In briefReAnchor: Reusing Discarded Verifier States to Speed Up Diffusion-Based Drafting
SummaryThis paper addresses speeding up speculative decoding, where a draft model proposes tokens that a larger target model verifies. Block diffusion drafters produce multiple draft tokens in one pass but only from a single anchor, and the target's verification hidden states beyond the first rejection are usually thrown away. ReAnchor is a training-free method that reuses these discarded states to reanchor a pretrained drafter at shifted positions, producing parallel draft blocks. At temperature T=0, it improves acceptance length by 19.4% and decoding speed by 17.9% on Gemma 4-12B across eight benchmarks.
In briefA pipeline building a semantic knowledge graph for vehicle intention in nuScenes
SummaryThis paper presents the nuScenes Intention Knowledge Graph (NIKG), an OWL-based semantic representation and construction pipeline for ego-centric vehicle intention context. It encodes ego-relative motion, spatial configuration, and time-to-collision risk as first-class properties of scene participants, using only ego pose, object annotations, and keyframe order as input. Applied to nuScenes mini, it yields 211,335 triples across 404 keyframes, with 92.5% forward consistency for ego-relative motion labels.
In briefA Fisher-geometry study showing multi-teacher distillation and model merging are closely linked
SummaryThe paper shows that multi-teacher on-policy distillation, to second order within a shared basin of related specialists, equals the Fisher-merging objective, so merging acts as a closed-form stage of distillation. It proposes RGBM, a two-phase method that first performs a task-vector-restricted Fisher merge and then uses a residual spectrum to decide which domains get isolated branch training and how updates are written back. According to the paper, on two open-source consolidation suites the merge alone matches or beats fully trained multi-teacher distillation, and RGBM improves further.
In briefBeyond Behavioral Accuracy: Individual-Level Fidelity of LLM Investor Simulation
SummaryThis study examines whether large language models can faithfully reproduce an individual's trading behavior and decision logic over time. Using a longitudinal paper-trading study with 80 participants and rolling next-day predictions, the authors find that no evaluated LLM reliably beats a simple recent-activity baseline for predicting trades. Buy-sell structure and asset selection also remain hard to reproduce.
In briefInterpretable Face Contrast Modeling for Smartphone Portrait Rendering
SummaryThe paper addresses how smartphone portrait rendering differs across manufacturers and proposes an interpretable framework that splits face contrast into overall facial tonal separation, eye-region contrast, and fine skin micro-contrast. Metrics for each component are validated through pairwise human judgments and correlate more closely with perception than traditional contrast measures. The authors also introduce a mannequin-based dataset of 508 captures from 29 smartphones and give a preliminary check on a smaller real-portrait dataset.
Yujin Cho, Beno\^{i}t Pochon, Jean-Michel Morel et al.
In briefA research agenda for structured financial foundation models
SummaryThe paper proposes a structured-data-first research agenda for modular, domain-specialized financial foundation models linked by a shared representation fabric for reuse across domains and tasks. It argues such models should maintain persistent financial state, capture its dynamics, simulate action-conditioned futures, and stay calibrated across regimes and tail outcomes. The authors also identify research challenges, directions, and capability-oriented evaluation criteria.
Natraj Raman, Saumitra Mishra, Vamsi Potluru et al.
In briefRouterWorld: A Benchmark for Latency-Aware Routing Among Chatbot Language Models
SummaryRouterWorld is a benchmark for choosing which large language model should answer a conversational query, balancing answer quality against response speed. It combines 371k single-turn responses and 76k multi-turn agentic trajectories, evaluated with a standardized LLM-as-a-judge protocol and hardware-based latency annotations. The paper evaluates 13 routing baselines and finds that most routers overemphasize accuracy while neglecting latency under partial observability. Code and data are publicly released.
