In briefSystematic study of design choices for language model data attribution
SummaryThe paper introduces CATT, a controlled benchmark that fine-tunes language models on fictitious knowledge so attribution methods have unambiguous ground truth, and can vary one dataset factor at a time. It shows that document redundancy and lexical distractors degrade attribution performance. The authors examine the gradient loss and how per-token scores are aggregated, finding that negative log-odds gradients and aggregating only top tokens improve ground truth recall by 35 to 72 percentage points.
In briefPropose-Filter-Release: a private filtering method that scales noise to typical record sensitivity
SummaryThis paper introduces propose-filter-release (PFR), a general differential privacy primitive. It privately filters out a few records that fit poorly with the rest of the data, so noise is scaled to the influence of typical records rather than worst-case ones. PFR needs only a concave friendliness score and a friendly-DP release mechanism, and it guarantees standard differential privacy under public conditions on the number of records. The authors illustrate it on scalar statistics, principal subspace estimation, KL divergence of user histograms, effective resistances of graphs, and treatment-rule selection.
In briefMicro-Edits: Tools for Fine-Scale Film Editing Support
SummaryThis paper studies micro-edits, small frame-level adjustments in film that control timing and direct visual attention. The authors first conducted a formative study with professional editors to understand how these edits are structured and used. They then formalized this structure mathematically and built AI- and optimization-based tools to make exploring micro-edits easier, validating them through user studies and video examples.
Nhan Tran, Hijung Valentina Shin, Abe Davis et al.
In briefRethinking Neural Thickets Without Weight Perturbations
SummaryThis paper reproduces the gains reported for the Neural Thickets hypothesis, which claims that random weight perturbations in pretrained language models find task-specific experts, and shows these gains do not require that geometric explanation. Perturbing a single decoder block or adding a random bias to output logits with a fully frozen model matches or exceeds full-weight perturbations on most reasoning tasks. The authors propose Perturb-Select-Vote (PSV), a general procedure built on output diversity, validation-based selection, and voting, and extend it to vision tasks.
Nicolas Dufour, Elouan Gard\`es, Amil Dravid et al.
In brief"Read, Allocate, Route": A Role-Coupled Tokenization Method for Pixel-Space Generative Transformers
SummaryThe paper introduces SpecSlot, a pixel-space rectified-flow Transformer that separates output detail from the length of the global attention sequence. Each spatial anchor is expanded into several slots, each covering a distinct band of DCT coefficients, and only the lowest-frequency slot participates in global attention while the others interact locally. The authors report comparable or better FID on ImageNet-256 and ImageNet-512 and a higher DPG-Bench score on text-to-image generation, at lower computational cost than the compared baselines.
In briefAdaptive scenario generation for KL distributionally robust optimization around a learned density model
SummaryThe paper studies distributionally robust optimization for two-stage stochastic programs when the estimated distribution is subject to error from a finite data record. It proposes a Kullback-Leibler ball centered at a density model fitted to the record, which lets the adversary up-weight unobserved costly realizations in proportion to the model's probabilities. To handle the limited number of scenarios, the method alternates between solving the finite problem and drawing new realizations from the model, steered toward the worst case at the current decision. On a multi-item newsvendor problem and a grid capacity expansion problem, it lowers cost under distribution shift when data is scarce and the model extrapolates well.
In briefLuggage re-identification: strategies for real-world use
SummaryThis paper studies luggage re-identification, which matches the same bag across disjoint camera views in airports, train stations and border crossings for security and lost-luggage recovery. The authors note that existing methods mostly work on clean cropped images, while real scenes involve occlusion, cluttered backgrounds and crowds. They evaluate an end-to-end pipeline that detects and re-identifies luggage from multiple video streams, comparing ways to select and merge views and shots and training on different environment datasets.
In briefFrom extreme weather to farm losses using agentic AI
SummaryThe paper presents WxCrop, an agent that judges agricultural hazards from physically interpretable weather indices checked by causal analysis, and combines weather and agricultural data to predict crop losses for past and forecast dates. It is demonstrated in Florida for hurricanes, freezes and dry heat. Reported checks include a hurricane index AUC of 0.98 for detecting cyclone days, held-out loss reproduction within 2%, and a 0.63 rank correlation with independent production records across counties.
WhyThe authors present an agent combining causal analysis of weather indices with crop loss prediction, reporting validation on Florida hurricanes, freezes, and dry heat.
In briefSonoWorld4D: Turning One Video into a Sound-Enabled 4D Scene
SummaryThis paper introduces Video-to-4DAVScene, a task that converts a single monocular video into an explorable, time-varying 4D audio-visual scene. The SonoWorld4D framework generates a 360-degree static background, reconstructs motion trajectories of sound-emitting moving objects, and synthesizes a sound field across sources via dual-space audio diffusion synchronization. Evaluations on real-world dynamic scenes and a user study indicate the resulting spatial audio is perceptually realistic and consistent with scene geometry and visuals.
In briefworld2vec: Learning mechanistic world models from latent dynamics
SummaryThis paper introduces world2vec, a method that learns world models directly from observations and actions without using reward information. The learned latent dynamics reproduce the true system faithfully over long prediction horizons, with at least an order of magnitude fewer parameters than DreamerV3, TD-MPC2, and LeWM. The authors report state-of-the-art control of continuous environments from pixels and accurate system identification across 71 attractors.
Tobias Schmidt, Rodrigo Gonzalez Laiz, Kriszti\'an Szuppinger et al.
In briefTesting whether vision models keep a persistent spatial memory of their surroundings
SummaryThis paper asks whether visual foundation models provide representations that maintain a persistent spatial state as an observer moves through a scene. Using frozen-feature probes on ten representations across streaming egocentric views of static indoor environments, the authors assess four levels: spatial structure, object persistence, predictive spatial update, and actionability. They find that strong readout of the visible scene does not guarantee accurate object persistence or future-feature retrieval.
In briefSymAnt: Long-Horizon Action Prediction Guided by Editable Symbolic Knowledge
SummaryThis paper introduces SymAnt, a neuro-symbolic method that combines neural networks with the structured reasoning of scene and knowledge graphs. From a short video, it jointly searches both graphs to build a symbolic context that guides an iterative process generating future action sequences. Experiments on five datasets (Breakfast, 50 Salads, EPIC-KITCHENS, EGTEA Gaze+, and Assembly101) show its effectiveness especially with short observation windows. Users can edit the underlying knowledge to adjust predictions without further training.
In briefShallowDefend: Using Shallow Safety Alignment to Defend Against Multi-Turn Attacks
SummaryThis paper argues that large language models concentrate their safety behavior in the first few output tokens, which is why harmful requests split across a multi-turn conversation still often succeed. The authors propose Assess→Decide→Respond, a framework where the model assesses user intent, commits to an explicit and parseable safety decision, and then generates a reply conditioned on that decision, with the decision trained by reinforcement learning. Across two model families and fourteen benchmarks, the defenders achieve the strongest multi-turn safety while preserving utility at a modest cost to helpfulness.
