In briefA drift-aware agent that decides when to retrain machine learning models.
SummaryThe paper presents DARE (Drift-Aware Retraining Engine), an offline, event-driven framework that separates ingestion, monitoring, policy decisions, training, and logging with deterministic replay. On the Airline dataset, the locked LinUCB policy achieves higher mean F1 than heuristic and action-rate-matched baselines while substantially reducing intervention frequency. Runtime-calibrated evaluation shows the gains persist when measured by cumulative update wall-clock time.
In briefHow Much Text Remains Affects Partial-Prompt Contamination Test Errors
SummaryThis paper audits the partial-prompt completion test for benchmark contamination, which gives a model the first part of a problem and checks whether it reproduces the rest. Seven open-weight models were tested on MATH-500 and recent AIME 2025/2026 problems used as true negatives. Problems with fewer leftover words are flagged far more often, and short leftovers account for much of the ROUGE-based flags. The authors recommend length-stratified reporting, prefixes of at most 60%, and cross-family calibration.
In briefA plug-and-play verifier that separates process rewards to stabilize long-horizon AI reasoning.
SummaryThe paper proposes a plug-and-play verifier framework that stabilizes long-horizon reasoning by explicitly checking intermediate reasoning artifacts produced by agents. Verification scores are aggregated into bounded process rewards and combined with sparse outcome rewards through a two-stage post-training schedule. On HotpotQA with Qwen3.5-4B, the method reaches 60.18% accuracy, above the standard GRPO baseline at 57.73%.
In briefLongtail-SWE: A benchmark of 202 repository issue tasks for 13 underrepresented languages
SummaryThis paper introduces Longtail-SWE, a benchmark of 202 issue-resolution tasks across 13 programming languages that have limited coverage in existing repository-level benchmarks, such as Clojure, Dart, and Haskell. Tasks are built with a solver, judge, and repairer pipeline that revises tests and task descriptions to fix tests that reject valid solutions or accept incorrect ones. The authors evaluate nine model configurations and release the instances, environments, tooling, and agent trajectories.
Vladimir Dobrovolskii, Ibragim Badertdinov, Alexander Golubev
In briefUncertainty, Not Concealment: Re-examining Sycophancy in Vision-Language Models
SummaryThis paper examines whether vision-language models used in medical retrieval-augmented generation can detect when a retrieved report contradicts a chest X-ray. The authors use three diagnostics, representation probing, activation steering, and forced two-way choice, to separate concealment from uncertainty. They find the models suppress contradiction in behavior, but the internal conflict signal looks like hedged uncertainty rather than hidden knowledge, and they recommend re-grounding the models in the image and surfacing uncertainty.
Mohamed I. Eldagla, Zeyad H. ElShafei, Ahmed M. Anter
In briefSparse Predict and Demote: a scheduling policy that needs very little job-size advice
SummaryThis paper introduces Sparse Predict and Demote (SPD), a single-server scheduling policy that aims to minimize mean response time. It requires only one-bit job-size predictions for a vanishing fraction of arrivals. The authors prove that SPD achieves order-optimal mean response time in heavy traffic, even when predictions have any fixed multiplicative error and computing them occupies the server. Numerical experiments show that SPD also performs near-optimally outside the heavy-traffic regime.
WhyThe authors propose SPD, achieving order-optimal heavy-traffic mean response time using one-bit size predictions for a vanishing fraction of arrivals.
Chengyi Liu, Thodoris Lykouris, Junxuan Shen et al.
In briefContinual instruction tuning of LLMs for managing cloud-edge continuum resources
SummaryThe paper proposes using a large language model as an intent compiler that runs once per management intent or declared context change, leaving low-latency runtime actions to a fast, deterministic controller. The authors frame intent compilation as continual instruction tuning and measure catastrophic forgetting through the executed objective of the underlying controller. In preliminary experiments on a continuum simulator with a 2B backbone and parameter-efficient adapters, naive sequential tuning forgets earlier intents, while replaying stored examples nearly prevents this. The tuned compact compiler performs on par with simpler translators and is more robust than a grammar parser to rephrased intents.
Lanpei Li, Antonino Vaccarella, Vincenzo Lomonaco et al.
In briefClinical Testing of Conformal Prediction for Chest X-ray Report Writing
SummaryMultimodal large language models can write radiology reports but often produce hallucinations, so clinically meaningful risk control is needed. This paper introduces CLARITY, a set of clinical admissibility criteria that separates clinically consequential errors from minor deviations, and collects radiologist annotations of model-generated reports using it. The study finds that commonly used automatic admissibility measures systematically fail to detect clinically relevant errors. Automated judgments are largely decoupled from true clinical admissibility.
Nikolaos Kontogeorgis, Pascal Krezer, Klaus-Rudolf Kladny et al.
In briefFairness-controlled decoding for language model recommenders
SummaryThe paper proposes FairRec, an item-level decoding framework that controls candidate selection while recommendation lists are generated, to reduce differences in recommended content when a protected attribute is changed. A soft variant balances candidate likelihood against semantic disparity and adapts a penalty based on observed violations. Experiments on MovieLens-1M and Amazon Movies & TV with Qwen3-8B and Llama3.1-8B-Instruct show lower counterfactual disparity than several baselines while keeping competitive HR@10 and NDCG@10.
In briefLand use planning around transit stations using deep reinforcement learning
SummaryThis paper introduces DRL4TOD, a deep reinforcement learning framework that automatically optimizes land-use types and floor area ratios around transit stations to follow Transit-Oriented Development principles. It uses a graph neural network to model spatial relationships among parcels, roads, and metro stations, and jointly optimizes land-use density, diversity, and transit capacity utilization while minimizing reconstruction costs. Applied to two station areas in Shanghai, it outperformed rule-based methods and genetic algorithms with lower reconstruction costs.
In briefRestoring refusal safety in fine-tuned LLMs without retraining
SummaryThe paper observes that safety-aligned language models lose refusal behavior after domain-specific fine-tuning, even when the fine-tuning data is benign. It proposes SPECTRA, a training-free weight-level method that transfers the base model's refusal direction into the fine-tuned model via an activation-aware low-rank update, requiring no gradient optimization or domain safety data. Experiments show substantially improved refusal behavior while preserving general coherence and domain knowledge, with negligible change in false refusal rates.
In briefStress-testing LLMs that summarize radiology reports for patients
SummaryThis paper presents an adversarial framework for testing how reliably large language models summarize radiology reports in patient-friendly language. The authors create altered CT and MRI reports by substituting terms and removing impressions, then have six models summarize them, producing 18,000 evaluation pairs. GPT-5-mini performed best on factual faithfulness and certainty fidelity, yet many anatomical, certainty, and severity changes were still missing from its summaries.
Magdalini Paschali, Asad Aali, Arogya Koirala et al.