In briefLearning Task Constraints, Not Timing: Reducing Spurious Coupling in Two-Arm Imitation Learning
SummaryThis paper studies how robots learning two-arm tasks from demonstrations can wrongly rely on one arm's progress to decide when the other should move, even when timing does not matter. The authors propose Constraint-preserving Temporal Recombination (CTR), a training-time data method that keeps each arm's recorded motions but retimes segments so that only task-required cross-arm constraints remain. Applied to Diffusion Policy, Action Chunking with Transformers, and π0.5, CTR improves worst-case success on real-robot initializations and sharply reduces task-constraint violations. Code is said to be forthcoming.](en)
In briefA Canary for Silent Image Classifier Errors: Catching Confident Mistakes
SummaryDeployed vision models often accept high-confidence predictions automatically, so confident errors can spread unnoticed. The paper analyzes a structural blind spot in Maximum Softmax Probability (MSP), the standard failure predictor, and proposes CANARY, a training-free detector that computes per-class kNN distances on PCA residuals of penultimate-layer features and fuses them with MSP. The authors report that CANARY matches or improves performance in 382 of 385 distribution-shift cells across five backbones, with the largest gains on natural shifts such as ImageNet-A/R and ObjectNet.
In briefScopeMem: Latent memory that learns where to revise frozen GUI agents
SummaryThe paper argues that latent memory gains for GUI agents mainly come from revising the model's default actions, not from encoding history. It proposes ScopeMem, whose encoder conditions the memory prefix on the model's initial answer so that it fixes errors while keeping correct decisions. Across AndroidControl, GUI-Odyssey, and zero-shot AITW, it outperforms the evaluated baselines, letting a frozen Qwen3-VL-8B match or surpass some GUI-specialized agents and Qwen3-VL-32B.
In briefReducing reward hacking in preference reward models using evolving penalties
SummaryThe paper introduces Dynalize, a method that mitigates reward hacking during policy training by periodically having a language model analyze training rollouts and write penalties for undesirable behaviors the reward model appears to reward. A judge applies these penalties to every rollout. In clinical counseling and creative writing, training on the reward model alone raised its score but reduced external benchmark performance, whereas Dynalize sustained longer training and improved downstream results.
Afonso Baldo, Ant\'onio Farinhas, Marcos Treviso et al.
In briefPDNet: Pruning and Distillation for Real-Time Worker Activity Recognition in Assembly
SummaryThis paper addresses real-time fine-grained human activity recognition in industrial assembly and proposes PDNet, which combines structured pruning and knowledge distillation for 3D convolutional networks. A 3D-CNN backbone is first trained, then channel-wise structured pruning produces a compact model, and an unpruned teacher is used for distillation to recover lost accuracy. On the MECCANO dataset with an X3D-M backbone, the pruned model without distillation reaches 25.54% Top-1 accuracy, while PDNet reaches 39.46% with a 6.90 MB model size and 3.78 GFLOPs.
In briefHow strength controls change comparisons of interventions across sparse autoencoder codes
SummaryChanging sparse autoencoder codes alters both the direction and magnitude of interventions, so strength must be controlled before comparing effects. The study compares norm matching and output calibration on fixed feature groups in Gemma and Llama. Full-output KL calibration leaves smaller mean gaps from the reference intervention and can serve as a diagnostic of residual disagreement.
In briefSPAR: Shrinking Robot Vision-Language-Action Models and Recovering Their Skill
SummarySPAR prunes the hidden and feed-forward dimensions of a robot model's language backbone, then uses reinforcement learning to recover manipulation performance. Applied to π0.5, it cuts the language backbone from 2.34B to 0.09B parameters and the full model from 3.37B to 1.12B. After RL, average success across four LIBERO suites rises from 84.55% to 93.45%, and inference is 1.98 times faster than the unpruned model.
In briefHow Wet-Biased Data Pipelines Mislead Training and Evaluation of Generative Precipitation Nowcasting
SummaryThis paper examines generative precipitation nowcasting models, which are mostly trained and tested on datasets biased toward wet conditions. The authors show analytically that selecting test cases based on target observations also shifts the conditional distribution the model learns, so matched training and test sampling can hide training bias. Using a fixed satellite-to-radar diffusion model, they show that the choice of test sampling can reverse the calibration conclusion.
Christoph Metzl, Stefan Kapsreiter, Kianusch Vahid Yousefnia et al.