WhyThe authors propose Assess→Decide→Respond, training an explicit parseable safety decision via RL with rule-based rewards; the paper reports leading multi-turn safety across two model families and four
In briefEvidence-based significance testing of clinical timeline hypotheses in language models
SummaryThe paper argues that language models often state fluent conclusions about how events unfold over time without any statistical test behind them. The authors introduce STRATA, a benchmark of 8,735 temporal hypotheses over clinical timelines of vaccine adverse-event reports, which grades each step from test plan and evidence set to p-values and verdicts. They also propose ART, a staged paradigm that makes each decision inspectable, and report that most failures occur in selecting relevant records.
WhyProposes STRATA, a benchmark grading each decision in temporal clinical hypothesis testing, and the staged ART paradigm; the authors report experiments on 8,735 hypotheses.
Chengyang He, Wenlong Zhang, Koduvayur Subbalakshmi et al.
In briefAn ODE-based sampler for discrete diffusion models.
SummaryThis paper introduces exponential residual flows (ERFs), a deterministic ODE-based formulation for discrete diffusion models analogous to the probability flow ODE for continuous diffusion. Each token and position gets a continuous residual variable that evolves by piecewise ODEs, and the discrete state changes when a residual reaches zero. The authors prove the construction preserves the marginals of the original diffusion and specialize it to masked and uniform diffusion. The deterministic trajectories support consistency distillation and report gains over prior methods on code and text generation benchmarks.
Samson Gourevitch, Yazid Janati, Eric Moulines et al.
In briefL3G-World: Real-time infinite world generation using a latent 3D Gaussian memory
SummaryThis paper introduces L3G, a latent 3D Gaussian memory that helps interactive world models keep spatial consistency during long exploration. Its memory grows with scene extent rather than stream length, and each Gaussian stores a VAE latent feature so it can be rendered directly into a video diffusion transformer's latent space. The authors report state-of-the-art generation quality and long-horizon consistency, with real-time generation at 9 FPS.
In briefSingle compact model that decides when to offer hints and generates them
SummaryThis paper studies proactive assistants that decide when to give short hints during a conversation and what those hints should say. The authors evaluate a gate and a generator jointly on 28,058 decision points. A single 1B decoder gates from an intermediate layer and generates in the same pass, matching or exceeding two-model pipelines on automatic end-to-end measures with 1.31 to 1.76 times less wall-clock time. Removing the memory profile sharply lowers hint quality, while retrieval and a larger generator add nothing on exact match.
In briefFight World Model: Rule-Based, Customizable Combat Video Generation
SummaryThis paper introduces Fight World Model (FWM), which separates rule-governed combat execution from video rendering. An executable engine tracks game state and resolves fights, and a compiler encodes the outcomes into an aligned interaction proxy that guides rendering. Without retraining the renderer, FWM supports modified rules such as damage scaling, limited-hit immunity, lifesteal, and multi-stage skills, along with character and arena customization. The authors report better action fidelity, hit reactions, health consistency, and appearance preservation than existing game world models.
In briefHairDeck: Generating Animation-Ready Stylized Hair Cards from One Image
SummaryStylized game and avatar hair is built from hair cards animated by spring bones. HairDeck treats each card as a single token and uses a set diffusion transformer to denoise all cards together, generating the cards and their spring-bone chains from one image. On CHARM's test set, it lowers Chamfer distance by 44% and card-orientation error by more than half, while running 33 times faster.
In briefEarly warning for AI agents: probing when a pressured agent leaks secrets
SummaryThe paper studies whether an AI agent can be pressured by a peer to break its operator's rule and reveal a secret, such as a deploy token, in a shared multi-agent setting. The authors train a linear probe on the agent's activations and read it at each model call before output is written. On held-out Qwen3-32B episodes, at a threshold flagging 5% of benign calls, the probe catches 82% of leaking calls, versus 26% for a zero-shot decision-model monitor.
In briefBeyond Ground Truth: An Unsupervised Engine for Checking Dental Segmentation Quality
SummaryThis paper addresses dental X-ray segmentation, where deep learning models can show high confidence even when their predictions are poor. It introduces HUS-QA, a three-stage quality-checking engine built on a SimCLR self-supervised ResNet-18 U-Net trained on the PRAD dataset, which assesses mask quality without ground-truth masks. Standard confidence methods fail in the experiments, while HUS-QA correlates positively with actual Dice scores and achieves a ROC-AUC of 0.8405.
In briefMultistatic integrated sensing and communication with deep learning clutter suppression for 5G and beyond
SummaryThis paper presents a multistatic integrated sensing and communication framework that uses 5G and beyond signals to estimate target position and velocity with distributed receivers. Dense, dynamic clutter is modeled with randomly located scatterers, and conventional signal processing baselines are shown to be limited under strong clutter. The authors introduce residual and attention-enhanced deep neural networks for clutter suppression, which outperform conventional approaches in position and velocity estimation across diverse clutter conditions.
In briefRADIOlogy: A Training-Free, Single-Pass Zero-Shot Anomaly Detection Method
SummaryExisting zero-shot anomaly detection methods often use complex prompt learning or costly sliding-window processing. This paper finds that multi-head attention pooling in vision-language models can localize and detect anomalies zero-shot, and proposes RADIOlogy, a single-pass method that scores images by comparing vision features with normal and abnormal prompts. It refines pixel-level segmentation using RADIO model features. The authors also show that this family of training-free methods can be applied to the medical domain with minimal changes.
In briefVeriCUA: An Agentic-Exploration Framework for Verifying Task Completion by Computer-Using Agents
SummaryThis paper presents VeriCUA, a platform-agnostic framework for checking whether computer-using agents have completed their tasks. It treats verification as dynamic evidence gathering: the verifier first inspects available context to build a task rubric, then selects GUI or code checks for unresolved requirements and combines the findings under explicit evidence rules. Experiments report better agreement with reference labels than compared baselines across online and offline settings, with the largest offline gains on hard cases and stronger rejection of false completion claims online.
In briefMulti-View MoE: Letting Every Router Shape Expert Computation
SummaryThis paper proposes Multi-View MoE, a sparse mixture-of-experts method that uses multiple independently initialized routers to select experts under a fixed active-expert budget. The routers' preferences are also used to mix learned latent views into each selected expert's input. On a 24-layer 10B-A1B model trained on 708B tokens, the full method improves the nine-task downstream mean by 1.18 percentage points over the baseline, and the multi-router component alone yields a 0.6-point gain.
