In briefWould an LLM Match into a Residency? Testing Open Models on Spanish Medical Exams
SummaryThis paper evaluates four open language models of 8 to 27 billion parameters on 1,552 questions from eight editions of Spain's MIR residency exam, scored under the exam's rules. The three larger models would rank roughly between 600 and 3,000 when answering every question. Medical fine-tuning reduced overconfidence but did not improve how well confidence separates correct from incorrect answers. Many correct answers rejected distractors for clinically unsound reasons, and removing or swapping the contextual images left accuracy unchanged.
Miguel Fernandez-de-Retana, Unai Zulaika, Ruben Sanchez-Corcuera et al.
In briefLearning reusable path objectives from trajectory observations for scientific simulations
SummaryThis paper studies action learning, which recovers a stochastic action from observed system trajectories to describe how dynamical systems evolve. The authors introduce LEAP, which anchors the unknown action to a tractable reference process and learns a correction by comparing observed and reference trajectories, yielding a differentiable objective on path space. The paper reports that LEAP recovers underlying dynamics and captures cross-time dependencies in stochastic physical systems, many-body orbital dynamics, and molecular simulations, and transfers across inference objectives and physical conditions.
In briefWho Is Sycophantic to Whom? Confident LLMs and Human Deference
SummaryMost sycophancy research asks whether LLMs tell users what they want to hear. This study reverses the question, asking whether users defer to a confident, consistent but wrong LLM. In an experiment with 32 participants learning science topics from five assistants built on GPT-5-mini, between 0 and 100% of facts were replaced with plausible falsehoods, all stated confidently. Quiz scores dropped from 7.4 to 2.9 out of 10, and participants held absorbed falsehoods with high confidence, while their confidence stopped distinguishing right from wrong answers.
WhyReverses sycophancy to measure user deference to confident wrong LLMs; a 32-person experiment shows absorbed falsehoods and collapsed confidence calibration.
In briefWhat Makes a Good Unmasking Policy? Controlled Study of Learned Unmasking in Diffusion Language Models
SummaryThis paper studies learned unmasking policies that decide which tokens to commit in masked diffusion language models. The authors separately examine the teacher's labeling of committable predictions, the states the policy imitates, the inputs it observes, and how much the number of function evaluations (NFE) must drop to offset policy overhead. They propose a policy trained on trajectory-contraction targets using confidence and hidden states, which achieves the best accuracy-NFE frontier on LLaDA-8B-Instruct and largely transfers to Dream-7B-Instruct.
In briefRethinking Sequence Compression with Test-time Regression
SummaryThis paper proposes TTR-Compressor, which treats input tokens as a regression problem whose fitted coefficients serve as the compressed representation. The regression is learned end-to-end while its coefficients are computed in closed form during the forward pass. It outperforms attention-based compressors on RULER for block-wise sparse attention, and for visual token merging in multimodal models it consistently beats the standard MLP merger, matching its performance with 55.6% fewer visual tokens.
In briefPaper overview: Why graph soft-prompt tokens in LLMs can collapse into generic prompts
SummaryThis paper studies G-Retriever and related methods, which use graph-neural-network tokens to inject graph structure into LLM prompts. The authors find that these graph tokens become nearly indistinguishable across inputs, a failure they call token collapse. A learned constant prompt that never accesses the graph matches or beats G-Retriever on three benchmarks, suggesting the tokens act mainly as generic soft prompts. The authors propose a two-stage procedure to avoid collapse and recommend interventions and decoding probes to evaluate graph-token alignment beyond task accuracy.
WhyIdentifies token collapse in G-Retriever and analyzes it via interventions and probes across multiple benchmarks, encoders, and LLMs, questioning prior graph-prompting claims.
Andrea Giuseppe Di Francesco, Rishabh Jain, Franco Maria Nardini et al.
In briefCellTrack-Bench: A Microscopic Cell Tracking Benchmark for Vision-Language Models
SummaryThis paper introduces CellTrack-Bench, a benchmark that tests how well vision-language models track cells over time compared with specialized cell-tracking algorithms. It draws on 15 standard datasets across four imaging types, containing 1,537 problems covering 9,041 cell-linking decisions. The authors also built a harder subset and evaluated 9 closed-source and 13 open-weight models, analyzing their failure modes.
Ilan Theodoro, Jord\~ao Bragantini, Lo\"ic A. Royer
In briefSquaring-the-Sphere: Seamless, distortion-free 360-degree panorama generation using a new CubeAtlas layout
SummaryThis paper presents Squaring-the-Sphere, a framework that generates 360-degree panoramas using CubeAtlas, a representation that unfolds cubemap faces into one compact canvas. A pretrained diffusion transformer models the full panorama in a single pass, using a layout-aware sparse attention mask to reduce computation. The authors report improved geometric fidelity, cross-view consistency, and high-resolution visual quality compared with previous methods.