In briefGaussian approximation of smooth functions for logistic classification on curved data manifolds
SummaryThis paper studies Gaussian kernel approximation and regularized logistic classification on a compact smooth manifold without boundary, measured by its intrinsic dimension. The authors build Gaussian approximants directly on the manifold, avoiding second-order saturation while keeping norm control tied to intrinsic dimension. They derive logistic excess-risk rates, and show that clipping predictions gives a faster rate whose exponent matches the minimax optimal squared-loss regression rate for bounded densities.
WhyOvercomes second-order saturation via manifold Bernstein inequality and extension-free Gaussian approximants, with rigorous proofs matching near-minimax rates.
In briefA review of 3D diffusion models for medical scan images
SummaryThis paper gives a comparative overview of 3D diffusion models developed for medical images such as CT, MRI and PET. It lists uses including data augmentation, domain transfer, reconstruction, inpainting and artifact correction, and notes that generation can be conditioned on patient data, reports or masks. The authors also discuss the high memory, compute and training data demands caused by volumetric data, and outline current performance, limitations and future directions.
Frank-Peter Schilling, Daniel Barco, Philipp Denzel et al.
In briefAfriGuard: Safety and Robust Refusal for African Languages
SummaryThis paper introduces AfriGuard, a multilingual dataset covering 10 African languages and English, for studying harmful intent detection, safety-category understanding, and refusal behavior. The authors find that guardrail models underperform general LLMs on safety-intent and category classification, and that models score below 60% F1 on understanding appropriate safety behavior. They also build efficient guardrail models and refusal detectors, and show that supervised fine-tuning substantially improves safety performance.
Israel Abebe Azime, Jesujoba O. Alabi, Iffat Maab et al.
In briefA simple, unified method for bounding regret by gradient variation in online learning
SummaryThe paper proposes a simple, unified method using surrogate losses with a quadratic penalty on decision movement to obtain gradient-variation bounds. It yields an optimistic-to-gradient-variation reduction for non-stationary online learning, matching a known rate. For universal online learning, it gives a simple algorithm matching optimal bounds while relaxing extended smoothness, and it also improves guarantees for online games.
In briefStudy on Learning Expand-and-Sparsify Representations
SummaryThis paper examines whether learning the linear expansion map by gradient descent, rather than using a random one, improves Expand-and-Sparsify (EaS) representations, and how that benefit depends on the expansion width. The authors compare methods across 73 OpenML classification and 48 regression tasks under a fixed protocol. Reported findings include that learning helps most at the narrowest expansions, that learned top-k representations match dense ReLU layers at the tested expansion factors, and that the sparse code performs better when binarized. The permissible sparsity level is also reported to rise as the expansion widens.
In briefVeriCUA: Learning Proactive Verification for Reliable GUI Agent Rewards
SummaryThis paper introduces VeriCUA, a proactive reward agent that verifies computer-use agents' task outcomes by running checks in the live GUI environment. It pairs a trainable Judge, which decides what to check and when to stop, with a frozen Probe that executes the checks, and optimizes the Judge with online reinforcement learning. Trained only on OSWorld, it achieves the highest reward agreement on OSWorld, MobileWorld, and WebArena-Verified, outperforming prompted models and improving downstream agent task success.
WhyProposes learning verification strategy via online RL, beating prompted large models across three benchmarks with fewer probe tokens, showing notable novelty and evidence.
In briefYUTELA: A Scalable Edge-Cloud AIoT System for Multi-Site Video Anomaly Detection
SummaryThis paper presents YUTELA, an edge-server-cloud AIoT system for long-term video anomaly detection across multiple sites, extending the earlier Ancilia framework. It uses NVIDIA Jetson Thor edge devices and TensorRT acceleration, raising throughput with eight cameras from 14.17 to 29.94 FPS. The system was run continuously for 170 hours at four public sites using 31 existing CCTV cameras, where mean end-to-end alert latency ranged from 6.8 to 15.4 seconds.
Shanle Yao, Babak Rahimi Ardabili, Amir Kimiyaie et al.
In briefGeoDiv KAN Solver for Time-Fractional Navier-Stokes Equations on Complex Domains
SummaryThe paper proposes GeoDiv KAN, a geometry-gated, divergence-free Gaussian radial basis function Kolmogorov-Arnold network, for solving time-fractional incompressible Navier-Stokes equations on complex domains. It uses geometry-dependent feature gating and a streamfunction formulation to preserve incompressibility by construction. On square, circular, and star-shaped benchmark domains, it reports higher accuracy than PINN, PIKAN, RBF KAN, and DF RBFKAN. It is further applied to two data-driven CFD cases: unsteady flow past a circular cylinder and turbulent RANS flow through a U-bend.