In briefPIVOT: Progressive Intervention via Online Teacher Correction
SummaryThis paper addresses sparse rewards in reinforcement learning from verifiable rewards (RLVR) on hard reasoning tasks, where failed rollouts provide no learning signal. It introduces PIVOT, which inserts teacher-generated corrections into student rollouts and applies token-level KL supervision to those segments alongside the RLVR objective. On hard logic games and math reasoning benchmarks, PIVOT outperforms existing methods, raising mean avg@8 on logic games from 19.5% to 35.6%.
In briefEach Token Gets Its Own Teacher: Multi-Teacher Adaptive On-Policy Self-Distillation
SummaryThis paper proposes MTA-OPSD, a method for post-training large language models that lets each token select its own teacher during training. Prior on-policy self-distillation methods use one fixed teacher or vary reference exposure only across responses. The authors find that the preferred teacher differs token by token within the same response, and identify it through majority voting and KL-based selection. On Qwen3 models for mathematical reasoning, MTA-OPSD outperforms existing approaches, with an average gain of up to +5.9 over the base model.
In briefGym-Anything for Robotics: A framework where AI coding agents build realistic robot simulations.
SummaryThis paper introduces Gym-Anything for Robotics, a framework in which coding agents create realistic robotics simulations to reduce the need for expert effort. The authors select simulations based on U.S. GDP data and build RoboWorld, containing 25 simulations with test and training reinforcement learning environments. Even the best frontier models fail on every test task, averaging a reward of only 0.06. The paper also proposes several methods that substantially improve model performance on RoboWorld, and the authors release their pipeline, environments, and benchmark.
WhyThe authors propose coding agents creating robotics simulations and build RoboWorld; the abstract reports frontier models average only 0.06 reward on test tasks.
In briefTraining Foundation Models with One Algorithm: Unifying Pretraining, Fine-Tuning, and RL into Likelihood Training
SummaryThe paper observes that foundation models are usually trained in separate stages (pretraining, midtraining, supervised fine-tuning, and reinforcement learning) that change both data and algorithm. The authors derive suitable RL objectives as maximum-likelihood fitting to model-generated target distributions, placing all stages under one likelihood-based method called unified training. Conventional staged training is presented as a special case of this framework.
In briefComparing Human and AI Role-Players in Counselling Training: Analysis of Counselling Patterns
SummaryThis study examines how well large language models can simulate counselling clients in educational role-plays compared with human role-players. Novice counsellors' sessions with human peers and with clients simulated by Mixtral 8x7b were recorded and analysed using a sentence-level counselling communication pattern system. The findings show both similarities and differences in counselling pattern use across the two scenarios, suggesting LLM-based role-play may help build counselling skills.
In briefFARMS: A Verifiable Foundation for Automated Computational Science
SummaryThis paper presents FARMS, a framework that turns computational mathematics into formal knowledge AI systems can interpret and reuse with explicit guarantees. It links formal statements to source context, algorithmic meaning and quantitative guarantees through source preparation, reusable representations and semantic review. Of 171 targets in randomized numerical linear algebra and numerical PDEs, 157 were accepted by the pipeline, with mean estimated costs reported per source.
In briefA skill-based agent for advancing clinical genomic analysis
SummaryThe paper presents ClinGeX, a large-language-model agent that composes 105 curated executable skills through task-specific retrieval and dependency-aware orchestration to support clinical genomic analysis. On 223 temporally held-out benchmark cases, it achieved 71.7% macro-averaged task-level concordance with expert classifications, versus 45.7% for the strongest comparator. In a human-AI collaboration study with 45 professionals, ClinGeX assistance raised case-level concordance and reduced reported interpretation time across scenarios.
In briefRamanujan Challenge Results: AI-Assisted Pipeline for Judging Multi-Format Proofs
SummaryThis paper analyzes submissions to the Ramanujan Challenge, a benchmark of research-level number theory problems, and compares the performance of different AI models, computer algebra systems, and formal tools. It also introduces an AI-assisted pipeline that evaluates natural language proofs, CAS code, and Lean formalizations for semantic and logical correctness. The authors report that, compared with one-shot LLM judgment, their pipeline reduces false-positive verdicts on natural language proofs from 23% to 6% and raises judgment accuracy on supporting code from 46% to 58%.
Alan Sorani, Ashvni Narayanan, Michael Shalyt et al.
In briefRetrieval as Traversal: Value-Indexed Memory for Stateful Long-Context Agents
SummaryThe paper proposes treating memory retrieval as a persistent local traversal through an address space, separating each memory's content from its address. Memories are indexed by a cost-regularised objective that weighs the information value a memory provides against the cost of making it locally accessible. As a proof of principle, a simple fixed-geometry traversal head on MiniGrid Memory reached higher mean return in fewer training steps than absolute addressing baselines, though learned geometry and repeated retrieval were not tested.
In briefDetecting Ghanaian Mobile Money Scam SMS with Machine Learning: Comparing Naive Bayes, SVM, and Random Forest
SummaryThis paper presents a supervised machine-learning pipeline for classifying scam and legitimate mobile money (MoMo) SMS messages in Ghana, using a labelled corpus of 220 messages across 18 subcategories. It compares Multinomial Naive Bayes, linear support vector machine, and Random Forest classifiers using accuracy, F1-score, and AUC-ROC. Naive Bayes and linear SVM achieved perfect held-out scores, suggesting interpretable classical classifiers are suitable for filtering scam text in low-resource settings.
In briefFundamental limits of predicting LLM output length from prefill states
SummaryThis paper studies how well a large language model's decoding length can be predicted using only hidden states from the prefill phase. The authors argue that current benchmarks mix predictor quality with task difficulty, and that the stochasticity of autoregressive decoding is an overlooked error source. They derive bounds that prefill-side methods must obey, and show that decode length is approximately normally distributed with an affine variance law, which matches their experiments. They conclude that system error dominates model error and propose a benchmarking protocol that fixes model, data, and sampling parameters.'
WhyThe authors model decoding as a stopped stochastic process, derive per-prompt normality and an affine variance law bounding prediction, and report agreement across predictors and settings.
In briefTRACER: A time-to-event EHR foundation model for detecting diseases across health systems
SummaryThe authors developed TRACER, a decoder-based foundation model pretrained on electronic health records from over 4.5 million patients at four health systems. It predicts 16 incident diseases at 1-, 2-, and 5-year horizons and was evaluated on more than 2 million patients from two health systems excluded from development. Performance was measured at low false-positive rates suited to screening, showing retained discrimination externally, but detection and false-alert burden varied across diseases and health systems.
Mara Pleasure, Ekaterina Redekop, Vedrana Ivezic et al.
In briefHomebrewed SLMs: Small Portuguese Language Models Trained at Home
SummaryThis tutorial shows how to train a language model from scratch, tokenizer included, on a single consumer GPU and evaluate it on specialist-domain tasks. It releases a tested Python package and two Brazilian Portuguese notebooks covering autoregressive and masked diffusion generation, plus a preregistered comparison protocol. On Brazilian prose and federal laws, the autoregressive mode memorises after 64 epochs while masked diffusion keeps improving up to 256. On legal tasks the model ties a 6-gram baseline in likelihood and beats it at detecting tampered text and filling cross-references, but loses at continuation, where larger public models lead.