In briefBroader Green and Red Concepts for AI, Big Data and ICT: A Multidimensional Framework
SummaryThis paper proposes generalized green AI and red AI concepts, extending the classification to big data, ICT, and electronic and electrical engineering. It uses a unified Environmental-Economic-Social-Technical/Operational framework to assess resource and environmental impacts. The authors analyze how risks spread across AI, big data, and ICT stacks using electricity, carbon, water, and e-waste indicators, and they organize countermeasures into a four-tier taxonomy.
In briefWhen to Follow the Teacher: A Method That Weights Teacher Guidance by Its Benefit to the Student
SummaryThis paper proposes Advantage-Based On-Policy Distillation (AOPD), a training method for students learning from a teacher model. Instead of imitating the teacher at every step, AOPD weights teacher actions by their counterfactual advantage under the student, reinforcing helpful actions and downweighting unhelpful ones. The objective can be estimated from standard student rollouts and terminal rewards, without extra rollouts. Reported experiments on math, coding, and MedQA show gains over distillation and reinforcement learning baselines.
Mahdi Sabbaghi, George Pappas, Hamed Hassani et al.
In briefDecentralized bi-level optimization for orbital transport of a rigid payload by spacecraft swarms
SummaryThis paper addresses moving large, non-cooperative objects in orbit with a swarm of small spacecraft instead of a single monolithic vehicle. It proposes a decentralized bi-level framework that splits the problem into an inner convex thrust allocation and an outer search over payload torque trajectories, using exact gradients. The authors report that the method yields fuel-efficient solutions quickly without full convergence, and they validate it through Monte Carlo simulations, including closed-loop recovery from in-flight faults.
Sidhdharth D. Sikka, Revathy Ravuri, Hei Yin Jowett Chan et al.
In briefAI in Mechanical Engineering: Opportunities, Challenges, and Responsibilities
SummaryThis paper, drawing on a round-table at COBEM 2025, offers a structured perspective on how artificial intelligence is reshaping mechanical engineering, from modeling and simulation to diagnostics and design. It examines how AI can enhance practice while preserving scientific rigor, covering physics-informed learning, uncertainty quantification, validation, certification, and the energy demands of AI computation. The authors conclude with principles for responsible adoption, stating that AI should complement, not replace, engineering judgment.
In briefDiffusion language models for more natural black-box attack prompts that are harder to detect
SummaryThe paper studies continuous-space attacks that find prompts triggering behaviors such as overthinking in black-box LLMs. Existing methods decode Gaussian latents with a random mapping, producing unnatural prompts that input guardrails can detect. The authors couple the same latent framework to a continuous diffusion language model, which learns the distribution over language. On RareTrap (sampling-based) and ThinkTrap (optimization-based), character-level perplexity drops from 10^3–10^7 to 3–5 and from 20–28 to 4.3–5.4 respectively, and risk estimates can differ by an order of magnitude between the two mappings.</br>
In briefContestable AI Agents for Recording Human Review of Governance Metadata
SummaryThe paper proposes contestable governance, in which specialist agents give scored, cited arguments on disputed metadata decisions and a human steward can accept, reject, modify, or add arguments before commit. A governance record keeps the evidence, interventions, and outcome, and a gate controls which verified decisions become precedent. Two HVAC use cases illustrate the approach: point classification using Brick Schema class labels, and an illustrative warranty failure-mode decision.
In briefAnalyzing catastrophic forgetting in continual learning through algorithmic stability
SummaryThis paper studies catastrophic forgetting, where learning new tasks degrades performance on earlier tasks in continual learning. Without assuming a specific model architecture, it uses an algorithmic stability framework to analyze forgetting under SGD, splitting it into generalization gaps and empirical forgetting. The authors compare constant and decaying learning-rate schedules, and experiments on Split CIFAR-100 and Split MiniImageNet examine the theoretical findings.
In briefPrivacy auditing without retraining, using shift-aware score transfer
SummaryThis paper studies observational privacy auditing, where the auditor cannot insert canaries, control data, or retrain the target model and must rely on known members and non-members. Because available non-members can differ in distribution from the training data, membership-inference scores become skewed. The authors learn how representations relate to scores using out-of-distribution non-members, transport this relationship to the target distribution, and add a conservative correction. Across several datasets with distribution shifts, the method yields more stable privacy estimates while keeping lower bounds tighter than baselines.
In briefSame Luck, Different Choice: Twin-World Credit Assignment for Agentic RL
SummaryThis paper addresses how to credit individual decisions inside an agent's episode in reinforcement learning. It proposes TWICE, which makes counterfactual branches inherit the factual episode's sampling noise so that outcome differences reflect the decision itself. On seven question-answering benchmarks, TWICE raises GRPO's average exact match by 5.3 and 4.5 points on Qwen2.5-3B and 7B, and uses 68% fewer auxiliary tokens than VinePPO.
WhyCoupled Gumbel noise twin worlds reduce counterfactual credit variance, improving sample complexity to Θ(Δ^-1); validated across seven QA benchmarks against strong baselines.