In briefTraining negatives still matter after calibration for answerability prediction
SummaryThis paper studies whether keeping negative passages, those without answers, when training answerability readouts changes calibrated predictions, ranking, and downstream answer retention. On SQuAD 2.0 and RepLiQA, models trained with negatives achieve lower Brier loss and better AUROC ranking than those trained without them. However, on SQuAD 2.0 they retain more answers to unanswerable questions. The authors do not claim a universal advantage for native training.
In briefResidual-based adaptive refinement for hybrid PINNs on singular elliptic problems
SummaryThis paper studies solving second-order elliptic PDEs with mixed Dirichlet-Neumann conditions on nonconvex or cracked domains, where solutions have singular gradients. The authors use a hybrid variational-collocation physics-informed neural network and add collocation points chosen by residual after each training stage. Tested on five two-dimensional benchmarks, it reduced relative L2 error by up to a factor of 32 compared with uniform point insertion under the same budget.
In briefProgressive multi-scale tumor-state memory for brain tumor segmentation with missing MRI sequences
SummaryThis paper addresses brain tumor segmentation when some MRI sequences are missing, proposing PMTM, a network that segments from any non-empty subset of T1ce, T1, FLAIR, and T2. It keeps a compact tumor state across encoder stages, combining modality evidence via a learned gate to guide residual fusion. Averaged over 15 modality subsets, PMTM reaches Dice scores of 75.31%, 77.54%, and 80.47% on BraTS2018, BraTS2020, and BraTS2023 GLI, higher than RobustSeg, RFNet, and mmFormer in the authors' experiments.
Milad Rostamian, Mohammadreza Mohammadi, and Mohsen Soryani
In briefALIGN: Aligning Condition Estimation with Action Prediction in Advantage-Conditioned Behavior Cloning
SummaryThe paper proposes ALIGN, a framework for vision-language-action robot policies that predicts categorical pairwise conditions directly in the policy's label space and injects them into a DiT action expert via AdaRMS modulation fused with flow time. It is evaluated on mixed-quality real-robot datasets combining standard demonstrations with retry-heavy trajectories for bowl stacking and key hanging. ALIGN reaches 90.0% and 56.7% closed-loop success, versus 56.7% and 33.3% for the strongest AC-BC baseline.
In briefLearning Which Experiences to Keep: Resource-Efficient Self-Improvement for Lifelong Agents
SummaryThis paper studies how agents can improve after deployment when generation, interaction, and memory resources are limited. The authors find that accumulating more experience does not reliably improve decisions, and some existing lifelong-memory methods perform no better than simple retrieval or no memory at all. They propose TokenWorth, which learns query-dependent memory worth from sparse execution feedback to guide retrieval and bounded memory management without updating the backbone model. Across three LifelongAgentBench environments, TokenWorth achieves the highest task success rate while using the fewest output tokens among the evaluated methods.
In briefWhen guidance becomes a guard: stopping unsafe text-to-image prompts early
SummaryText-to-image diffusion models can produce unsafe images, and output filters only act after an image is generated. The authors show that classifier-free guidance exposes strong safety signals in conditional model responses. They propose Conditional Semantic Trajectory Interception (CSTI), which compares conditional and unconditional activations at early checkpoints to detect risky inputs without attack samples or generator changes. CSTI performs best overall among six baselines at a matched 5% false positive rate, and reduces generation latency by over 60% on intercepted paths.
In briefOddin: Latent visual reasoning in unified models via self-distillation
SummaryUnified multimodal models can reason visually by interleaving text and image generation, but producing intermediate images is computationally costly. Oddin replaces these intermediate images with compact latent tokens handled entirely by the understanding branch, reserving the generation branch for the final image. The model is trained through offline self-distillation using final image generation, text distillation, and cross-entropy losses for special tokens. Across seven VBVR-Pro visual reasoning tasks, Oddin outperforms interleaved generation baselines and closed-source vision-language models on average.