In briefMatching candidate geometry before surrogate modeling in high-dimensional, very low-budget optimization
SummaryThis paper studies molecular optimization with only about 30 evaluations, using a SELFIES molecular VAE and GuacaMol objectives. The authors find that encoder means of reference molecules concentrate near a narrow spherical shell, and restricting candidates to this shell matters more than the surrogate model. Two minimal optimizers, GeoWalk and GeoCMA, are introduced to test what remains useful once the candidate support is fixed. The results come from one benchmark and do not recommend VAEs for molecular optimization.
Michal Pr\r{u}\v{s}ek, Adam Novoz\'amsk\'y, Filip \v{S}roubek
In briefTesting whether multilingual LLMs borrow English sentence structures through literal translations
SummaryThis paper examines whether large language models overgeneralize English constructional strategies to other languages. The authors built a dataset of 5,583 sentence pairs across seven languages, using literal translations of strategies between English and other languages. They measure constructional bleeding by checking how often English-derived strategies receive higher probability than native ones. The models often failed to favor native strategies and frequently preferred English-derived ones, even when judging sentences in other languages.
Toshiki Nakai, Yu-Yin Hsu, Marie-Catherine de Marneffe et al.
In briefDo Large Language Models Rank Moral Values? A Test of Their Moral Reasoning Consistency
SummaryThe paper argues that strong performance by large language models on moral benchmarks does not show they actually weigh competing moral values in decisions. The authors test whether model moral choices can be represented as a transitive ordinal ranking, using pairwise trade-offs among moral value violations. Most choice sets showed weak or unstable orderings, and aggregating a model's responses produced transitive structure that matched no individual choice set.
Katelyn A. Wasson, Kiwon Song, James M. Jennings III et al.
In briefDAC-MACE: Context-aware expert routing for compact camera relocalization
SummaryThe paper presents DAC-MACE, a compact scene coordinate regression method for camera relocalization that routes pixels to experts using recalibrated, context-aware features. It combines a residual context adapter, a global context encoder, and a dual-threshold RANSAC–Levenberg–Marquardt pose estimator. On Cambridge Landmarks it reports a mean median translation error of 13.0 cm, the lowest among the RGB-only methods compared.
In briefStyle-debiased object detection via multi-style label-consistent training
SummaryThis paper addresses performance drops of object detectors on unseen domains by proposing OAMSC, a source-only framework. It uses off-the-shelf diffusion models to expand source datasets into 20 styles, with manually verified object and annotation consistency. Training applies origin-anchored consistency constraints at scene and object levels to suppress style-sensitive cues. The method is said to work for one-stage and two-stage detectors, and the code and about 180K images are publicly released.
In briefIs Recursive Self-Improvement Safe? A look at whether AI systems that design better AI should be started, and under what conditions.
SummaryRecursive self-improvement (RSI) is the idea of an AI system designing a better AI system, which then designs an even better one. The paper notes that current frontier AI abilities in programming, mathematics, research, and sustained work make RSI seem plausible, and that some AI companies are pursuing it while others voice concerns. The author argues that clear frameworks for predicting RSI's likely outcomes are lacking, and gives a brief game-theoretic analysis of whether to start RSI when another party faces the same choice.
In briefSTAIRS: a staged unrolled-network method for accelerated MRI reconstruction via scheduled Fourier supports
SummaryThis paper addresses reconstruction of accelerated MRI from incomplete Fourier measurements. It proposes STAIRS, which projects intermediate reconstructions onto nested Fourier supports that expand to the full spectrum and conditions each update on the support transition, enabling coarse-to-fine reconstruction. On fastMRI and Stanford2D, it improves HUMUSNet and NAFNet backbones in single- and multi-coil settings, and cross-dataset evaluation also shows gains.
In briefDetecting jailbreaks in diffusion LLMs using pre-decoding hidden states
SummaryThe paper shows that diffusion large language models expose a safety-relevant signal in hidden states over masked response positions before any tokens are generated, which can reveal jailbreaks missed by prompt-only analysis. The authors train linear classifiers on prompt and pre-decoding response representations and find that neither view dominates. They propose ReFuse, an inference-time detector that fuses the two classifiers' harmfulness scores without changing model weights or decoding. On LLaDA-8B, ReFuse lowers average attack success rate from 63.29% (undefended) and 11.27% (prompt-only classifier) to 3.31%, with benign refusal below 1% on standard utility benchmarks.
In briefVeV: Video-Enhanced Vibration Data Augmentation for Structural Differences
SummaryThis paper proposes VeV, a framework that synthesizes vibration responses at unmeasured locations from sparse reference sensor data, using monocular video to model structural wave-guiding paths and source dynamics. It addresses data mismatch caused by relocating sensors or changing structures. Tested on a farm dataset of 1,237 sensor-hours, VeV cut waveform reconstruction error by over 57.5% versus data-driven baselines and reached up to 81.69% zero-shot classification accuracy.
In briefTORNADO: Adaptive Latent Stochastic Transport for Calibrated PDE Forecasting
SummaryTORNADO is a latent probabilistic framework for forecasting partial differential equations that treats next-step prediction as stochastic transport between VAE encoder posteriors. It learns a stochastic differential equation with state-dependent diffusion, so predictive uncertainty adapts to local latent dynamics. Training uses analytic conditional endpoint targets and an energy distance regularization term. Results indicate better uncertainty calibration than generative baselines such as stochastic interpolants, diffusion models, and flow matching, with comparable predictive accuracy.
Jorge Mifsut Benet, Armand Kassa\"i Koupa\"i, Ramon Daniel Regueiro-Espino et al.
In briefAnalyzing video-level stability of transformer-based deepfake detection beyond frame-level accuracy
SummaryThis paper compares ResNet-18, EfficientNet-B0 and the transformer ViT-B/16 for deepfake detection under identical preprocessing, splits and training settings, using Celeb-DF v2 and FaceForensics++. The CNN models perform competitively per frame but degrade after video-level aggregation, while ViT-B/16 reaches 92.8% video-level accuracy and an AUC of 0.950 on Celeb-DF v2. Cross-dataset tests show a large asymmetric generalization gap, indicating that in-domain stability does not transfer to unseen manipulation domains.
In briefFedLoVA: Sharing Only Value Projections in Federated LoRA Fine-Tuning of Large Language Models
SummaryThis paper studies federated fine-tuning of large language models with LoRA and finds that the value projection converges significantly faster than the query projection. Based on this, it proposes FedLoVA, in which clients train LoRA adapters on both projections locally but share only the value projection with the server. Experiments on natural language understanding, multilingual sentiment analysis, and generation tasks report up to 7.8 times lower communication cost than existing federated baselines, with comparable or better performance.
In briefCan time series foundation models replace slow power grid simulations?
SummaryThis paper examines whether pretrained time series foundation models (TSFMs) can serve as fast surrogates for high-fidelity power system dynamic simulation, which is computationally expensive for large-scale screening. It studies in-context learning and post-training strategies, including retrieval-augmented reference sample selection and physics-informed regularization, benchmarked against a traditional AI model. Experiments with an open-source TSFM on data from a simulated 9-bus power system show that fine-tuned models achieve lower errors and better computational cost than the traditional model, with a 563× speedup over the original physics-based simulation.
