThis paper presents AgenticRV, a runtime-verification architecture for agentic AI systems that separates observation boundaries from the monitoring formalism. It formalises conditions under which observations suffice, when replacing a monitor preserves decisions, and when blocking prevents unsafe effects. A Python implementation supports LangGraph, MCP, and A2A, and a controlled LangGraph experiment with ablations shows that both an external temporal monitor and a programmatic report monitor are needed for the selected requirements.
In briefModel or User: Which Values Win? Inference-Time Value Alignment in Multi-Turn Chats
SummaryThis paper studies whether language models can be steered toward particular values during inference in multi-turn conversations, without retraining. The authors compare four prompt- and activation-based steering methods when the user and assistant push toward the same or opposing values. They report that when the two pull in opposite directions, the side matching the assistant's baseline bias often prevails, and early apparent success can fade or reverse after a few turns.
Enrico Liscio, Anne Arzberger, Aditya Shankar et al.
In briefWhen Depth Helps: Layerwise Inference in Tabular Foundation Models
SummaryThe paper studies all 24 in-context-learning layers of the Transformer-based tabular foundation model TabPFN-3 across 272 classification datasets. It combines representation geometry, prediction trajectories, truncation, and structural interventions. The authors report that stronger late dimensional contraction is associated with larger gains in class-prototype separation, and that repeating a pretrained block or perturbing layer order reduces predictive performance on average.
In briefVaryST: A Generalizable Spatio-Temporal Forecasting Model Beyond Fixed Settings
SummaryThis paper introduces VaryST, a compact spatio-temporal forecasting model trained from scratch that builds predictive context from available long-term observations. It derives node representations from long histories, dynamically selects relevant support nodes beyond predefined neighborhoods, and selectively masks historical patches during training. Experiments on six real-world datasets against seventeen baselines report strong in-domain accuracy and improved zero- and few-shot generalization under spatial and temporal changes.
In briefCalibrating and Budgeting Risk for Automated Yes/No Decisions in LLM Systems
SummaryThis paper studies how to set thresholds for binary decisions made by LLM workflows, such as allowing or blocking inputs, when human review is unavailable. Uncertain cases fall back to a conservative default action, and false omission risk is controlled within the inputs assigned the utility-maximizing action. Across six datasets covering moderation, safety, and reranking, the authors compare general System One models, task baselines, and a local open-weights fine-tuned model. Naive threshold selection exceeded the target risk in about half of settings, versus 7.8% for Learn Then Test (LTT).
In briefReproBAIT: Can AI coding agents rebuild biological AI tools from limited descriptions?
SummaryThis paper introduces ReproBAIT, an evaluation that tests whether coding agents can design, implement, and train published biological AI tools when given only limited descriptions. The agents run in a sandbox without internet access and are scored on held-out data using tool-specific metrics. The tools tested include ProteinMPNN, RFdiffusion3, and a closed-source molecular design tool. The authors find substantial reproduction for some tools, limited success for a highly complex one, and moderate evidence of memorisation, suggesting that publicly released tool information could undermine commercial value or managed-access strategies.
Shiv Muthupandiyan, Tolga H. Dur, Mac Walker et al.
In briefEfficient contextual partial monitoring with general function approximation
SummaryThis paper studies finite stochastic partial monitoring games with side information, where the learner receives only partial feedback symbols. Building on the Estimation-to-Decisions meta-algorithm, the authors propose E2D.PMSide, a computationally efficient framework with two tractable exploration rules that achieve optimal regret orders in T. The method is evaluated on a synthetic label-efficient prediction game and an LLM routing task, where it outperforms prior approaches.
Kapilan Balagopalan, Hao Qin, Kwang-Sung Jun et al.
In briefImproving category-aware active domain adaptation using uncertainty-guided region coverage
SummaryThis paper studies active domain adaptation, which aims to reduce annotation costs and improve unsupervised domain adaptation performance. To address the sensitivity and redundancy of point-wise selection, the authors propose a region-based method that maximizes positive manifold coverage and add an uncertainty score to learn a better target-domain decision boundary. Experiments across multiple domain adaptation benchmarks show it outperforms the first category-aware baseline and existing active domain adaptation methods.
In briefAsymptotically exact negative guidance for diffusion models via positive-unlabeled learning
SummaryThis paper proposes a negative-guidance method for diffusion models that steers generation away from an undesired condition, specified by reference samples or a conditional score. The method samples from an explicit target density whose avoidance strength controls how much probability mass is suppressed near that condition. It estimates membership in the undesired condition at each timestep as a positive-unlabeled learning problem and corrects residual bias with Sequential Monte Carlo. On Gaussian mixtures it approaches the ground-truth target and outperforms existing methods on trade-offs between avoidance, quality, bias, and diversity.
In briefDelta Tokenizer: Fast CPU tokenization for repeated agent turns and long prompts
SummaryThis paper presents Delta Tokenizer, a CPU front end that speeds up tokenizing LLM requests. For multi-turn agent sessions, it re-encodes only the appended text and the affected tail, reusing the previous token sequence. For long prompts, it splits text at positions where tokenizer rules keep pieces independent and encodes them in parallel on multiple cores. The authors report outputs identical to the reference tokenizer and lower time to first token in SGLang.