In briefTeleASR-Amharic: A Benchmark for Amharic Speech Recognition Under Real Telecom Conditions
SummaryThis paper introduces TeleASR-Amharic, a benchmark for testing Amharic automatic speech recognition under varied acoustic and telecommunication conditions. It covers channel conditions, environmental disturbances, and telecom-domain variations drawn from real telecom recordings. Results show severe telecom degradation sharply reduces accuracy, with the best model's WER rising from 0.23 to 0.79, while scaling data and telecom-aware augmentation improve robustness. The telecom-domain data is released alongside the benchmark.
Dagnachew Mekonnen Marilign, Enrique Molero, Abrham Haile et al.
In briefWhich "correct" emotion label? Acted versus listener-perceived ground truth in speech emotion recognition
SummaryThis paper examines how ground truth is defined in speech emotion recognition, using Emozionalmente, an acted Italian corpus with both the intended emotion and five listener judgements per clip. It compares three label authorities: the expressed condition, the listeners' majority plurality, and the full vote distribution, using frozen XLS-R, HuBERT and WavLM representations. The authors find that none of these labels replaces the others, since expressed labels describe production, pluralities give a single-class decision, and vote distributions capture perceptual uncertainty. They recommend stating which label authority is predicted and using distributional scores when listener disagreement matters.
Marco Giordano, Stefano Giacomelli, Claudia Rinaldi et al.
In briefWenSight-Decision-1: A Model for Direct E-commerce Rule-Based Decisions
SummaryThe paper presents WenSight-Decision-1, a shared model that scores supplied options directly from natural-language rules and case information across six e-commerce operational tasks. It fine-tunes Qwen3.5-9B using reference options as supervision. On 1,440 synthetic cases reserved for final evaluation, accuracy rose from 61.60% for the base model to 92.92%.
Mingwen Liu (WenSight AI Lab), Jun Sun (TooChuang), Sichen Wang (WenSight AI Lab)
In briefRethinking LLM fine-tuning by reparameterizing weight space to keep safety intact during downstream adaptation
SummaryFine-tuning large language models on downstream tasks often weakens their previously aligned safety behavior. The paper proposes WSR-Tune, which builds a safety-conditioned basis and reparameterizes weight matrices so safety information concentrates in a few directions. It freezes these safety-critical directions and updates the remaining ones for adaptation. Experiments reportedly show a better trade-off between downstream performance and safety retention.
In briefR2-GP: Variance-calibrated priors for additive Gaussian process models
SummaryThis paper introduces R2-GP, an R2D2-style prior for additive Gaussian process models that calibrates the explained variance (R²) of Matérn kernel components. The authors note that independent amplitude priors overestimate variance because Matérn function variance depends on both amplitude and length-scale, especially in higher dimensions. R2-GP rescales each component using its length-scale and observed inputs, modelling each component's variance as a share of the function variance induced by a global R² prior. The method is also extended to Hilbert space approximations for large sample sizes. Simulations and real-world case studies report better generalisation and parameter recovery than independent variance priors.
Javier Enrique Aguilar, Soham Mukherjee, David Kohns
In briefPredicting Memory Retrieval Failures in AI Agents from Embedding Geometry
SummaryThis paper tests whether geometric relationships among conversation turns, memory texts, and queries can predict whether an answer will be correct before generation. A logistic classifier using 25 features reaches a pooled AUROC of 0.640, below a question-difficulty baseline of 0.655. The authors conclude the results show associations rather than a reliable deployment diagnostic, since embedding proximity cannot verify fact coverage or separate memory, retrieval, and generation errors.
Iliya Zhechev, Radoslav Ralev, Chris Coleman et al.
In briefRecursive AI model for summarizing long videos into text
SummaryThis paper introduces RVLSM, the first recursive video language summarization model designed for turning long-form videos into text summaries. It splits videos into semantically coherent segments based on scene boundaries and handles each segment with video language model sub-models and external tools such as action recognition and speech recognition. The approach improves over zero-shot video language model baselines on the VideoXum and MixTyVSum benchmarks and is competitive with a state-of-the-art fine-tuned model. Its summaries also contain named entities that match the ground-truth summaries more closely than those of the compared methods.