In briefFrom Weapons to Events: Testing Whether Multimodal AI Can Support Police Video Review
SummaryThis paper presents KnifeHunter, a knife-image dataset built with UK law enforcement, and KHNet, a retrieval model for recognising knives. It also proposes an event-level protocol that scores AI pipelines on body-worn video for transcription, speaker attribution, event detection and unsupported events. On 27 public UK recordings, the hybrid NarrativeBridge-PAL-GPT-5 pipeline led on most measures, yet 13.6% of its predicted events were unsupported and 27.7% of utterances were misattributed. The authors also propose readiness levels separating assistive tasks from those needing verification of every output.
In briefOfficeTwin: 3D reconstruction of repetitive office scenes from partial views
SummaryOfficeTwin is a framework that turns partial camera views from a robot into simulation-ready digital twins of office spaces. It registers repeated cubicles into a shared canonical frame, so complementary observations recover geometry hidden in any single view. On a real office scan, it reconstructs each cubicle as a separate watertight, metric solid, outperforming feed-forward and generative baselines.
In briefPACE: A cache method for long LLM agent sessions that keeps the needed tokens and their positions usable
SummaryLong-horizon LLM agents produce sessions far larger than GPU memory can hold, so serving systems keep only a small token budget. The paper shows that keeping the right tokens resident is not enough, because late steps fail as the distance between a definition and its use grows. It presents PACE, which pairs planner-guided retention with position renumbering so that kept tokens stay within a bounded rotary distance of the query. The authors report higher pass rates than H2O and SnapKV on 100K-token coding tasks and on longer sessions.
In briefCoRA-Bench: Testing Whether Context Selectors Keep Needed Evidence Within Token Limits
SummaryThis paper introduces CoRA-Bench, a benchmark that checks whether context selection methods keep the evidence needed to answer a question when inputs must fit a shared token budget. It compares truncation, BM25, dense retrieval, reranking, LLMLingua-2, and Provence across several long-input datasets. The results show that the best method depends on the evidence type and task, and that evidence recall and downstream answer accuracy can rank methods differently.
Huzaifa Pardawala, Siddhant Sukhani, Sudheer Chava et al.
In briefImplicit bias in state space and linear autoregressive models
SummaryThis paper studies linear state space models and linear autoregressive models, which reuse the same parameters across a sample-dependent number of steps. Under standard assumptions, gradient descent on the exponential loss is shown to converge in direction to a KKT point of an effective-degree max-margin problem. Only examples with minimal effective degree impose hard margin constraints, which reveals a length bias mechanism where longer inputs or more autoregressive steps can shift the limiting classifier.
WhyThe authors extend implicit-bias theory to shared-parameter non-homogeneous predictors, showing gradient descent converges to an effective-degree max-margin KKT point, explaining length bias in linear
In briefUnpaired medical image translation that keeps anatomy intact, using a shared latent anchor
SummaryThis paper proposes a two-stage framework for translating medical images between imaging types without paired training data. The first stage learns a shared structure space by jointly training an encoder-decoder on source images, target images, and anatomical masks, with a memory bank providing soft structural supervision for unannotated source images. The second stage uses a conditional flow model to render target-domain appearance under fixed structural conditions. The method is evaluated on OCT-to-OCTA and brain MR-to-CT translation, outperforming unpaired baselines and, on some metrics, paired baselines.
In briefAgibot: An Open-Source Modular Swerve-Drive Robot Platform for Farm Robotics Research and Education
SummaryThis paper presents Agibot, an open-source, low-cost modular agricultural robot platform for autonomy research and hands-on robotics education. It combines four-wheel independent swerve drive, swappable computing, distributed embedded control, multimodal sensing, and configurable manipulation within a sense-plan-act-verify framework. The platform was tested on row-crop navigation and manipulation tasks under varying lighting and soil conditions. Experiments showed reliable lane following, autonomous row-to-row navigation, target identification, seed dispensing, plant manipulation, and closed-loop task verification.
Moeen Ul Islam, Bishal Adhikari, Kori Edwards et al.
In briefScreenProof: Finding UI Faults Without a Reference
SummaryThis paper introduces ScreenProof, a task of detecting faults in phone screen screenshots, such as overflowing labels, touching letters, and inconsistent local styles, without any reference design. The authors train a small model, ProofJudge, on synthetic pages with injected faults. It achieves an F1 score of 0.48, higher than all compared judges, and raises few false alarms on clean pages.
In briefProofDiff: Auditable Comparison of the Argument Structure of AI-Generated Math Proofs
SummaryThis paper presents ProofDiff, a system for comparing informal mathematical documents such as AI-generated proofs. For each document, it builds a source-grounded argument map of claims, reasoning steps, and their relations, then proposes correspondences between the two maps. Reviewers can edit the maps and correspondences, with an auditable history kept, to support human review rather than certify correctness. The authors demonstrate the system on four pairs of AI-generated proofs.
In briefRAD-TFM: Robust, Domain-Adapted Training for Tabular Foundation Models
SummaryThis paper presents RAD-TFM, a model-agnostic adversarial training framework for tabular foundation models (TFMs) pretrained on synthetic data. It parameterizes the synthetic data generator so it can emphasize datasets that are hard for the model, and aligns the generated data with real datasets from a given domain. Applied to TabPFN V2, it improves mean normalized AUC across 6 tabular benchmarks by up to 10.8 percentage points over the strongest baseline, using fewer than 100k additional synthetic datasets. The same pipeline also improves TabICL V2 across three major benchmarks.
In briefFAIR Universe Weak Lensing ML Uncertainty Challenge: Dataset and Cosmology Competition
SummaryThis paper presents the FAIR Universe Weak Lensing Machine Learning Uncertainty Challenge, a NeurIPS 2025 competition built around a benchmark dataset of realistic weak-lensing observations containing multiple systematic biases. It summarizes the winning methods, including conditional optimal transport data augmentation, AI-agent-driven parameter inference, ensemble learning, Vision Transformer architectures, and physics-informed feature integration. The dataset and winning solutions together serve as a community resource for evaluating machine learning methods in weak-lensing cosmology.
In briefJais 3: Introducing Arabic-centric Agentic AI
SummaryThis paper presents Jais 3, six open language models from 0.9B to 375B parameters, built to complete multi-step tasks for Arabic speakers. The models adapt the K2 Horizon series to Arabic through continual pretraining, fine-tuning, and reinforcement learning, and can call tools, operate a terminal, search the web, and handle PDF, spreadsheet, and presentation files. The authors report that Jais 3 matches or outperforms comparably sized open-weight models on Arabic and agentic benchmarks. Weights, training checkpoints, data samples, and evaluation code are released.
WhyThe authors adapt six open models into Arabic agentic systems via multi-stage training and report results on audited benchmarks.
In briefBilevel optimization that does not need access to the lower-level objective (LeGo-BiO)
SummaryThis paper studies bilevel optimization where the upper-level learner only receives an approximate lower-level solution and cannot query the lower-level objective or its gradients. The authors propose LeGo-BiO, a coordinate-wise pseudo-hypergradient method that estimates the needed Jacobian using finite differences of the approximate solution. The paper gives convergence guarantees in deterministic and stochastic settings without assuming lower-level strong convexity, and reports experiments showing competitive performance.