In briefBURGER: Testing How Robots Generalize Through Language-Guided Drawing
SummaryThis paper presents BURGER, a benchmark where robot arms draw objects or spatial arrangements described in language, using a scalable pipeline that generates drawing trajectories in simulation and the real world. The authors used 28,200 simulated demonstrations across 282 drawing tasks to evaluate five representative vision-language-action (VLA) models on data scale, language grounding, and multimodal behavior. They also deployed a trained policy on a Franka Panda arm that can draw reactively under human intervention.
In briefPhysCount: Do AI Models Count Visible Appearances or Real Physical Objects?
SummaryThis paper introduces PhysCount, a benchmark testing whether multimodal large language models can tell apart what appears in an image from what physically exists in the scene. It includes 546 real-world images and 4,595 human-written questions across 651 target groups. Tests on 40 models show that counting visible appearances is much easier than identifying which appearances belong to the same entity, with the best model answering every question correctly in only 13.36% of groups.
In briefNEOPIC: A neural operator framework for simulating plasma with particle-based methods
SummaryThe paper introduces NEOPIC, a neural operator that computes electric fields directly from particle positions, without using a grid, mesh, or tree structure. Because it does not deposit particles onto a mesh, it avoids aliasing of particle noise and can serve as a plug-and-play field solver with any time integrator. Numerical experiments on electrostatic plasma benchmarks, including Landau damping, two-stream and bump-on-tail instabilities, a cyclotron, and a Penning trap, support the theoretical results.
WhyThe authors propose a neural operator acting directly on particle empirical measures, avoiding grid deposition, and report validation on electrostatic benchmarks including Landau damping and a Penning
Sriramkrishnan Muralikrishnan, Chelsea Maria John, Jeffrey Kelling et al.
In briefCrossWeave: Retrieving 3D Scenes and Objects from 2D Images via Sparse Local Alignment
SummaryThis paper presents CrossWeave, a framework that retrieves corresponding 3D scenes or objects from query 2D images, which is useful for robotics and AR/VR spatial reasoning. It learns from sparse 2D-3D patch-point correspondences rather than whole-scene or instance contrastive objectives. An optimal transport-based aggregator clusters structurally redundant regions to improve descriptors. On ScanNet and 3RScan, a single model achieves state-of-the-art results for both scene- and instance-level retrieval.
Aadith Warrier, Gnana Prakash Punnavajhala, Siddharth Tourani et al.
In briefIndicIPR-QA: An Extractive Question Answering Benchmark for Indian Intellectual Property Law
SummaryThis paper introduces IndicIPR-QA, an extractive question answering benchmark built on Indian intellectual property rights laws. It contains 4,462 question-answer pairs drawn from 1,116 statutory passages across seven Acts, generated with GPT-4o and checked through automatic, LLM-based, and human validation. Experiments show BM25 reaching a token-level F1 of 20.21 while fine-tuned InLegalBERT reaches 85.02, indicating simple word matching is insufficient for these questions. Performance also differs substantially across the seven Acts.
In briefScaling Multilingual Safety Training for LLMs Using Templates
SummaryThis paper notes that safety alignment for large language models focuses mainly on English, which hurts performance in low-resource languages. The authors propose a template-based augmentation pipeline that extracts safe templates from harmful prompts, translates them, and fills them with culturally authentic entities to create multilingual refusal and benign corpora. They release MSafeTemp, a validated corpus covering 60 languages, and show that English-only evaluation can misjudge multilingual safety while template-augmented training improves safety beyond English.
In briefA low-rank subspace captures how LLM representations align with human brain activity during listening
SummaryThe paper uses ridge reduced-rank regression to identify the subspace of large language model representations that linearly predicts neural responses. Across three naturalistic language-listening datasets (intracranial electrophysiology and fMRI) and six model conditions, full linear encoding performance is recovered using only 1-3% of the embedding dimensionality. These subspaces generalize across held-out participants and differ from the principal axes of variation in the embeddings.
In briefFinding Interpretable Acoustic Features in Autoregressive Voice-Cloning TTS Models
SummaryThis paper proposes a feature-based interpretability framework for autoregressive voice-cloning text-to-speech (TTS) models, using sparse autoencoders, PCA, and random directions. The analysis relies on sparse probing, steering, and quantitative auto-interpretability, focusing on Qwen3-TTS and extending probing and steering to five other models. The models encode concepts such as pitch, energy, noise, and accent, but sparse prediction and steering work well only for some of them. The authors also report that sparse autoencoders suit probing human-labelled concepts and discovering new ones, while PCA performs competitively for dense probing and steering, and that auto-interpretability identified 684 validated features covering speaker clusters, speaking style, and recording conditions.
In briefAlgebraic World Model: AI that builds mathematical models of unknown environments and proves what is reachable or impossible
SummaryThe paper introduces the Algebraic World Model (AWM), a framework that explores unfamiliar environments, builds executable Python hypotheses, and formalizes the resulting model in Lean 4. AWM produces machine-checked proofs that certain goals are unreachable, including in two impossible game levels where baseline coding-tool models produced none. Across the 25 public ARC-AGI-3 games, the authors report that the models let Lean replay the winning trajectories of all 183 levels and predict unseen transitions.
In briefNumerical integration over bounded domains using learned reflected diffusions
SummaryThis paper studies high-dimensional integration of a non-negative function over bounded convex domains, estimating the integral while keeping samples inside the domain. The authors build an importance-sampling proposal from reflected diffusion dynamics and train a neural score model to reverse a process that moves the target toward the uniform distribution on the domain. They derive a training objective without extra boundary terms, an unbiased integral estimator, and a REINFORCE-based gradient estimator for non-smooth integrands.
Kirill Korolev, Artur Goldman, Timofei Gritsaev et al.
In briefLucidLyra: Feed-forward 3D scene generation with built-in semantic understanding
SummaryLucidLyra is a framework that generates 3D scenes from video latents in a single feed-forward pass, making them directly queryable for semantics and object instances. It distills knowledge from a camera-controlled video diffusion model and frozen 2D vision foundation models into one model that jointly predicts 3D Gaussians encoding geometry, appearance, open-vocabulary semantics, and class-agnostic instance cues. The method needs no per-scene optimization or 3D annotations and extends to monocular video for dynamic scenes.
Oliver Hahn, Rodrigo Marcuzzi, Sangeetha Grama Srinivasan et al.
In briefAI in EGI's e-infrastructure: past experience and future strategy
SummaryThe EGI Federation has provided federated e-infrastructure services to science for over 15 years. This paper reviews EGI's early AI work in the interTwin and iMagine Horizon Europe projects, and notes that since 2023 the focus of AI in science has shifted from model training to adapting it to specific domains and questions. It then presents an AI strategy within the EGI Federation Strategy 2026-2030, built on three enablers: AI-ready infrastructure, human skills, and trust.