Alessia Intini, Lorenzo Vaiani, Enrico Chen et al.
In briefEarly Prediction of Ventricular Reentry Using Physiology-Aware Graph Attention
SummaryThis paper proposes a graph attention classifier that predicts sustained ventricular reentry early, using simulated cardiac electrophysiology. Each simulation is a graph built from early nodal voltages and edge attributes such as geometry, stimulation frequency, tissue labels, amplitude, and diastolic interval, computed from only the first two paced activations. On 1000 independent test simulations, it reaches 94.6% accuracy and 0.987 ROC-AUC. Ablation experiments identify diastolic interval as the most critical descriptor.
Ma{\"e}lis Morier (EPIONE), Jairo Rodr{\'i}guez Padilla, Patrick Gallinari (ISIR et al.
In briefHarnessRL: A training harness that co-evolves with the policy for open-ended RL
SummaryThis paper addresses reward failures in rubric-based reinforcement learning for open-ended text generation, identifying non-discriminative rewards and spurious rewards as two failure modes. It proposes HarnessRL, which co-evolves task sampling, rollout guidance, and rubric criteria with the policy, using a proposer and just-in-time critic loop to validate updates. The authors also introduce Rubric-Mix-10k, a five-domain training and evaluation suite covering medical, science, writing, role-play, and instruction-following tasks.
In briefSNO-PI: Few-Label SNOMED CT Entity Linking Using Pre-trained Encoders and Curated Terminology
SummaryThis paper presents SNO-PI, which ranked first on SNOMED International's SNOMED CT Entity Linking Benchmark as of 18 October 2026, with a macro character IoU of 0.5387. The system combines high-recall span proposal, SapBERT candidate retrieval, and listwise ensemble reranking across three BERT-family encoders, pre-trained on unlabelled MIMIC-IV notes and on SNOMED CT synonyms and hierarchy. The authors find that curated knowledge helps most for concepts not seen during training, which nonetheless remain the main source of errors.
In briefGoverning the Rule Space: Ruliology and Anticipatory Regulation of Quantum Technologies
SummaryThe paper argues that regulation should be understood as exploring possible transformations rather than merely catching up with technology. It proposes a Ruliological Framework for Anticipatory Regulation that distinguishes rules, transformations and trajectories to assess alternatives and possible developments. The Quantum Europe Strategy is used as a case study.
In briefConsistency of Entities and Citations in a Deep Research AI Across Repeated Identical Prompts
SummaryThis paper examines how consistent a commercial deep research system is when the same prompt is repeated. Each report is represented by the named entities it mentions, grouped by a custom schema, and the citations it retrieves. Across ten runs of one open-ended cybersecurity prompt, the median pairwise entity overlap was 0.12 and citation overlap was below 0.02, with discovery curves that did not visibly level off.
In briefLocal data augmentation methods for imbalanced binary classification
SummaryThis paper addresses imbalanced binary classification, aiming to maximize recall while limiting false detections in the minority class. The authors compare several rebalancing strategies, from uniform subsampling to local data augmentation, and prove that merely rebalancing class proportions cannot change classification performance, whereas local augmentation does. On synthetic data, subsampling anchor points near the classification boundary combined with a Gaussian augmentation kernel outperforms SMOTE-like procedures, and a diffusion-based kernel performs better when the support's geometric structure matters.
Harold Gu{\'e}neau (SAMM), Alain Celisse (LPP, MODAL) et al.
In briefPolicy-reward co-evolution for code generation and why training reward can mislead
SummaryThis paper tests whether SPARK-style rollout recycling, which trains a model to both generate and evaluate code, transfers to code generation and adds value through execution-grounded reflection. Compared with execution-only GRPO on Qwen2.5-Coder-3B-Instruct, co-evolution raised training performance but did not generalize to held-out problems. The authors describe this as single-adapter policy drift and argue that higher training reward alone is not evidence of generalizable self-improvement.
In briefWhere Rhetorical Distortion Lives: Measuring Influencer Content Across Platforms and Languages
SummaryThis paper studies rhetorical distortion in influencer content, where accurate facts are framed as more significant or alarming than warranted. It measures 13,060 posts from 330 accounts on six platforms using a three-tier pipeline across five dimensions. Platform is more strongly associated with distortion than content category, with X/Twitter scoring highest, and the reply layer carries about 3.5 times the distortion of the tweets it answers.