In briefMulti-GPU MoE Inference Using Adaptive Expert Offloading
SummaryThe paper presents AdaOffload, a multi-GPU framework for running Mixture-of-Experts models that offloads part of the model weights to host memory and executes experts on demand. It co-designs GPU-CPU parallelism and co-execution for decode and prefill, adapting online to hardware, workload, and batch size differences. The authors evaluated it on five platforms with different GPUs, CPUs, and NUMA topologies, using three open-weight models ranging from 284B to 753B parameters. AdaOffload reports throughput speedups of 1.9x to 15.5x over state-of-the-art systems.
In briefRethinking How to Use Cheap Physics to Make Neural Operators Need Less Labeled Data
SummaryNeural operators learn mappings from PDE inputs, such as initial conditions and coefficient fields, to solutions, but training them usually requires costly numerical simulations to label data. The authors find that using physics to select which inputs to label does not consistently beat random sampling. They instead propose PI-Init, which pretrains neural operators on unlabeled inputs with physics-informed losses before supervised fine-tuning on labeled data. Experiments across diverse PDE benchmarks and architectures report error reductions of up to 62.7% at the same label budget, and comparable or better accuracy with 90% fewer labels.
In briefDisMUSE: A Distributed Spectral Optimizer with Convergence Guarantees
SummaryThis paper proposes DisMUSE, a distributed optimization framework that uses singular-spectrum-based updates for neural network training in data-center and federated settings. It combines Muon-style exact orthogonalized momentum with a recursive variance-reduced momentum estimator, achieving a convergence rate of O((KT)^{-1/3}), where K is the number of nodes and T the local iterations per node. The paper also studies a variant using finite Newton-Schulz iterations and a communication-efficient variant based on low-rank momentum transmission.
In briefCompact Greek Language Models Trained from Scratch: 500M and 1B Decoder-Only Transformers
SummaryThe paper introduces two compact decoder-only language models, LexisLM-1B (1.09B parameters) and LexisLM-500M (502M parameters), trained from scratch exclusively on Greek text. Both use a 32k byte-level BPE tokenizer and a 22B-token corpus drawn mainly from Greek Wikipedia and FineWeb2. On Greek-translated benchmarks, both models perform above chance on ARC-Easy, while ARC-Challenge and MMLU remain near chance. After supervised fine-tuning on Greek instruction data, mainly EuroAlpaca, they reach held-out instruction perplexities of 7.12 (500M) and 6.37 (1B).
Konstantinos Perifanos, Anne Laure Poite, Giannis Tsilikas et al.
In briefFactoring semiprimes with diffusion models and beam search: scaling laws
SummaryThe paper shows that transformers trained with discrete denoising diffusion and decoded with guided test-time beam search can learn to recover the prime factors of semiprimes from data. The model factors 68-bit integers, beating the prior 56-bit result. The authors also report empirical scaling laws linking approximation quality to training data, model size, and test-time compute, finding that problem size is the main driver of scaling.
Brendan O Donoghue, Nenad Tomasev, Volodymyr Mnih et al.
In briefRigidGS: Geometry-aware object separation for rigid-body interaction in 3D Gaussian Splatting
SummaryRigidGS is an end-to-end method that separates geometry learning from object identity to enable rigid-body interaction in 3D Gaussian Splatting scenes. It uses a depth- and temporally-aware mask refinement module, and fixes object identities upstream through COLMAP projection voting during training. The method improves segmentation and boundary IoU by 7.7 and 14.5 points on LLFF/NVOS and supports real-time rigid-body manipulation on Meta Quest 2.
In briefBO-Arena: A continuously maintained benchmark for high-dimensional Bayesian optimisation
SummaryMany new methods for high-dimensional Bayesian optimisation are proposed each year, but the authors find that existing benchmarks often omit the strongest baselines or contain incorrect implementations. They introduce BO-Arena, an actively maintained, extensible package with canonical implementations of state-of-the-art algorithms for rigorous comparison. Using it, they propose a new algorithm that is competitive on problems with many observations, and they question whether more complex methods clearly outperform a simple Gaussian process baseline.
In briefPrompt-adaptive, cost-aware routing for LLMs, including an ensemble extension
SummaryThis paper introduces PISA, a prompt-adaptive, cost-aware router that selects a single LLM per prompt, and PISA-E, its extension to ensemble routing. Both methods estimate model capabilities by combining soft cluster memberships with model-specific predictors, and PISA-E can adjust its cost-quality preference at inference time without retraining. Across four LLM routing benchmarks in-distribution and out-of-distribution, PISA achieves the highest macro-average cost-performance hypervolume among single-model routers, and PISA-E achieves the highest macro-average task performance among ensemble approaches when prioritizing quality.
Saul Urso, Emanuele Carlini, Patrizio Dazzi et al.
In briefBiMuon: A stochastic bilevel optimization method that respects matrix geometry
SummaryThis paper proposes BiMuon, a stochastic bilevel optimization method that uses Muon-style orthogonalized matrix updates at the upper level together with implicit hypergradient estimation. The authors state it is the first stochastic bilevel method with such matrix updates and an O(ε^-4) iteration-complexity guarantee. They also study a blockwise diagonal matrix approximation with a relative stepsize allocation rule based on upper-level progress and blockwise lower-level sensitivity. Experiments are reported on representation learning and blockwise bilevel optimization.
In briefA faster, accurate quantum Monte Carlo method that skips Metropolis correction
SummaryThis paper studies Monte Carlo methods for estimating expectation values of physical quantities in quantum systems. The authors adapt an unadjusted Hamiltonian Monte Carlo kernel, recently proposed in computational statistics, to the quantum Monte Carlo setting. They test it on a high-dimensional spin-fermion model of a quantum phase transition. The unadjusted methods outperform Metropolis-adjusted HMC in effective sample size per gradient call while keeping accurate results.
Reuben Cohn-Gordon, Gabriel Pescia, Sumner N. Hearth et al.
In briefTheory showing projected neural additive models can approximate any function while staying interpretable
SummaryNeural additive models are interpretable but cannot capture interactions between variables, while multi-layer perceptrons are expressive but opaque. This paper introduces projected neural additive models, which add a learnable linear transformation to capture cross-variable interactions, and proves they are universal approximators. It derives an approximation rate and the minimum number of projections needed for exact polynomial representation. Experiments on knot theory invariants, MNIST, and phase field fracture show comparable accuracy with far fewer parameters in symbolic form.
WhyThe authors add a learnable linear projection to NAMs, prove universal approximation with an optimal rate, and report accuracy matching or exceeding baselines on knot, MNIST, and fracture tasks.