Gergely Sipos, Sergio Andreozzi, Marco Rorro et al.
In briefAgentic University: Collaborative Agents for Personalized Learning
SummaryThe paper introduces Human-Participatory Multi-Agent Organization and its implementation, Agentic University (AU), where learners work with specialized agents as organization members. Together they maintain a revisable Capability Program that links personal goals, multimodal learning activities, and layered assessment, adjusting support based on learner feedback and performance. Evaluation reports larger mean assessment gains (12.19 vs. 5.61 points on a 50-point scale) and stronger cognitive support for users.
In briefrTaste: Do Research Agents Have Taste? Judging Which Early Experiments to Run
SummaryThis paper tests whether AI research agents can judge which experiments are worth running before their results are known. Six frontier models, working alone or in teams, proposed minimum viable experiments for seven recent NLP papers, forecast their outcomes and ranked them, and a separate executor ran all of them for review by three model auditors. Of 415 proposed experiments, 194 were judged sound, valuable and new, but only 73 were also judged beyond the seed paper and non-obvious. The agents' rankings showed limited improvement, and their forecasts were not measurably more accurate than a fresh instance of the same model.
In briefMetaRefine: Refining Image Quality Scores Using Multi-Level Features and Original Image Geometry
SummaryThis paper introduces MetaRefine, a score refinement architecture that uses lightweight residual heads to correct scores from a frozen image assessment model. The corrections are predicted from original image geometry, multi-level visual features, and the model's original score, without changing the original scoring pipeline. Across 100 configurations on 21 datasets, it improves PLCC and SRCC in most configurations and reduces MAE and RMSE, with a median parameter overhead of only 1.11%, and it can further improve state-of-the-art models.
In briefThe Precision Wall: A Boundary Where Optimizers Stall on Power-Law Measurement Spectra
SummaryThis paper studies low-rank matrix sensing, where replacing isotropic Gaussian measurements with power-law covariances causes division-normalised first-order optimizers to stall above a critical exponent. The authors attribute the failure to a numerical precision limit: the smallest non-gauge expected Hessian eigenvalue decays with dimension, and once it falls below the optimizer's denominator floor, normalisation noise overwhelms rare-direction gradients. They give a leading-order boundary formula that predicts Adam's onset within 0.45 across a held-out grid of 16 configurations.
In briefKeeping LLMs Useful While Repeatedly Removing Training Data via Orthogonal Projection
SummaryThis paper introduces Continual Orthogonal Projection (COP), a method for removing the influence of sensitive training data from large language models through repeated unlearning requests without retraining from scratch. Each update is restricted to the orthogonal complement of activation subspaces cached from earlier forget sets, storing only a low-rank basis per layer. Experiments on the TOFU and RWKU benchmarks show that COP preserves utility across sequential requests and forgets more effectively than continual-unlearning baselines.
Juan Belieni, Ana Carolina Erthal, Eliezer de Souza da Silva et al.
In briefFIRME: A Memory Graph Anchored to a Registry for Evolving Investor Knowledge
SummaryThe paper introduces FIRME, a registry-anchored memory graph that links observations from investor calls, cash-flow records, and S&P 500 earnings-call transcripts to persistent canonical relationships while keeping source, time, and context. Persistent aliases direct recurring observations to the same relationship history across periods. Compared with a baseline that lacks alias memory, FIRME raises final-period Correct Match Rate by 7.8 to 12.1 percentage points across four dataset-backend settings.
In briefAVPS-360: A spatial audio-visual benchmark for segmenting and tracking sound-producing people
SummaryThis paper introduces AVPS-360, a benchmark for segmenting and tracking people who produce sound in 360-degree videos, using first-order Ambisonics spatial audio. Built on STARSS23, it contains 230 recordings totaling 10.5 hours, with 31,678 manually reviewed instance masks linked to sound-event labels. The authors also propose a baseline that combines spatial audio and visual features, which they report substantially outperforms existing methods on long recordings with intermittent sound.
Yifu Sun, Parthasaarathy Sudarsanam, Tuomas Virtanen et al.
In briefMRI-free PET quantification using CT-to-MRI synthesis with reliability estimates
SummaryThis paper addresses amyloid PET SUVR quantification, which normally requires a subject's own T1-weighted MRI that is not always available. The authors synthesize MRI from CT using a diffusion model and apply conformal prediction to estimate a plausible range for the real MRI-based SUVR. On OASIS-3 11C-PiB PET, synthetic-MRI cortical SUVR closely matched real-MRI results and gave the same amyloid-positivity call in all 87 test subjects. Zero-shot testing on an external cohort worked without retraining, though conformal calibration needed recalibration on the target cohort.
In briefTraining AI chatbots in simulated settings to conduct psychiatric intake interviews
SummaryThis paper trains language models with multi-turn reinforcement learning in simulated environments to improve structured text-based psychiatric intake, covering history-taking, differential diagnosis, and initial management planning. In a randomized, double-blind crossover study, 32 psychiatrists role-played standardized patients and 15 separate psychiatrists scored the transcripts. The trained agent outperformed its base model in 76.7% of matched scenario pairs and more than halved under-triage of acute safety risks.
Samuel Schmidgall, Tim Strother, Alex Bijamov et al.
In briefRural Readiness: Domain generalization when the hard domains are chosen adversarially
SummaryThe paper studies how a model trained on several source domains can perform on any unseen target domain, aiming to match the worst hypothesis consistent with the source data. It characterizes the sample complexity of this goal up to logarithmic factors. Relaxing the goal to compete with hypotheses that do well on 99% rather than 100% of source domains reduces the sample complexity from Ω(dk) to O(d), where d is the VC dimension and k the number of source domains.
In briefInfantNet: A large dataset for estimating infant body pose and shape
SummaryThis paper introduces InfantNet, the largest real-image dataset of infants to date, with 108,902 RGB images of infants aged 6 to 18 months. Every image has 2D keypoint annotations, and 11,642 images also include 3D pose and shape labels with full SMIL parameters. Baseline experiments show that adult pose estimators generalize poorly to infants, but fine-tuning on InfantNet gives consistent improvements.
In briefA shared benchmark for aerial-enhanced self-driving: the KITScenes Aerial dataset
SummaryThis paper introduces KITScenes Aerial, an extension of the KITScenes Multimodal dataset that adds synchronized aerial imagery and higher-resolution neural representations. It provides a common basis for map construction, aerial-assisted online HD mapping, and cross-view localization. Mapping is evaluated across all 53 map classes, showing that aerial context especially helps fine-grained, extended-range mapping and that higher resolution improves results further. Under geographically disjoint conditions, cross-view localization achieves a mean translation error of 5.13 m.
Daniel Lengerer, Richard Schwarzkopf, Jonas Merkert et al.