In briefDecentralized real-time planning for multi-robot teams via negotiated dispersive trees
SummaryThe paper proposes WarpWeaver, a decentralized real-time planner for coordinating multiple nonlinear robots toward a near-joint-optimal solution. Robots negotiate rents on shared space through decentralized column generation over dispersive command trees, yielding an agreement certified near-optimal at quiescence. Simulations and six-robot onboard hardware experiments are reported.
WhyExtends a single-robot method to decentralized column-generation negotiation with certified near-optimality, backed by simulations and six-robot hardware experiments.
Shashank A. Deshpande, Blake T. Edwards, Jonathan P. How
In briefPYTHIA: a transferable machine learning method for optimizing defects and adsorption sites together
SummaryThis paper presents PYTHIA, a framework that uses foundation models and transfer learning to apply machine learning to under-studied materials with few samples. A complex model is first trained on a large dataset of chemically similar molecules, then its weights are frozen and Transformer and Gaussian process layers are added for subspace alignment and uncertainty quantification. The resulting model serves as a surrogate for Bayesian optimization and was applied to adsorption and defect optimization in two distinct systems. The authors report that it identified optimal adsorption sites efficiently, including for complex cases.
In briefLACToS: Local Anomaly Detection Using Superpixel Conditional Typicality
SummaryThis paper proposes LACToS, a method for finding small anomalous or out-of-distribution regions within an otherwise normal image. The authors note that existing methods, mostly designed for industrial defect detection, perform poorly when normal data is multimodal. LACToS is based on statistical typicality, grouping pixels into superpixels and using an efficient proxy to estimate typicality. It is evaluated on medical imaging, human faces, and X-ray security screening, where it performs more robustly than prior methods.
Jihoon Hong, Alireza Kheirandish, Rohan Sanda et al.
In briefEmpirical study of recent LLMs for knowledge graph-to-text generation
SummaryThis paper benchmarks recent large language models against fine-tuned encoder-decoder models on converting knowledge graph triples into natural language. It uses reference-based metrics and graph-grounded, reference-free LLM-based evaluation validated against human annotations. Recent LLMs substantially narrow the gap with fine-tuned models on reference-based metrics and perform better on graph-grounded metrics.
In briefDiona: A Regression Method Using Direct Inference on Admissible Sets
SummaryThis paper proposes Diona, a regression method that estimates a function's state and derivatives at a query point directly from samples, with admissible bounds, instead of long global training. It treats values and derivatives at the query and neighbouring samples as latent variables in a constrained optimisation problem and establishes finite-sample coverage and, for noiseless data, convergence under dense sampling. On a six-function benchmark of three to eight dimensions, Diona was more accurate than a Gaussian process on noiseless, denser data and its bounds contained the true value at over 95% of query points for four functions. Its time to first prediction after a data update was much shorter than the GP's and a neural network's, but this advantage holds only when few predictions are needed between updates.
In briefNonparametric autoregressive joint distribution estimation using parallel gradient boosting
SummaryThis paper studies estimating joint probability distributions by factorizing them autoregressively into conditional distributions estimated with flexible nonparametric methods. For loss functions with a subadditivity property, the authors derive convergence-rate bounds for the autoregressive estimator from those of its conditional estimators. When conditioning sets are sparse, the difficulty is characterized via variable elimination in an undirected dependence graph. An autoregressive estimator based on parallel gradient boosting shows competitive performance on various simulated and real datasets.
In briefSynthesizing Complete 3D Prostate MRI Scans with Diffusion Models
SummaryThis paper presents a conditional 3D diffusion framework that generates complete prostate biparametric MRI from anatomical masks. It supports sequential synthesis, where T2w and diffusion images come from separately trained models, and joint synthesis, where a single model generates all modalities at once. Five configurations were trained on the PI-CAI, FastMRI Prostate, and PROSTATEx datasets, and both pathways were shown to be feasible. Directly synthesizing ADC and high-b-value DWI performed better than deriving them from synthetic diffusion-weighted images.
In briefBeyond Causal Masks: Detecting Hidden Future-Token Leakage in Adaptive Sequence Models
SummaryThe paper studies a hidden flaw in autoregressive models where a sequence-level routing summary lets prefix positions see future tokens, even though the causal attention mask is correct. An anomalous 24-layer checkpoint lowered validation loss from 3.5958 to 1.4259 but failed a prefix-invariance audit and had much higher reference-model perplexity during generation. The authors traced the leak to a full-sequence summary broadcast through the dynamic routing-query path, and two independently trained 12-layer causal repairs removed it, bringing loss and perplexity back to control levels.
In briefJev-DLC: Learning from typed user corrections to decisions over time
SummaryThis paper studies how user feedback given after deployment, by accepting a decision or revoking it in favor of another, can persist and shape later requests. Each correction is treated as evidence about a recurring action and as a directed link between the preferred and rejected choices, stored in bounded channels above a frozen scorer. The method keeps no raw interaction history and adds no scorer calls. The authors also introduce TCPA-Bench, a chronological benchmark that credits a correction only through its effect on later decisions.