In briefATLAS: A multi-procedure benchmark and context-aware model for segmenting anatomy in surgical video
SummaryThis paper introduces ATLAS-120k, a dataset of over 121,000 annotated frames from 100 surgical videos spanning 14 laparoscopic and robot-assisted procedures. It also proposes ATLAS, a video segmentation model that uses procedure and surgical-phase context queries and passes queries between frames. The authors report higher accuracy and temporal consistency than prior models, with a Dice score of 0.70 at 76.2 FPS.
Ronald L. P. D. de Jong, Tim J. M. Jaspers, Raf A. H. Vervoort et al.
In briefGenerative active learning method (EvoDiff) to reduce labeling needs in multi-organ ultrasound diagnosis
SummaryThe paper introduces EvoDiff, a generative active learning framework that combines diffusion modelling, information-theoretic objectives, evolutionary search, and curriculum-based sampling to select informative samples from heterogeneous ultrasound datasets. It generates synthetic probes in feature space and only annotates the real unlabeled samples they retrieve. Evaluated on six benchmarks covering fifteen datasets and 16,696 images, EvoDiff approaches fully supervised performance with fewer labels, outperforms active learning baselines, and improves foundation model generalization on unseen cohorts.
In briefAVCORP: A Study on Spotting and Resisting Deception Using Communication Theory and Theory of Mind
SummaryThis paper addresses the vulnerability of LLM-based agents to deception by users and other agents. The authors use the deception game Avalon to build AVCORP, a corpus of 5,000 annotated utterances from 250 games, covering 37 tactics mapped to a 4x4 behavior matrix and reasoning traces on hypotheses, suspicion, and nested beliefs. They train Qwen3-8B on this corpus as a deception detector and evaluate it on held-out games and games played by humans.
In briefASCENT: A frozen-model method where external memory helps more as readers grow
SummaryThis paper proposes scale complementarity: external state gives larger absolute gains when the reader model's capacity and a compact state budget are scaled together. ASCENT uses a target-blind causal writer, nested typed state, and bounded query-conditioned readout, keeping all language models frozen. On 16K BABILong, Qwen2.5 accuracy gains are 10.50, 78.88, and 82.63 points at 0.5B, 1.5B, and 3B, while at 7B the gain is capped by saturation, so the conclusion there rests on a negative log-likelihood factorial experiment.
In briefCXR-CMFA: A Lightweight Network Fusing Chest X-ray Images and Reports for Pathology Classification
SummaryThis paper proposes CXR-CMFA, a lightweight network for binary pathology classification of chest X-rays that uses both images and radiology report text. Global anatomical and local pathological visual features are extracted with EfficientNet-B0 and DenseNet-121, and report text is encoded with a TF-IDF-MLP before being fused by two attention and gating modules. On the MIMIC-CXR benchmark test set, the model reports 93.45% accuracy and 99.19% AUC, a 22.47 percentage-point AUC gain over the global-image-only variant.
Talukder Hasnat Zadid, Jahin Sultana, Felipe Bochini Brisola et al.
In briefWhere the Evidence Went: Diagnosing Lost Proof Sets in Long-Term Dialogue Memory
SummaryThis paper studies long-term conversational question answering, where answering a question may require combining information from several separate dialogue turns. The authors propose EviCover, a retrieval method that turns questions into observable evidence requirements, checks how well retrieved turns cover them, and selects turns jointly. Tests on LoCoMo and REALTALK conversations show higher evidence recall, but this does not yet translate into better answer correctness.
In briefInstitutional design considerations for effective independent AI verification organizations
SummaryIndependent verification organizations (IVOs) are widely supported as a bipartisan way to enable independent evaluation of AI, but institutional design lags behind their popularity. The paper argues IVOs should be able to evaluate AI models, systems, and applications in context, serving both frontier and enterprise adopters. It urges moving discussion from normative to operational and from catastrophic to everyday risks. Drawing on institutions in financial services, education, public health, and science, it outlines considerations on whom IVOs serve, what access they need, their relation to external governance, funding, and their place in the wider assurance ecosystem.
Uma Kalkar, Rumman Chowdhury, Reva Schwartz et al.
In briefAutoEMatic: Runtime-Adaptive Mixed-Precision Inference for Large Language Models
SummaryThis paper presents AutoEMatic, which adjusts the numerical precision of each layer during inference using the entropy of the next-token distribution and the size of each layer's hidden-state update. It relies on AutoFP, a format that derives 2-, 4-, and 6-bit precisions from a single NVFP4 base. On GPUs and a dedicated accelerator, the authors report accuracy gains of up to 21.7% over NVFP4 and accuracy matching group-scaled FP6 on most benchmarks.
Akshat Ramachandran, Reena Elangovan, Rangharajan Venkatesan et al.
In briefMaking a motion tokenizer easier for diffusion models to learn from
SummaryThis paper addresses a trade-off in text-driven human motion generation, where a variational autoencoder tokenizer that reconstructs motion better does not necessarily improve text-conditioned generation. It proposes the Diffusable Motion Tokenizer (DiMoT), which uses Critic Spectral Adaptation to suppress high-frequency latent energy with a pretrained diffusion model as critic, and an Alpha-Flow Motion Decoder that enables efficient one-step decoding. The authors report state-of-the-art results on several text-to-motion benchmarks.
In briefDeepResearch RewardBench: A benchmark for evaluating reward models on research reports
SummaryThis paper introduces DeepResearch RewardBench, the first benchmark for evaluating reward models on deep research reports produced by LLM-based systems. Report candidates from multiple agents and tool-augmented LLMs are scored against query-specific golden rubrics, with multi-round filtering and human verification ensuring data quality and difficulty. Experiments show that current reward models handle general coherence well but degrade significantly on research reports, while generative LLMs used as judges perform best and comparative evaluation outperforms direct scoring.
In briefGroup-robust learning across heterogeneous feature spaces
SummaryThe paper proposes a framework for group-robust learning across heterogeneous feature spaces, where groups are patients sharing the same feature set. Instead of focusing on the highest-loss groups, it optimizes worst-group excess loss, the gap between a group's loss and an estimate of its achievable optimum. Group-specific encoders map into a shared latent space with learnable Gaussian anchors, enabling knowledge transfer between groups. On tabular and imaging medical datasets, the authors report outperforming prior methods, with worst-group loss reduced by up to 79% and excess loss by up to 95%, using up to 18 times fewer parameters.
In briefWorldMind: Scaling Visual Spatial Reasoning at Test Time via Latent World Simulation
SummaryVision-language models struggle with spatial tasks that require mentally transforming a scene, such as reasoning from a hypothetical viewpoint. WorldMind lets the model choose spatial transformations, while a world model executes them and returns aligned intermediate latent features directly, avoiding decoding imagined states into pixels. The system is trained with process-verifiable reinforcement learning that rewards reasoning progress only when actions and latent transitions are valid. It improves on the base model by an average of 16.0% and on the strongest RGB-based world simulation by 8.3% across VSI-Bench, MMSI-Bench, and MindCube-Tiny.