In briefGlance: A Generator That Also Ranks Search Results
SummaryRetrieval-augmented generation usually runs separate models for reranking and answering, duplicating model memory and passage processing. Glance adds a lightweight decision adapter and pointer head to the generator, scoring candidate passages from one shared prefill without encoding each query-passage pair separately. Using a single 9B backbone fine-tuned on MS MARCO for about 39 minutes, it outperforms bge-reranker-v2-m3 on SciFact and NFCorpus in nDCG@10.
In briefPhysInsert: Physically Plausible Object Insertion with Geometric Priors
SummaryPhysInsert is a framework for inserting objects into images based on a reference, adjusting both the object and its surrounding scene so the result looks physically plausible. The authors build the Insert World dataset of strictly aligned Blender-rendered counterfactual pairs across viewpoints and lighting, and add geometry-guided adaptation with 2D object-scene spatial supervision. They also introduce the PhysInsert-Bench benchmark, and experiments on it and DreamBooth show improved image fidelity, while a user study reports better placement, lighting integration, perspective, and pose plausibility.
In briefA review of reinforcement learning for controlling partial differential equations
SummaryThis review connects reinforcement learning with PDE-constrained optimization, surveying model-free, model-based, and multi-agent approaches to controlling systems governed by partial differential equations. It emphasizes sample efficiency, generalization across parametric variations, and scalability to high-dimensional problems. Two examples, stabilizing a chaotic system and a flow control problem, illustrate the performance and pitfalls of RL agents, and the paper discusses evaluation practices, current limitations, and open research directions.
Nicol\`o Botteghi, Matteo Tomasetto, Andrea Manzoni
In briefLearning cell development dynamics using family tree information
SummaryThis paper introduces lineage inpainting, a generative model that uses information about a cell's relatives, such as sisters and cousins, to predict how cell states change across lineage trees. It combines tree-structured transformers with stochastic interpolants. Applied to a lineage-traced transcriptional atlas of mouse embryogenesis, the model recovers dynamics that gene expression alone does not reveal, including cases where distinct lineages converge to a common state. Trained only on unperturbed development, it can also simulate counterfactuals to generate hypotheses about perturbation responses, and in many cases these agree with observed phenotypes, including a Tbx5 knockout measurement.
Gokul Gowri, Nic Fishman, William N. Colgan et al.
In briefPrefix Executability: Judging Tool-Using AI Agents by Each Step, Not Just the End Result
SummaryThe paper argues that tool-use benchmarks scoring only final task success can hide whether an agent's intermediate tool-call sequence was actually executable. It proposes a prefix executability protocol that checks whether every prefix of a predicted tool-call sequence stays valid under tool schemas, state preconditions, and data dependencies, summarized with several reliability metrics. Experiments on BFCL-v3, ToolSandbox, and τ³-bench show that rankings based on final success diverge from prefix-level rankings.
Amir H. Rezaeian, Weiyi Sun, Yassine Benajiba et al.
In briefFramework-constrained generative model for discovering new inorganic crystal materials
SummaryThis paper introduces OMatG-FC, a generative method for de novo inorganic crystal materials that is constrained by known structural frameworks, meaning unit cell shapes and particle positions. It uses discrete flow matching for atomic species and stochastic interpolants for unit cell volume scaling, which reduces the problem's dimensionality. The authors also analyze duplicate frameworks in the MP-20 dataset and report state-of-the-art stable, unique, and novel rates on LeMat-GenBench.
Maya M. Martirossyan, Hillary Pan, Philipp H\"ollmer et al.
In briefA pretrained model that filters noisy signals across different dynamical systems
SummaryThe paper introduces FilterFM, the first pretrained model for Bayesian filtering, which estimates latent states across heterogeneous dynamical systems using a single checkpoint. It uses a hierarchical two-level observation sequence encoder and a prior-to-posterior state-reconstruction module, and models uncertainty via quantile prediction. Reported results show gains in in-domain accuracy, runtime, and noise robustness over baselines, along with the first zero-shot filtering on unseen dynamics.
In briefSoil moisture estimation for farming using knowledge distillation from simpler single-input models
SummaryThis paper proposes Soil Moisture Distillation (SMD), a knowledge distillation framework that uses unimodal teacher models to improve multimodal soil moisture estimation. It also finds that crop images provide better visual cues than soil patches. SMD improves MAPE by 2% for soil patches and 1.2% for crops over state-of-the-art baselines.
Mohammed Rakib, Vrinda Nagula, Arunkumar Bagavathi
In briefAdaptive Search with Path-Constrained Mobile Sensors Using Multi-Armed Bandits: A method for energy-limited sensor teams to choose routes adaptively to find regions of interest.
SummaryThe paper studies how a team of energy-limited mobile sensors can detect regions of interest from noisy online measurements. Sensors move only along precomputed feasible paths that respect mobility limits, require returning to charging stations, and avoid known obstacles. The authors model path selection as a multi-armed bandit problem and propose an iterative search algorithm based on confidence bounds, with finite-time and anytime probabilistic correctness guarantees and non-asymptotic bounds on detection and labeling times.
Abraham P. Vinod, Parth K. Thaker, Gautam Dasarathy et al.
In briefUncertainty-aware feedback loop diagnosis combining reinforcement learning and conformal prediction for wafer defect classification
SummaryThis paper proposes UFLD, an algorithm for detecting defects in wafer maps. It uses reinforcement learning to adaptively choose measurements and reports a prediction uncertainty level after each step, stopping once enough information is obtained. The method uses conformal prediction to set the stopping time, without needing a separate learned confidence network. On the MixedWM38 dataset, UFLD reduces acquisition time by about 40% and reaches 98.4% inspection accuracy.
In briefNECA: Borrowing Event Timing from Similar ICU Patients to Predict Organ Dysfunction
SummaryThis paper proposes NECA, a model-agnostic method that finds training patients with similar early ICU records and converts their later organ dysfunction event onset times into cumulative event-curve features over 6, 12, 24, and 42 hours. These features are appended to each patient's representation and fed to standard predictors. On PhysioNet 2012 and a 4,000-stay MIMIC-IV subset, NECA improves average PR-AUC and ROC-AUC across four organ dysfunction event tasks, with the largest gain for neurological dysfunction.
Chenyang Wang, Jiawei Theresa Tang, Xinran Liu et al.
In briefIMPACT: Inverter-Informed Multimodal Learning for Photovoltaic Power Forecasting
SummaryThis paper presents a physics-informed multimodal deep learning framework for probabilistic photovoltaic power forecasts 10 to 110 minutes ahead. It combines plant-level temporal history, inverter-level spatial cloud dynamics, solar geometry, and current weather through gated cross-attention, with online conformal calibration of quantile outputs. The model outperforms all 14 baselines in forecast skill score, and retains useful skill under overcast and variable sky conditions.