In briefThe Geometry of Metaphor: Probing and Steering How LLMs Represent Figurative Language
SummaryThis paper uses mechanistic interpretability techniques to identify causal subspaces linked to metaphor in large language models across different architectures and sizes. The authors show that steering these subspaces changes outputs on classification tasks spanning different parts of speech. They also report that the same techniques can control how metaphorical the generated text is.
In briefFrom Prompt Rules to a Golden Data Model for Compliant Pharma Claims
SummaryThis paper is a production case study asking whether moving compliance and audience personalization signals from prompt-level rules into structured context at retrieval time improves LLM-generated pharmaceutical claims. It uses Composer, an internal tool built on a Golden Data Model of 1,867 claims across four products. The authors report gains in indication, content type, target audience, a six-dimension quality score, and claim reuse, plus a 29.5% reduction in weekly API cost during deployment.
In briefTraining user simulators that stay on task while sounding natural
SummaryThis paper proposes a reinforcement learning method for user simulators that rewards both adherence to complex task intents and conversational naturalness. Naturalness is measured by a lightweight discriminator trained adversarially and alternately with the simulator. The authors release UserLM-2.0, which produces valid simulated conversations 46.2% of the time, compared with 14.1% for the strongest prior user simulator.
Mustafa Omer Gul, Philippe Laban, Tanya Goyal et al.
In briefDuoCook: A Visual Planning and Coordination Benchmark Inspired by Overcooked
SummaryThis paper introduces DuoCook, a benchmark inspired by the game Overcooked for testing how vision-language agents coordinate two players in cooking tasks. It uses 20 maps built in Unreal Engine, where one model controls both players from top-down visual observations. Four models are compared on task score, order completion, and other metrics, and a memory ablation examines the effect of a shared self-authored notebook.
In briefPiano2Pop: turning piano demos into full pop arrangements without paired training data
SummaryThis paper introduces Piano2Pop, a task that converts piano MIDI into instrumental pop productions. The method conditions a pretrained flow-matching music generation model on piano rolls, training on synthetic piano-instrumental pairs instead of aligned piano-pop data. On 476 real piano-cover/pop pairs, the pair-supervised approach achieves the best Frechet Audio Distance and related scores, and all fine-tuned systems significantly outperform zero-shot baselines.
Yun-Chen Cheng, Chih-Pin Tan, Tzu-Hung Huang et al.
In briefFairFRESH: A fair, explainable neuro-symbolic framework for household food recommendation
SummaryThe paper treats household food recommendation as constrained fair division, since members share a limited pantry and follow different dietary rules. FairFRESH learns per-member utilities, allocates foods under explicit welfare objectives, enforces dietary rules with a deterministic filter, and explains results using Shapley and Banzhaf attributions. On NHANES-derived data, its Recall@20 of 0.293 beats five learned recommenders but not a previous-day repetition heuristic (0.321). On held-out four-person households, Nash allocation cuts Gini disparity by 42% versus conflict resolution, and all recommendations satisfy the encoded dietary rules.
Kodur Sai Vinay Sathvik, Kakoli Bora, Tejaswini R. M
In briefEvaluating Uncertainty-Aware Ensembles of Hippocampal Segmentation Software
SummaryThis study evaluates ensembles of five hippocampal segmentation methods (InnerEye, ASHS, FastSurfer, HippoSeg, FreeSurfer) across two dementia datasets and one epilepsy dataset. A sub-ensemble of InnerEye, FastSurfer, and HippoSeg was found to be both accurate and efficient. Ensemble-derived uncertainty measures flag low-confidence cases, which may help route ambiguous segmentations to expert review. The uncertainty also correlates significantly with motion artefact severity and modestly contributes to Alzheimer's disease classification.
Gabriel Oliveira-Stahl, Anna Schroder, James Moggridge et al.
In briefStructure-preserving reduced-order models of thin panels in supersonic and hypersonic flow
SummaryThis paper builds structure-preserving reduced-order models for two AePW-4 High-Speed Working Group problems: a clamped panel at Mach 1.92 (RC-19) and a cantilever plate at Mach 5.85 (HyMAX). Each problem uses a Ritz-Galerkin reference model and a learned model that keeps the reference potential and trains only a non-energy-adding dissipation term. On RC-19 the predicted pressure ratio of 1.521 matches shock relations and pressure-sensitive paint measurements, and the mean dominant frequency lies within 1.0% of the measured 236 Hz.
WhyProposes an energy-consistent learned reduced model, validated against shock theory, PSP measurements, and a published model on two flutter cases.
In briefSABLE: A Scalable Bayesian Method for Optimizing in High-Dimensional Latent Spaces
SummaryThis paper introduces SABLE, a Bayesian optimisation method for searching high-dimensional generative-model latent spaces when evaluations are expensive and the budget is small. It uses a Bayesian affine surrogate and splits each query into an in-span component chosen by expected improvement and an off-span component set by a finite-horizon schedule. The authors report high sample efficiency on synthetic benchmarks and latent-space optimisation tasks, with decision times reduced by several orders of magnitude at large budgets.