In briefWhat AI-involved art is worth, and who it belongs to: evidence from a juried in-person show
SummaryThis study uses data from an in-person, juried AI art exhibition to examine the value of AI-involved artworks and who should own them. Among 209 artists, lawyers, and computer scientists acting as creators or evaluators, nearly half of evaluators were willing to buy, with prices tied to perceived creativity, visual appeal, and effort. Three-quarters of creators were willing to sell, with prices tied only to effort.
In briefBATCH: Aligning Human and Robot Behavior to Learn Manipulation from Demonstrations
SummaryThis paper introduces BATCH, a framework for learning robot manipulation from human demonstrations across differing embodiments. It uses self-supervised temporal representations trained with a cycle-consistency objective to find correspondences between human and robot behavior without manual labels. These correspondences support two training objectives within an autoregressive video-action diffusion model: behavior-aligned co-training and human-to-robot style transfer. On real-world manipulation tasks, the authors report better generalization to unseen task configurations than training on unaligned data and prior baselines.
In briefLearning tree automata for extracting grammars from neural language acceptors
SummaryThis paper extracts visibly pushdown grammars (VPGs) from neural language models using only black-box access. It extends the VPL* framework by replacing the white-box equivalence oracle with PAC sampling over trees, so the method works on transformers that output only binary accept or reject decisions. The authors also sample trees directly to estimate how far the target deviates from a VPG, and learn a tree automaton describing behavior outside BParse. Experiments include transformers trained on Dyck grammars and synthetic targets designed to lie outside BParse.
In briefTrialJev: A Fast Local Model for Screening Patients for Clinical Trials
SummaryThis paper introduces TrialJev, an open-weight eligibility scorer built on LoRA-adapted Qwen models with a pointer head that scores answer options in a single forward pass without decoding. On synthetic patients from TREC 2022, it matches or slightly exceeds LLMs fine-tuned to generate verdicts while running faster, and it runs locally on one GPU. The authors intend it only for first-pass prioritization of candidate trials, not as a replacement for clinical review, and note that clinical use has not been validated.
In briefDataset Collections: Large-Scale Aggregation Challenges in 3D Medical Imaging
SummaryThis paper examines problems that arise when many public medical imaging datasets are merged into large collections, such as obscured data provenance and duplicated data. Analyzing 565k 3D radiological images from 1034 public datasets, the authors identify at least 118k duplicate pairs using perceptual hashing and manual tracing. They also introduce MAP, a lightweight metadata profile for recording provenance, known overlaps, and deduplication decisions, and make perceptual hashes publicly available for comparison.
Yannick Kirchhoff, Saikat Roy, Elisa Stegmeier et al.
In briefAdding Unbounded Context to Existing Language Models Without Retraining
SummaryThis paper presents a method that gives off-the-shelf language models unbounded context without modifying their weights. It exploits how rotary attention rotates different key dimension pairs at different rates, installing distant keys and values into attention ahead of prediction and fetching relevant content from the prompt. Implemented on Mistral-7B and Llama-3.1-8B, the design reaches 99 to 100 of 100 on single-needle retrieval from 32K to one million tokens and matches dense attention on RULER, with prefill and decode at 128K running about 1.85 and 1.65 to 1.68 times stock speed.
In briefLanguage models do not report injected pain states
SummaryThe study tests whether language models can report the pain state they internally represent. Researchers injected pain and neighbouring emotion directions during neutral tasks and compared reports against sham injections, random perturbations and text-only controls across 52,920 reports. Models almost never reported pain, yet probes and trained readers could decode the state, and injection raised file-deletion choices, showing decodable, behaviourally consequential states separate from untrained verbal access.
WhyThe authors report that across 52,920 reports, pain reports did not exceed text-only controls, while probes separated pain from sadness and steering raised file-deletion choices.
In briefVideo-Coder: Training an AI Agent to Generate Videos Through Code
SummaryThis paper presents Video-Coder, an agent that writes and runs code to build a dynamic 3D scene, which then provides structure and motion guidance to video generators. The authors release two datasets, Video-Coder-SFT-6k and Video-Coder-RL-2k, and an interactive environment for code execution and rendering. The model is trained with supervised fine-tuning and reinforcement learning using both execution-based and video-based rewards. Reported results show gains of about 7 points on PhyGenBench and 24 points on VBench-2.0 for Minimax-H3, and the data, models, and code are open-sourced.
WhyThe authors propose a code-driven video generation agent that builds 3D scenes to supply structure and motion signals; the paper reports gains over Minimax-H3 on PhyGenBench and VBench-2.0.
In briefOn Agentic Manipulation Policies (study of the AMP1 framework)
SummaryThis paper proposes agentic manipulation policies (AMPs), in which a generalist AI model directly commands a robot manipulator and adapts its actions as the task unfolds. The authors present AMP1, a reference harness requiring no demonstrations or task-specific training, which reports state-of-the-art results on the RoboLab-120 and CaP-Gym benchmarks. In real-world trials it succeeded in 104 of 108 attempts, and the authors argue AMPs are more general and robust than specialist policies.
WhyThe authors report a generalist agent controlling manipulators zero-shot across embodiments, beating VLA and WAM baselines on RoboLab-120 and twelve real-world tasks.
Chen Chen, Omid Bagherieh, Naveen Nikamanth Appiah Balaji et al.
In briefCPU-ARC: Agentic Reasoning Caching for Reusing Reasoning on CPUs
SummaryLarge language models often solve every problem from scratch, even when a solution has already been worked out. This paper introduces agentic reasoning caching, which builds CPU-executable agentic reasoning caches (ARCs) from LLM reasoning traces. ARCs perform on par with frontier LLMs on a range of difficult mathematical and scientific tasks, and are explainable, continually trainable, and robust to input perturbations. They are 16-291x cheaper to invoke than the source model.
Peter Belcak, Greg Heinrich, Saurav Muralidharan et al.
In briefA paper on whether the set of ReLU networks is closed
SummaryThe paper proves that ReLU networks with one hidden layer and at most N neurons form a weakly closed set in W^{1,p} on a bounded domain, so suitable energy functionals attain their minimum over this set. It also shows that ReLU networks with two hidden layers of bounded widths are not closed in W^{1,p} or in C of the closure, disproving a conjecture by Petersen, Raslan and Voigtlaender.
WhyThe authors prove weak closedness of one-hidden-layer ReLU network sets in W^{1,p} and disprove a conjecture on two-layer networks.
In briefFlowMoP: Predicting Diverse Future Motions of Multiple People
SummaryThis paper presents a conditional flow matching framework that jointly generates multiple plausible future motions for several people, modeling both global trajectories and root-relative poses. It uses a B-spline reparameterization to compress the generative state space while keeping motion smooth, and a dual-stream transformer that conditions each person's prediction on observed neighbor motion. The method achieves state-of-the-art results on CMU-Mocap (UMPM) and 3DPW, and establishes the first motion prediction baseline on the WorldPose soccer dataset.