In briefOCCAMS: A Multi-Task Learning Study of CMB Power Spectra
SummaryThis paper uses multi-task learning with variational autoencoders, without any built-in cosmological model, to fit the cosmic microwave background temperature and polarisation power spectra with a small number of latent variables. Compared with Planck 2018 data, six latent variables give the best or near-best fit, matching the number of free parameters in the standard model-dependent approach. The latent variables correspond partly to standard cosmological parameters, suggesting that deep learning can provide a compact, data-driven parametrisation of CMB spectra.
In briefLinking hidden neurons in audio classifiers to human-understandable acoustic concepts.
SummaryThis paper examines whether hidden neurons in deep audio classifiers can be linked to human-understandable symbolic acoustic concepts. It presents a model-agnostic neurosymbolic framework that uses symbolic concept induction to analyze the final 32-neuron hidden layer. The framework is tested on music genre classification with GTZAN and musical instrument classification with IRMAS. Several neurons were associated with meaningful concepts, and most confirmed pairs were statistically supported by the Mann-Whitney U test.
In briefRetrieving from Your Own Mistakes: On-Policy Training of Embedding Models for Multi-Hop Retrieval-Augmented Generation
SummaryThis paper addresses multi-hop retrieval-augmented generation, where answers need evidence spread across several documents. It finds that chain embedding retrievers trained only on gold histories perform strongly, but their mistakes at test time reduce their ability to find missing evidence. The authors propose on-policy state training, which trains the encoder on the states it reaches itself, and report that this retrieves complete evidence more often and transfers to harder long-chain variants.
In briefSCARL: Sequence Confidence-Aligned Reinforcement Learning
SummaryThis paper proposes SCARL, a reinforcement learning method for language-model reasoning that adjusts each sample's advantage based on the gap between its group-relative confidence and the outcome its advantage implies, without changing the advantage's sign. The authors report that on Qwen3-4B-Base it raises average Pass@32 over DAPO by 2.4 points. The method is also reported to improve 14 of 16 settings across four estimators and four backbones while preserving policy entropy.
In briefContinuous-Action World Models for Transferring Robot Dynamics Across Embodiments
SummaryThis paper introduces Continuous-Action World Models (CoAWM), a framework that learns spatiotemporal dynamics priors usable across different robot embodiments and control frequencies. It uses a shared latent function space, first pretrained on trajectories from heterogeneous robots and sampling rates, then connected to a world-model backbone that conditions on observations, language, and robot state. A single decoder returns end-effector actions at any queried time.
In briefScaling math proof training data using structured textbooks
SummaryThis paper introduces TOME-Math, a corpus converting 2,118 university mathematics textbooks into about 380K structured proof-supervision records that keep chapter order, proposition-proof links, dependency references, and source provenance. The authors use these structural signals to explore curriculum learning across mathematical subjects and to study scaling with supervision volume and source breadth. At full scale, the paper reports that fine-tuning on TOME-Math raises the unweighted mean score by 16.7 percentage points over the base model and exceeds three public instruction-data baselines in aggregate.
In briefCombining and guiding several diffusion models with a generalized Feynman-Kac correction method
SummaryThis paper introduces generalized Feynman-Kac correctors (GFKC), a framework for sampling from an arbitrary target function defined over several pretrained diffusion models. The authors show that geometric interpolation between an initial function and the target satisfies a Feynman-Kac PDE and derive its potential in closed form. The method runs in the log domain, supports reward-based steering, and corrects generation on the fly without retraining. It is validated on analytic Gaussian models and on diffusion models trained on MNIST.
In briefTesting AI language models on randomly chosen unsolved math conjectures
SummaryThe authors introduce OpenConjecture, a dataset of conjectures from arXiv mathematics papers published since February 2026. Frontier GPT models were applied to 200 randomly sampled conjectures over three passes, producing candidate solutions for roughly 60% of them, with counterexamples about 1.3 times more common than proofs. The paper also examines solution novelty, literature citation, and whether resolutions rely on errors or underspecification in the conjecture.
In briefLLM agent reasoning for predicting pediatric diabetes outcomes from multimodal medical records
SummaryThis paper tests large language model agents for clinical prediction in children, using records from 2,489 youths with type 2 diabetes and 4,018 with type 1 diabetes. Data included structured records, social determinants from clinical notes, and glucose monitoring data for type 1 patients. Using 18 months of follow-up, the authors predicted seven type 2 diabetes outcomes at 24 months. The evidence-citing CEDAR pipeline outperformed trained machine learning models in both cohorts, and two pediatric endocrinologists reached the same prediction as CEDAR in about nine of ten cases.
In briefHiFOR: Hierarchical Foresight for Robot Vision-Language-Action Models
SummaryThis paper proposes HiFOR, a framework for vision-language-action robot models that predicts future states at multiple time scales. A Hierarchical Dependency Mask jointly predicts short-term state changes, semantic milestones, and final goal states. The method organizes future latent representations along time, view, and spatial dimensions, and retrieves relevant future information guided by evolving action representations during denoising. Experiments were conducted on two simulation benchmarks and six real-world robotic manipulation tasks.
In briefThe Athlete as a Scientist of One: Agentic Bayesian Experimental Design for Individualized Physiology from Consumer Wearables
SummaryThis paper presents EpiCoach, an agentic system that studies one cross-country mountain biker as an auditable N-of-1 experiment. Claims about the athlete are kept in an evidence-gated registry, each requiring a probabilistic prediction before a session and updated only through logged statistical fits. Predictions come from a digital twin built from heart rate, GPS and barometric altitude, and the agent designs each next training week as a constrained experiment. Evaluated on eighteen months of data via replay benchmarks and a live prospective phase, the authors release the registry, benchmark and protocol.
In briefCurriculum or Easy Start? A Benchmark of Unsupervised Environment Design for Quadrupeds
SummaryThis paper proposes a benchmark that compares 14 unsupervised environment design, domain randomisation and robotics curriculum methods under one protocol on five 3D quadrupeds in MuJoCo MJX. Terrains are described by eight interpretable parameters, and all curricula train the same PPO student with matched environment-step budgets. No method ranks in the top group on every robot and level set, and methods with easier training levels often lead on in-distribution random terrains. On three robots the leader on random levels differs from the leader on held-out terrain.
In briefYour LR Schedule May Not Be Principled: Effective Learning Rate Scheduling in Weight-Normalized Training
SummaryThis paper argues that the pretraining loss gains of weight-normalized training mainly come from its effective learning rate trajectory rather than from normalization alone. It proposes AIR (Affine Inverse-square-Root decay), a schedule combining a 1/√t decay with an affine adjustment to meet set peak and terminal learning rates. In 128B-token Mixture-of-Experts pretraining, AIR reduces terminal training loss by about 0.01 relative to cosine, and applying the same effective learning rate trajectory to weight decay reproduces similar training dynamics.