Erik Bodin, Lisa Maria Kreusser, Donney Fan et al.
In briefAgentVQA: A Unified Benchmark for Testing Vision Models as Agents
SummaryThis paper introduces AgentVQA, an offline benchmark for testing vision-language models on agent-style tasks across five domains: web agents, robotics, egocentric videos, games, and spatial planning. It contains 1000 challenging real-world examples drawn from 23 source datasets, and its model rankings match those from online evaluations. The best of 22 evaluated models reaches only 58.5% accuracy, with errors mainly caused by failing to accurately track and represent the current and changing visual state.
In briefDirCast: Live esports directing and commentary with one streaming video model
SummaryDirCast is a single streaming video-language model that directs live esports broadcasts by deciding which camera to show and commentates on it, using only past footage. It scans all cameras at low resolution, decides whether to switch, and reads the chosen camera at high resolution before speaking. Tested on 27 professional Dota 2 matches, it directs better than rule-based and zero-shot baselines including GPT-5, and its commentary wins 29.2 percent of comparisons against GPT-4o.
In briefSelf-supervised sonar image restoration that removes speckle and blur without clean reference images
SummaryThis paper proposes a self-supervised method to jointly reduce speckle noise and blur in multibeam imaging sonar images, without needing clean reference images. It uses a physics-based coherent image-formation model as a forward operator and trains on registered multi-frame sonar sequences. Experiments on two real multibeam sonar datasets report a good balance between forward-prediction agreement, speckle suppression in homogeneous regions, and preservation of structural boundaries compared with baselines.
In briefTraining small models to pick the right tools for AI agents
SummaryLLM agents must choose among more than ten thousand tools, which is hard for complex multi-step instructions. This paper trains a small open model with reinforcement learning to do this selection, a role usually handled by large commercial LLMs driving a retriever. On ToolBench, the post-trained Qwen3-8B outperforms Llama-4-Maverick (402B). The trained model transfers directly to different retrievers and generalizes to new tool corpora in other domains.
Zhaoyi Zhou, Akshayaa Magesh, Jeongyeol Kwon et al.
In briefTesting whether LLM inference speedups designed for NVIDIA servers also work on Apple Silicon laptops and desktops
SummaryThis paper examines which large language model inference optimizations still help when run on Apple Silicon, a platform for local inference where most methods were developed for server-class NVIDIA GPUs. The authors test four families of techniques—KV-cache pruning, speculative decoding, ported GPU kernels, and post-training quantization—on four Apple machines under single-request decoding. They report that KV-cache pruning can reach 1.75 times higher throughput on Apple but can lower it on desktop NVIDIA GPUs, and that all tested speculative decoding methods reduced throughput on Apple. The authors conclude that a speedup depends jointly on the technique, runtime, hardware, and workload rather than on the technique alone.
WhySystematically evaluates four optimization families on Apple Silicon, showing gains depend on hardware and workload; multi-machine experiments with RTX 4090 comparisons give solid evidence.
Kyrylo Yemets, Andrii Yaroshevych, Victor Muryn et al.
In briefMoG-VLM: Motion-guided vision-language models for efficient video anomaly detection
SummaryThe paper proposes MoG-VLM, a framework that uses motion as a lightweight signal to decide when a vision-language model should run, reducing the cost of real-time video anomaly detection. It introduces Temporal Gradient Magnitude for content-aware frame sampling and a hierarchical dual-stream design that activates multi-frame VLM reasoning only around likely anomalies. The authors report state-of-the-art zero-shot, training-free AUC-ROC on three of four benchmarks and a 1.31x inference speedup on XD-Violence.
Khaled Waleed Dawoud, Muhammad Zaigham Zaheer, Khan Muhammad et al.
In briefTraining LLMs to Make Their Reasoning Easier to Monitor
SummaryThis paper studies how to detect harmful behavior in large language models from their chain-of-thought (CoT). The authors show that the existing metric TPR_mc can overestimate how often a monitor detects behaviors actually caused by an intervention, and they derive a complementary lower bound for a tighter evaluation. They build paired control-intervention training data and train models with three intervention-specific methods. Across most model-monitor setups, these methods generally improve CoT monitorability over the base models.
In briefTraining-free object identity reasoning for video-language models
SummaryVideo-language models recognize objects in single frames but struggle to keep track of the same object over time, performing near chance on real shell-game videos. The paper proposes Track, Then Bind, a training-free method that uses a promptable video tracker to assign track numbers to candidate objects, then a frozen vision-language model binds the target to a track using identifying evidence. The answer is read out from the tracked state. The method reports strong results on VET-Bench, Perception Test cup games, and TOC-Bench.
WhyA training-free track-then-bind method separating identity binding from temporal tracking beats strong baselines across several benchmarks with solid ablations, influencing video reasoning.