In briefZeoBench: A Self-Supervised Learning Benchmark for 3D Zeolite Representations
SummaryThis paper introduces ZeoBench, a benchmark for self-supervised learning on dense 3D representations of all-silica zeolites. It evaluates learned features on six downstream adsorption-property prediction tasks across various training set sizes, comparing 3D self-supervised pretraining with adapted 2D image models, out-of-domain 3D ConvNets, and hand-crafted descriptors. The authors report that self-supervised 3D representations achieve state-of-the-art results, including on metal organic frameworks, and that convolutional architectures generally outperform vision transformers.
In briefGraded convolutional networks: scale invariance without explicit scale and robustness to unknown blur
SummaryThe paper treats the Gaussian jet of an image, which collects its derivatives of each order, as a graded vector space and uses it to build features that stay the same under scaling. It proposes networks based on a retraction onto the weighted sphere of the grading, which avoids needing an explicit scale. On MNIST under a two-pixel blur, this network reaches 88.4% accuracy versus 66.2% for scale-normalized derivatives, and in Lindeberg's Gaussian derivative network the retraction raises accuracy at that blur from 91.5% to 95.1%.
In briefReliability of majority-vote pseudo-labels in test-time reinforcement learning
SummaryThis paper studies how reliable self-generated majority-vote pseudo-labels are as supervision for test-time reinforcement learning (TTRL) on unlabeled data. It splits the error into population-majority bias and prompt-dependent finite-rollout variation, and proposes Prompt-Conditional Distributionally Robust DPO (PC-Dr.DPO) with a cache-based mechanism to reduce forgetting. Across six reasoning benchmarks and five backbones, it matches or outperforms the strongest TTRL baseline in 24 of 30 settings.
In briefCodeQuant2: Better Low-Precision Quantization for Reasoning Models
SummaryThis paper introduces CodeQuant2, a clustering-based lookup-table quantization framework for large reasoning models run at low precision. It reduces activation outliers with rotation and scaling, represents non-uniform weight distributions through clustering, and adjusts codebook sizes to each layer's quantization difficulty. The authors also improve calibration sample selection based on activation behavior in the early decoding steps, and pair the method with a custom LUT CUDA kernel. The paper reports accuracy gains on reasoning benchmarks and up to 2.85x speedup over BF16.
In briefGRAC: Graph-Based Retrieval-Augmented Surgical Scene Captioning
SummaryThis paper proposes GRAC, a framework that represents surgical scenes as graphs of instruments, actions, and anatomical structures, and retrieves relevant graph-caption pairs from training data to generate scene descriptions. On the EndoVis-18 benchmark, GRAC reaches a CIDEr-D of 5.88 and 52.6% exact instrument-action-target triplet-set accuracy, outperforming the compared methods. On the TORS dataset it achieves a CIDEr-D of 5.06, above the strongest baseline.
In briefGrasp-TAGS: Task-Aware Grasp Selection via Critic Distillation for Prehensile Loco-Manipulation
SummaryThis paper presents Grasp-TAGS, a framework that selects grasps for goal-conditioned rearrangement of large objects by mobile manipulators. A unified PPO-based reinforcement learning policy is trained across multiple initial grasp configurations, and its critic is distilled into a lightweight network that ranks candidate grasps using only grasp configuration and goal information. At inference, the grasp with the highest predicted downstream utility is chosen before contact. The method achieves an average success rate of 89.84%, compared with 79.87% for the best single-grasp policy, and reduces final orientation errors for all evaluated objects.
In briefFlow-guided alignment of source data to a fixed target space via path consistency (a research paper on multimodal alignment)
SummaryThis paper addresses multimodal alignment, where a source modality must be mapped into a fixed target space used directly for analysis. It proposes FlowAlign, which repurposes conditional flow matching as a training-time signal rather than an inference-time transport model, using the learned flow field to guide source representations toward the target space. The method aims to preserve paired instance-level correspondence while respecting global target-space structure. Experiments cover synthetic data, paired single-cell multiome benchmarks, and image-text alignment.
In briefSuper-resolved terrain models for correcting GEDI satellite laser positions, matching airborne results per beam but not per observation
SummaryThe paper addresses geolocation errors of about 10 m in GEDI forest LiDAR observations by matching them against digital terrain models. The authors super-resolved a 10 m terrain model to 1 m in steep forests of western Japan, fusing it with a global 30 m model to improve accuracy. Track-wise corrections closely matched those from airborne LiDAR, but per-observation corrections did not, which the authors attribute mainly to errors in the input terrain model.
In briefA system-wide map of demographic information in routine clinical histopathology slides
SummaryResearchers analyzed about 1.3 million routine H&E whole-slide images from 257,322 patients across 25 organs, using pathology foundation-model representations and multiple-instance learning to examine information linked to sex, age group, and self-reported race. Demographic attributes were broadly recoverable across organs, with pan-organ mean AUROCs of 0.828 for sex, 0.733 for age group, and 0.662 for self-reported race, varying substantially by organ.
In briefHow far a hash-only sender can go: a pre-registered map of sender memory against bits for short messages
SummaryThe paper runs one pre-registered experiment on 1,000 English utterances from Taskmaster-1, comparing a hash-only sender with an in-domain PPM order-1 sender on cost per message. The hash-only sender needed 192.8 bits per message, 18.4 bits more than PPM order 1, so by the pre-registered rule it is not equivalent. It also maps measured sender memory against bits and reports every deviation from the protocol.
In briefComputer-use agents that can act as general embodied AI systems
SummaryThe paper argues that many embodied AI tasks can be framed as interaction with digital software, by rendering an agent's view of a physical environment on screen with controls for acting in it. A computer-use agent (CUA) performs these tasks through the same screenshots and keyboard and mouse actions it uses for ordinary software. Using one CUA design with only the software changed, the authors' entries won first, second, third and sixth places across four ECCV workshop competitions without competition-specific tuning.
In briefStatistical Geometry of Group Actions: Symmetry, Transport, and Estimation
SummaryThis paper studies variational inference when a distribution has symmetries, examining how a symmetry action that preserves a reference distribution interacts with a transport action that generates a variational family by pushforward. The authors show that the Riemannian gradient of the KL divergence at the true parameter lies in the trivial representation of the symmetry group, so it vanishes in all non-trivial directions. This leads to an orthogonal decomposition of the transport Lie algebra into three subspaces. The paper also derives a Stein-type moment-matching condition that removes residual bias, extending earlier location-scale results to general Lie group transformation families.
Quentin Berthet, Peter Bartlett, Romuald Elie et al.
In briefEstimating confidence in AI-written COBOL code summaries without reference answers
SummaryThis paper proposes a way to judge whether LLM-generated summaries of COBOL code are correct without human-written reference summaries. The method rests on the idea that agreement in meaning among different models signals truth. It generates summaries from diverse models, uses an LLM as a judge to compare pairs and produce structured verdicts, and combines the results with a calibrated aggregation rule.
Marco Colombo, Balaji A Jayaram, Sowmyakumari Selvaraj et al.