In briefDon't Lose the Thread: Making LLM Agents Manage Their Own Memory on Long Tasks
SummaryThis paper addresses LLM agents' struggles on long-horizon tasks, where cluttered working memory dilutes attention and hinders planning. It proposes CORAL, a paradigm in which agents checkpoint task progress and verified facts, then reset cluttered context and resume from a clean state, trained with a multi-episode reinforcement learning algorithm. On the GAIA benchmark, an 8B model with CORAL reaches 42.8% average accuracy, using 6-7 times fewer input tokens than a ReAct agent.
In briefA unified Hessian-guided method for post-training quantization of language models
SummaryThis paper proposes a unified framework for post-training quantization of large language models. Using a tractable Fisher approximation, it derives a Hessian-guided estimate of the output change caused by joint weight and activation quantization. The method jointly optimizes quantization ranges, linear transformations, and weight rounding under a shared output-sensitivity objective. The authors report that it outperforms state-of-the-art transformation and rounding methods in both channel-wise and block-wise settings.
Mads Toftrup, Marco Federici, Boris van Breugel et al.
In briefWizard: Real-time, agent-controlled manipulation of any virtual world
SummaryThis paper presents Wizard, an agentic game world model that enables real-time, precise control of character motions and generalizes zero-shot to new scenes, characters, and actions. The authors build a dataset covering 750 scenes and 178 characters, using skeleton and geometry information as structured representations to condition a video generation model. An adaptive masking mechanism improves alignment between these conditioning signals and generated video, while an agentic framework explicitly maintains and updates game states so players can freely configure their virtual worlds.
In briefRFW-A: Automatically generating 400K articulated 3D objects across 500+ categories using AI agents
SummaryThis paper presents RFW-A, a method where AI agents build reusable category-level generators that then produce articulated objects through procedural execution without further model calls. The authors use it to create RFW-A-400K, a dataset of over 400K simulation-ready articulated assets spanning more than 500 categories. Experiments report high structural and kinematic validity, and training on this data improves part-level articulation estimation.
WhyThe authors combine agentic program synthesis with procedural generation, reporting 400K articulated assets over 500 categories and zero-shot policy transfer to real objects.
In briefExplaining multistep, multivariate wind power forecasts with GradientSHAP
SummaryThis paper adapts the GradientSHAP method to explain multivariate, multistep wind power forecasting models, an area that has been largely unexplored. The authors create new visualizations that give both local and global explanations, showing relationships among environmental factors, site conditions, and temporal patterns. They find that with careful handling of SHAP values and domain expertise, the method yields useful insights into model behaviour. The authors also note that its complexity makes practical deployment difficult.
In briefHow pretrained language models jointly encode two independently changing temporal variables
SummaryThe paper studies how language models keep track of several contextual variables that change independently within one document. Using controlled documents with separate language-switch and topic-change timers, the authors find that both timers can be linearly decoded from the same hidden states across ten pretrained checkpoints. The results support an approximate factorization of these representations in overlapping subspaces.
In briefRadiology guideline-based decision graph for clinical decision support
SummaryThis paper introduces GDGraph, which uses a large language model to split radiology guidelines into sections and build an executable decision graph from extracted subgraphs. At inference, a fast model answers localized clinical questions while deterministic traversal resolves Boolean logic, producing a traceable decision path. On 347 cases covering renal cyst, adnexal mass, and thyroid nodule guidelines, accuracy rose from 57.3% to 90.8%.
In briefHow to Loop a Transformer: Scaling, Specialization, and Parallel Inference
SummaryThis paper studies looping in transformer models, which reuses the same weights by applying a transformer, or part of it, repeatedly to add depth without adding parameters. The authors train dense decoder-only models up to 1B parameters, varying which parts are looped, the number of loops, and how information passes between loops. They find that looping the full model is strongest at a fixed parameter budget, while looping a middle block is better under limited training compute, and that gains from more loops can saturate or become unstable. They also add loop-specific low-rank adapters and modify Parallel Loop Transformers so that quality is preserved while decoding runs in parallel.
Chanakya Ekbote, Artidoro Pagnoni, Devendra Singh Sachan et al.
In briefAccuracy and calibration of Decision and System One AI models on rheumatology exam and diagnosis questions
SummaryThis study evaluates how accurately several Decision and System One models answer rheumatology examination questions and diagnostic case vignettes, and how well their confidence matches their correctness. On Spanish specialty-training exam questions, most models scored highly, with Jev at 93.0% in Spanish and 91.9% in English, below GPT-6 Sol at 99.4%. On diagnostic vignettes, all System One models scored below GPT-6 Sol at 76.9%, and Jev reached 50.0% while showing overconfidence.
Alfredo Madrid-Garc\'ia, Diego Benavent, Fabian Lechner et al.
In briefVampire-Hammer: calling an automated prover from the Lean proof assistant
SummaryThe paper presents Vampire-Hammer, a tool that lets the Lean interactive theorem prover hand goals to the Vampire automated theorem prover and then checks the resulting proofs with the Lean kernel. It processes goals and premises with the Lean-auto framework and uses standard TPTP and annotated TSTP formats to support proof reconstruction. On 196,083 eligible Mathlib theorems with a 10-second budget, it solves 21,158 problems that the other evaluated tools did not solve within the same time.
In briefTextual Response-Aware Guidance for Zero-Shot Adversarial Robustness in Vision-Language Models
SummaryThis paper studies how to make vision-language models more resistant to adversarial attacks while keeping their zero-shot performance. It argues that two common guidance methods each have limitations, and proposes TRAG, which constrains visual representation drift using the pre-trained text encoder's responses to local input changes. Experiments across architectures and tasks report state-of-the-art clean and robust performance, and further analysis indicates TRAG helps preserve pre-trained image-text relationships.
In briefRimuru: Unified Memory Management for LLM Serving via Elastic Expert Offloading
SummaryThis paper addresses device memory pressure in serving large language models, where fixed memory partitions for model weights and attention states cannot adapt to changing workloads. It presents Rimuru, which uses a shared page pool, memory views with per-tensor page sizes, a multi-level page allocator, and a page-table-aware fused MoE kernel to share memory elastically across heterogeneous tensors. A workload-aware adjuster reallocates pages at runtime. Implemented on vLLM, the system is reported to achieve up to 19.03x baseline throughput and a 98.1% average reduction in TTFT.
In briefDeep learning with causal constraints for early warning of rapid cyclone intensification in the western North Pacific
SummaryThis paper presents CausalVisionRI, a deep-learning system that warns of rapid intensification of tropical cyclones over 24 hours in the western North Pacific. It combines infrared satellite imagery, atmospheric and oceanic fields, and a causal graph discovered from observations as a training prior. On a forward holdout test covering 2020 to 2024, its five-member ensemble achieves a Brier skill score of 0.259, outperforming several reference systems. Ablation experiments show that the causal graph pathway improves performance.