In briefMeasuring the verifier's floor: how much apparent rare-word speech synthesis error comes from recognition errors
SummaryText-to-speech systems are usually scored by transcribing their output with a speech recognizer, so recognizer mistakes also count toward word error rate. The authors measure these errors on human recordings of the test sentences. On LibriSpeech and FLEURS, the evaluation recognizer misses very rare words even in human speech, and checking with more recognizers lowers the apparent error, though very rare words remain the hardest.
Kirill Borodin, Vasilii Kudryavtsev, Nikita Vasiliev et al.
In briefKeeping Clinical Meaning Intact in AI-Written Patient Health Summaries
SummaryThis paper studies how large language models can turn structured menstrual health data from wearable devices into readable summaries for patients. The authors combine structured evidence conditioning with retrieval-augmented generation and evaluate 4,500 summaries from three models using NLI, atomic factuality, and readability metrics. Explicit grounding constraints improve clinical alignment, especially uncertainty preservation, but they also increase reading difficulty and medical terminology.
Bayram Ayadi, Florian Geissler, Johanna Schmidhuber
In briefAnyGameRenderer: Entity-consistent generative rendering for agent-built games, with an entity-level benchmark
SummaryAI coding agents can build small games quickly, but their visuals look like placeholder programmer art. AnyGameRenderer is a video-to-video generative model that turns such footage into high-quality visuals while keeping layout, camera, and each entity's position and timing. It is LoRA fine-tuned from MiniMax-H3 ref2va on about 10k paired clips and evaluated on a new entity-level benchmark of 59 cases and 753 entities. Compared with the base model, temporal match rises from 50.9% to 62.9%, and missed count changes fall from 34.7% to 21.3%.
In briefRheOPT: Few-shot prediction of non-Newtonian flows by pretraining on Newtonian simulations
SummaryThe paper introduces RheOPT, a transformer-based neural operator that is pretrained on inexpensive Newtonian fluid simulations and then fine-tuned on a few non-Newtonian samples. It is evaluated on a 4:1 contraction channel and on the oscillatory wake behind a triangular obstacle. The authors report that pretraining improves few-shot prediction and recovers shedding frequency and wake fluctuation amplitude that a from-scratch model suppresses.
In briefAdaptive Skill Granularity for Coding Agents: Runtime Routing and Recovery from Macro to Micro Steps
SummaryThis paper introduces Adaptive Skill Granularity, a framework that chooses how coarse or fine a coding agent's skill should be based on task structure, execution state, and macro-skill failures. It combines a feature-based granularity controller with a finite-state recovery executor. Evaluations show fixed macro execution is efficient when applicable but brittle under mismatch, while runtime routing and explicit macro-to-micro recovery improve robustness.
In briefHow Different, or Different How? Stance Granularity in Viewpoint Selection for Dialog Systems
SummaryThis paper studies how dialog systems that help users engage with alternative viewpoints should decide how far those viewpoints differ from the user's own. The authors propose a policy in which a bandit sets a maximum distance and the policy selects the most novel viewpoint within that bound, comparing a scalar stance encoder with Perspectivized Stance Vectors that split positions by concept. A between-subjects study (N=63) found that concept-level stance produced more varied exposure sequences and broader topic coverage, but did not change justification depth or reconstruction accuracy. The authors also found users reconstructed viewpoints least accurately on concepts where one side took no position.
In briefMixPT: a point cloud transformer alternating window and global attention
SummaryMixPT is a point cloud transformer that interleaves window attention and global attention blocks to balance computational cost with scene-wide information exchange. For scenes with many tokens, global attention is applied to spatially pooled tokens, and its residual updates are returned to the original-resolution stream. The authors report competitive accuracy with lower latency than representative window and global attention baselines on indoor and outdoor benchmarks, and show that the interleaved pattern can be repeated to scale depth.
In briefWhen Can an Agent Trust Its Memory? Action-Conditioned Verification for Long-Lived Agents
SummaryThis paper studies how long-lived agents should verify memories carried across sessions before taking irreversible actions. It frames the decision as a sequential stopping problem that balances execution risk, abstention, and verification cost, and proposes AMV-Guard, a gate that selects which memory premises to re-check using value of information. On ToolSandbox, the method reduces mean verification cost and increases Proceed coverage compared with a static policy given the same inputs.
In briefConsMSA: Semantic Distribution Consistency Learning for Sentiment Analysis from Multiple Modalities
SummaryThis paper proposes ConsMSA, a method for multimodal sentiment analysis that combines text, video, and audio to predict human emotions. Because raw multimodal inputs often contain redundant or conflicting information within and across modalities, the method models semantic distribution consistency using an intra- and inter-modality consistency score. This score is used to regularize training, reweight tokens, and prune redundant ones. Experiments on CMU-MOSI, CMU-MOSEI, and CH-SIMS report state-of-the-art results, and accuracy stays comparable when only 10% of tokens are retained.