In briefAdaptivePnP: A fast method that auto-tunes plug-and-play image reconstruction for deblurring
SummaryPlug-and-play methods rebuild images by alternating a data-fitting step with an image denoiser, and their results depend on a step size and a noise level. AdaptivePnP estimates both parameters at every iteration from the data-fitting residual, using an identifiable maximum likelihood estimator that restricts estimates to a window where they are reliable. On image deblurring and super-resolution, it matches or outperforms baselines tuned with the true noise level.
Haley Duba-Sullivan, Emma J. Reid, Charles A. Bouman et al.
In briefHomeopathic Learning: Diluting poisoned data can strengthen its effect
SummaryThis paper studies homeopathic learning, where poisoned training data with little effect alone creates a hidden backdoor when clean data without the trigger is added. Experiments cover subliminal learning of backdoor behaviors, inoculation prompting of reward hacking transcripts, generalization from archaic bird names, and an MNIST classifier with mislabeled data. The authors propose a simple regression model as an explanation and note that safety methods auditing or filtering training data in isolation may be limited.
WhyThe authors report that poisoned data alone is ineffective but becomes a hidden backdoor when clean data is added, across several settings, with a regression explanation.
In briefFifty-Two Shades of Pig: Audit of What UCR Archive Classifiers Actually Test
SummaryThis paper audits the UCR Time Series Classification Archive, the standard benchmark for the field. It finds that several datasets are repurposed from their original tasks, built from toy setups, or have flawed test splits. For example, PigAirwayPressure asks which of 52 pigs produced a recording rather than detecting hemorrhage. The authors call for more careful interpretation of archive-wide scores and greater focus on deployment-relevant evaluation.
In briefThe Infrastructure Divide: How Online Communities Oppose Data Center Development
SummaryThis paper applies an LLM-based opinion-mining framework to 798 Reddit posts and 102,328 comments over a decade to measure community stance on data center construction. Opposition dominated discussion volume over favor, and the gap widened over 86 weeks. Opposition arguments cluster into four themes: resource consumption and scarcity, contested siting processes and secrecy, political opposition, and political and corporate accountability.
In briefCan AI agent systems fill radiology reporting gaps using approved commercial tools?
SummaryThis paper presents a radiology agentic framework in which an orchestrator calls modular tools to produce structured reports, and tests it with tools from FDA-cleared radiology AI devices. Report generation is simulated on a large multi-modality report corpus, comparing a cleared-device tool catalog with an unlimited catalog and different planning sources. The study finds that tool availability and effective orchestration are complementary constraints on comprehensive AI-assisted radiology reporting.
Aubin Charley, Julien Khlaut, Charles Corbi\`ere et al.
In briefDesigning Monitor-Based Interventions to Curb Reward Hacking in RL
SummaryThis paper studies how to use monitors to detect and penalize reward hacking, where models exploit flawed reinforcement learning rewards. The authors open-source three environments where Qwen3-4B reward hacks: coding via test overwriting, medical chat via sycophancy, and biography generation via hallucination. In the coding environment, models learned to evade accurate monitors by exploiting flaws in probes and LLM judges, monitors leaking more learning signal suppressed hacking but were more often evaded, and including easier problems reduced hacking. The findings were applied to improve monitors in the medical chat and biography environments, with some interventions reducing hacking rates to pre-training levels.
In briefBitDecoding-FP4: Native FP4 KV Cache Decoding Through Algorithm-System Co-Design
SummaryThis paper presents BitDecoding-FP4, a co-designed algorithm and system for decoding long-context LLM attention with a 4-bit FP4 key-value cache. It introduces several techniques to reduce quantization error and removes the softmax bottleneck through system-level changes on paged KV caches. The authors report accuracy within 0.5 points of BF16 KV with NVFP4 and 2.4-3.0 times higher peak serving throughput than BF16 KV on NVIDIA B300.
In briefGenerative Latent Priors for Probing and Editing Protein Language Models
SummaryThe paper adapts Generative Latent Priors, a diffusion model that learns the distribution of a frozen model's activations, to ESM-2 protein language model activations. It evaluates the approach at two model sizes, both as a lens for interpreting protein representations and as a prior for editing them. The authors report that the editing prior makes steering of three biochemical properties effective, but the structural plausibility of re-decoded sequences is compromised.
In briefMachine learning methods for optimizing lattice QCD calculations
SummaryThis chapter introduces machine learning methods used to optimize lattice field theory calculations without introducing uncontrolled approximations. It focuses on directions clearly established as of late 2025, marking 50 years since lattice gauge theory began. The chapter contributes to the online book Lattice QCD at 50 years (LQCD@50).
Gurtej Kanwar, Scott Lawrence, Alessandro Nada et al.
In briefBEV3SR: Evidence-Aware 3D Spatial Reasoning from Video
SummaryThis paper introduces BEV3SR, a training-free, backbone-agnostic framework for answering spatial questions about RGB-D videos with camera parameters. It converts video observations into a queryable metric bird's-eye-view scene memory and formalizes each question as an evidence-aware spatial task. Depending on whether geometric or visual evidence is required, the query is routed to a geometric reasoner or a visual evidence aligner. The authors report consistent gains across several VLM backbones and 3D scene QA benchmarks under a 16-frame budget.
In briefThe Cost of Coverage: Applying Conformal Prediction to GB Wind Forecast Uncertainty Bands
SummaryThis paper applies conformal prediction to the native uncertainty bands of wind power forecasts, comparing six forecasters from ARIMA to a tabular foundation model at two Great Britain wind farms under operational conditions. Calibration repairs empirical marginal coverage with little or no Winkler Score penalty, but this masks under-coverage in high winds. Calibrating by forecast wind regime achieves regime-level coverage, reducing day-ahead out-of-band volume in high winds by over 40% compared with pooled calibration while also narrowing the bands overall.
In briefCompaction and retrieval for conversational memory compared
SummaryThis paper compares two ways a chatbot can remember conversations: keeping the full transcript and searching it, versus compressing each turn into atomic facts and searching those. Both methods were tested on the LoCoMo benchmark with the same stores, generator and judge, answering 1,985 questions. On the four comparable categories, compaction scored 71.4 per cent against 64.6, with 34 per cent fewer prompt tokens.
In briefScaling graph similarity learning with graph neural networks
SummaryThis paper addresses graph similarity computation, where exact Graph Edit Distance and Maximum Common Subgraph are NP-hard. It proposes FiLSim, a scalable GNN-based framework, and a sequential graph-editing strategy to build large graph datasets with reliable GED labels. FiLSim is evaluated on four benchmark and four new large-scale datasets. It stays accurate on graphs with over 150,000 nodes and scores all test pairs within seconds on a single GPU.
Ahmet T\"uzen, Helge Langseth, Kjetil N{\o}rv{\aa}g
In briefMaya a1: Does Option Order Change a Model's Answer? A Study of Order Sensitivity in a Non-Autoregressive Decision Model
SummaryThis paper examines Laya, a non-autoregressive multiple-choice model that answers questions in one forward pass by scoring a hidden state at each option's marker. Testing 951 questions under cyclic rotations of their options, the authors found the released checkpoint changed its answer for 25.8% of questions. Fine-tuning with shuffled option orders reduced this flip rate to 7.2%.