In briefEvolve Together, Fall into Place: How Team Social Structure Shapes Self-Evolving AI Agent Groups
SummaryThis paper studies whether the social structure of a self-evolving multi-agent AI system, including team composition and how members depend on each other, affects how the team improves. It introduces AGORA, a framework that varies these social factors across three evolution mechanisms. Tested on web search, codebase understanding, and document QA, the right social structure outperformed a strong agent evolved alone by up to 25 points.
In briefLearning perception-to-action rules via reinforcement learning to improve theory-of-mind reasoning
SummaryThe paper observes that large language models can infer an agent's mental state in theory-of-mind tasks but often fail to use that inference when predicting its actions. The authors generate concise if-then rules linking perception, belief, and action, and train Qwen2.5 and Llama3 models with Group Relative Policy Optimization (GRPO) across four ToM benchmarks. Rule-generating GRPO outperforms prompting, supervised fine-tuning, and think-style reasoning, and rules from models of 3B parameters or more improve GPT-4o-mini's downstream performance.
Soyoung Oh, Nursulu Sagimbayeva, Nicholas Jennings et al.
In briefStateCraft: adding abstract-state reasoning to video world models
SummaryVideo world models simulate environments with realistic appearance and dynamics, but must track how past actions change the world. This paper proposes StateCraft, which separates an updatable abstract state from the visual state and pairs a language model with a video diffusion model. On a benchmark for generative interfaces, the authors report high self-consistency and generation quality over long trajectories, and a proof-of-concept 3D experiment.
In briefJetSpec++: Foresight Parallel Drafting with Self-Refinement for Faster Speculative Decoding
SummaryThis paper studies speculative decoding, where a small draft model proposes tokens that a larger target model then verifies, aiming to speed up text generation. The authors introduce JetSpec++, which combines intra-block self-refinement, inter-block foresight drafting, and pipelined drafting and verification to hide most of the drafting cost. The paper reports a wall-clock speedup of up to 7.45 times over autoregressive decoding and 1.54 times over DSpark on GB200, and says the target model's output distribution is preserved.
WhyCo-optimizes acceptance rate and drafting cost in speculative decoding via self refinement and pipelining, with substantial speedups reported across seven benchmarks.
In briefOne agent per zone: decentralised building climate control on low-cost edge hardware
SummaryThis paper proposes a decentralised controller that gives each building zone its own active-inference agent, with no messages exchanged between agents. Each agent estimates unobserved coupling through a latent disturbance and other floors' metered power. In EnergyPlus simulations of an office building over five unseen weather years, the fleet achieved the lowest mean building peak (88.3 kW) among the compared controllers. The whole fleet runs on one Raspberry Pi 5 with a p95 loop latency of 442 ms.
Panagiotis Kasnesis, Lazaros Toumanidis, Christos Chatzigeorgiou et al.
In briefDo More Self-Distillation Methods Yield Better Agents? A Controlled Comparison in Agentic RL
SummaryThis paper compares ten self-distillation methods against GRPO for training AI agents on ALFWorld, ScienceWorld, Search-QA, and WebShop across three model sizes. The study reports 132 training runs totaling 12,124.4 recorded H800 GPU-hours, with repeated evaluations over eight rollout seeds. The authors find that gains over GRPO are limited, no method consistently wins, strong final scores can hide unstable training, and extra training cost does not reliably improve performance.
In briefLiFi: LiDAR Generation from Multi-View Images with Geometric and Semantic Guidance
SummaryThis paper presents LiFi, a framework that generates LiDAR point cloud scenes aligned with multi-view camera images. It guides a latent diffusion process through geometry and semantic branches, and adds a dual-branch balanced classifier-free guidance strategy. The authors report that experiments show LiFi outperforms existing methods in producing high-fidelity LiDAR scenes.
In briefFPV-CRASH: A dataset for locating crash moments in drone first-person-view videos
SummaryThis paper presents FPV-CRASH, a dataset of 1,052 drone first-person-view clips, evenly split between crash and normal flight, with single-frame crash-moment annotations for crash clips. The authors also provide a heatmap-based reference model that estimates a continuous crash timestamp from sparsely sampled frames. Under a fixed clip-level protocol, the model reduces mean absolute error from 1.12 s for the strongest timing-prior baseline to 0.56 s.
In briefReinforcement learning with supervised-style training: a stable policy method that uses advantage guidance
SummaryThe paper introduces Advantage Guidance, a policy extraction method that keeps a simple supervised learning structure conditioned on advantage. It adds contrastive diffusion guidance during sampling to target the KL-regularized optimal policy. The authors report that it significantly outperforms prior state-of-the-art RL methods on 26 long-horizon, sparse-reward tasks.
In briefLearning Who Benefits Most: Causal Ranking for Health Stratification
SummaryThis paper presents PROMETHEUS, a framework that treats population health stratification as a multi-treatment causal ranking problem. It orders eligible patient-treatment opportunities by expected incremental benefit to help prioritize care under resource constraints. The method combines doubly robust causal signals, a Siamese neural ranking model, and contrastive pretraining, and is tested on synthetic and semi-synthetic data using administrative covariates from ASL Benevento. The authors describe the approach as preclinical and state that validation with observed treatment-outcome data is required before clinical use.
Rocco Zaccagnino, Martina Girolamo, Alberto Lombardi et al.