In briefMulti-token prediction and alternative tokenization help models learn shortest-path search on larger graphs
SummaryThis paper asks whether a language model can generate the shortest path between two nodes on an unseen graph, a task prior work found models fail at as graphs grow. The authors attribute the failure to two causes: the supervision signal and the tokenization, and propose multi-token prediction and alternative tokenizations to address them. The two methods are shown to be orthogonal, and together they let models consistently learn the task at scales far beyond previous limits.
Arvid Frydenlund, FuTe Wong, Katherine Lambert et al.
In briefDIAL: Language models whose instruction-following level you can dial in
SummaryThis paper studies steerable language models that give fine-grained numerical control over how closely an LLM follows text instructions, such as how short a response should be. The authors propose DIAL, which combines classifier-free guidance distillation, a lightweight steering adapter with a text-conditioned classifier, to achieve fine-grained control, efficiency, and scalability together. They also introduce SteerIF, a dataset of 98,625 instructions, and report competitive balance between concept expression and fluency across several LLM families.
Thanh Q. Tran, John Russell Himawan, Arun Verma et al.
In briefConservative Meta-Agent Search: Safely Adapting AI Agent Designs Over Time
SummaryThe paper proposes Conservative Meta-Agent Search (CMAS), a method for updating multi-agent system architectures in changing environments. New designs are compared to the current one via shadow replay and are promoted only if they clear a variance-penalized margin, followed by a simulated canary stage and rollback. The authors introduce the MetaInvest-Bench simulator and report results across 96 streams, including regret and utility gains plus a 3.4% false-promotion rate, while noting the limits of these claims.
In briefRate-controlled KV re-selection that refreshes visual token caches when attention drifts, for long multimodal generation
SummaryMultimodal LLM KV cache compression typically picks visual tokens once at prefill and reuses them throughout generation, which the authors show performs poorly on long-form outputs. They propose Rate-Controlled Refresh (RCR), where attention drift decides when and where to re-select tokens while a per-sample controller sets how often. Across model-task settings, RCR outperforms the per-step retrieval baseline InfiniGen on most tested configurations.
In briefA sparse map graph for cross-session indoor navigation
SummaryThe paper proposes a sparse topometric graph that represents indoor navigable connectivity with nodes encoding local 3D geometry and distance-weighted edges, avoiding fine-grained matching. A learned graph-matching network aligns subgraphs across sessions. It is evaluated on cross-session point-goal navigation, reporting 77.1% success on Gibson and 40.0% under 3D-FRONT scene changes, with memory reduced by more than two orders of magnitude versus point-level representations.
In briefDimension-free behavior of unadjusted Langevin for ridge-separable models, via subspace robust Wasserstein analysis
SummaryThis paper studies the bias of the Unadjusted Langevin Algorithm, a common tool for high-dimensional Bayesian inference. It measures error with subspace robust Wasserstein distance, which controls means and low-dimensional marginals rather than the stricter Wasserstein-2 distance. For ridge-separable potentials built from i.i.d. data, high-probability bounds show a bell-shaped mixing profile as the sample size N varies relative to dimension d. The authors also prove that dimension-free guarantees are impossible in general near the proportional regime N of order d.
WhyIntroduces subspace robust Wasserstein analysis giving dimension-free ULA bounds for ridge-separable potentials, plus a proven impossibility result near proportional regime.
In briefPhysTrace: Interpretable Time-Based Checking of Physical Consistency in Video Generation Models
SummaryThe paper presents PhysTrace, a training-free framework that aligns expectations generated by a multimodal large language model from video history with observed video windows. It temporally localizes physical violations and gives interpretable descriptions without task-specific probes or classifiers. The authors also introduce IntPhys2-Timing, which adds violation intervals to the IntPhys2 benchmark. PhysTrace reports a Hit Rate of up to 70.7% across physical principles on IntPhys2-Timing and higher overall possible-impossible accuracy than compared methods on IntPhys2.
In briefPredictable performance is suboptimal: what it means for model routing
SummaryRouting methods choose a language model for each query based on predicted correctness, and many train separate performance predictors to guide that choice. The paper shows a predictor beats a simple self-confidence baseline only when the candidate model is suboptimal on the task or the predictor adds substantial capacity. Experiments with 16 language models on four benchmarks found the predictor's advantage grows for less optimal models and shrinks after finetuning. Two-model cascades show the same patterns.
In briefFocusVLA: Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models
SummaryThe paper proposes FocusVLA, a stage-aware multi-view routing framework for vision-language-action models in robotic manipulation. It predicts future interaction intent from historical actions and uses it to dynamically route global and local views during action generation. On RoboTwin 2.0, it improves average success rate by 10.10 points over the strongest baseline π0.5. Its intent prediction and routing modules can be integrated into existing VLA frameworks without changing their core architectures.
In briefHow Well Do Decision Models Transfer to Vietnamese? A Text Classification Study
SummaryThis study tests four specialized decision models on Vietnamese zero-shot text classification across 13 settings in six task groups. It compares English, Vietnamese, and bilingual label descriptions and examines robustness to teencode, typos, and diacritic removal. With English label descriptions, JEV performs best overall, and diacritic removal causes the largest performance drop.
Phat Huu Ha, Nguyen Xuan Phuc, Nong Nhut Quy et al.
In briefHow depth growth and looping shape in-context learning in language models
SummaryThis paper studies how language models that reuse computations across depth, through depth growth or looping, acquire, retain, and use in-context learning (ICL). Experiments cover pretrained models up to 3B parameters and controlled tasks. Grown and looped models rely more on in-context information, acquire ICL earlier than a full-depth dense reference, and retain it more strongly when fixed associations can also be learned in the weights. Depth growth propagates early induction heads across layers, and both model types depend more on induction-ranked heads for few-shot and later-token prediction.
Ferdinand Kapl, Kaitlin Maile, Stefan Bauer et al.
In briefPost-hoc feature selection layer for explaining neural networks
SummaryThis paper adapts the Feature Selection Layer into a post-hoc version that optimizes feature weights after training. It is a lightweight module attached to frozen pre-trained models on tabular data to highlight features the original model considers important. The original model's architecture and parameters remain unchanged. Experiments on synthetic and real-world datasets compare it with embedded FSL and methods such as Integrated Gradients and SHAP, reporting that it identifies relevant features while preserving predictive performance.
Jo\~ao Kenji Suwa, Jo\~ao Pedro Silveira e Silva, Bruno Iochins Grisci
In briefWhen and How Muon Preconditions: The Role of Structured Gradient Noise
SummaryThis paper studies why Muon, an optimizer popular for training large language models, works. It shows that under classical isotropic noise, Muon gives no acceleration beyond rescaling the learning rate. The authors analyze structured noise from mini-batch sampling in an associative memory model with uneven task frequencies. They report that Muon can reduce the dependence on task frequency from linear to square-root, and they note momentum helps reduce sampling noise.
In briefSyRA-CS: Meta-learning for adapting synthetic code-switched speech recognition to real speech
SummaryThis paper addresses code-switched speech recognition, where real mixed-language speech is scarce. The authors propose SyRA-CS, which uses Reptile meta-learning to train on synthetic CS-FLEURS speech and then adapts with one hour of labeled real target speech. Across four architectures and Arabic, French, German and Hindi paired with English, it improves WER over continual fine-tuning in all 16 model-pair comparisons.
Fadel Thior, Ahmer Tabassum, Haania Siddiqui et al.
In briefMajor Revision, Minor Accountability: LLMs Reviewing Top-Journal Papers
SummaryThis study has four large language models write independent peer-review reports on 200 papers published in 2026 in 20 journals, including Nature, Science, and Cell. Of 800 reviews, 75.6% recommended major revision or rejection, yet the rate varied from 17.5% to 100% across models. When one methodological or statistical flaw was planted in each manuscript, detection rates ranged from 17.0% to 60.5%. The authors conclude the outputs are an uneven, framing- and model-dependent signal.
Bastien Le Guellec, Su Hwan Kim, Raphael Bentegeac et al.
In briefSQuAT: Spherical lattice vector quantization for training low-bit language models
SummaryThis paper studies vector quantization-aware training (VQ-QAT) for low-bit large language models. It proposes SQuAT, which splits weights into direction and magnitude, quantizes directions with an E8-derived spherical codebook and magnitudes with scalar quantizers, using separate straight-through estimators for each. A distributed exact nearest-neighbor scheme cuts training time by 12.8× and scales training to 10B tokens. Under identical quantizers and initialization, SQuAT outperforms standard VQ-QAT STE and reaches state-of-the-art VQ-QAT accuracy on several models.
Sebastian Loeschcke, Ivan Koryakovskiy, Tycho F. A. van der Ouderaa et al.
In briefSame Entity, Different Script: Why Language Models Judge Names Differently Across Writing Systems
SummaryThis paper studies how multilingual language models give different identity judgments when the same person's name is written in Devanagari versus Latin script. Using 601 audited identity pairs, the authors find Qwen2.5-7B's same-person accuracy drops from 35.1% in Latin to 8.5% in Devanagari under Marathi prompts, while different-person accuracy stays above 98%. The effect replicates on 374 Malayalam pairs, and activation patching localises causal sensitivity to the second entity in middle layers.
Athira Gopal, Patil Prathamesh Pravin, Ankush Kadu et al.
In briefHERD: A hybrid method for continual learning that evolves in-context skills and distills them into weights
SummaryThis paper proposes HERD, a hybrid continual learning framework. On the harness side, it refines reusable skills and keeps hard examples the skills miss in bounded memory; on the parameter side, it uses drift-constrained distillation with dual anchors to reduce forgetting. The authors report that on the 25-stage TRACE+17 task stream, under fixed memory and compute budgets, HERD matches the strongest parametric baseline in new-task learning while showing the least forgetting and the best retention of general capabilities.