In briefCALIPER: Adaptive Process Rewards for Spatial Reasoning in Vision Models
SummaryThis paper addresses image-based spatial reasoning, where the stage limiting accuracy shifts during training and answer-only rewards cannot pinpoint failures. The authors propose CALIPER, an adaptive process-reward framework, along with SpatialTrace, a corpus of stage-decomposed spatial reasoning traces. Experiments across eight benchmarks and three backbones report the highest mean score under both one-shot and harness training, with consistent gains over fixed reward allocation.
In briefExpert-level evaluation of English learners' speech on a consumer PC: accuracy, fluency, and prosody
SummaryThe paper presents oollim, a system that runs locally on a consumer computer to evaluate second-language English learners' pronunciation accuracy, fluency, and prosody. It combines pretrained XLS-R, XEUS, and PhoneticXeus representations with reference-text features and ridge regression models trained on several public corpora. On 305 independently held-out recordings rated by three experts, several ridge configurations substantially outperform archived Microsoft Azure Pronunciation Assessment API data on Spearman correlation and quadratic weighted kappa across all three constructs.
Richard Rose (Hankuk University of Foreign Studies)
In briefA simple, general method for top-k selection under a limited sampling budget
SummaryThis paper studies Bayesian top-k identification, where a decision maker must select k items of unknown quality using a fixed sampling budget allocated adaptively, with each sample giving only a noisy quality observation. The goal is to minimize the expected opportunity cost of the selected set. The authors propose allocating samples in batches by minimizing a surrogate objective derived from an upper bound on this expected cost. They prove the method's posterior expected opportunity cost decays at an exponential rate of at least half the best achievable by any sampling rule. Simulations show it is competitive with, and often better than, the best existing algorithm across a range of settings.
In briefLaMER: Language-Guided Multi-Expert Routing for Weakly Supervised Referring Expression Comprehension
SummaryThis paper proposes LaMER, a framework that locates image regions described by natural language without box-level supervision. It dynamically combines four frozen expert encoders through a language-conditioned routing fusion module, with a matching head and multi-objective loss for training. The paper reports improvements over DViN and WeakMCN on RefCOCO, RefCOCO+, and RefCOCOg, with ablations supporting each component.
Yu Yu Chang, Tzu-Hsuan Weng, Youssef Mohamed Amin et al.
In briefOne Fact, Many Places: Redundant Evidence Breaks Single-Gold Evaluation in Long Visual Documents
SummaryThe paper observes that facts in long visual documents often appear in several places, such as charts, text, and summaries, yet benchmarks score answers against a single gold location. The authors annotated 420 questions from MMLongBench-Doc and MMDocRAG with all locations that independently establish each fact, finding frequent and often cross-modal copies. They show that this misjudges retrieval, citation, and masking-based evaluations. The annotations and redundancy-aware metrics are released.
WhyShows single-gold evaluation misjudges redundant evidence in visual documents, with 420 annotated questions and multi-model validation; impact is mainly within document understanding evaluation.
In briefHow graph construction in benchmarks affects multi-hop knowledge-graph navigation
SummaryThis paper audits two multi-hop knowledge-graph question answering benchmarks, PathQuestion and PathQuestion-Large, treating them as graph-navigation tasks. The authors keep questions, answers, and path annotations fixed but swap each task-filtered graph for the larger source knowledge graph, which substantially lowers accuracy against released answers and trajectory fidelity. They conclude that benchmark graph construction is part of the task specification and can materially shape measured results.
Eduin E. Hernandez, Sergio A. Diaz, Luis F. Garcia et al.
In briefBridged Self-Distillation for On-Policy Learning: A unified way to transfer reasoning from various teachers to smaller models
SummaryThis paper proposes Bridged Self-Distillation (BSD), a method for transferring reasoning skills from a teacher model to a smaller student model. The teacher first produces a reasoning trace as extra context, and a copy of the student that sees this context scores the student's own rollouts at the token level. A verifier supplies the outcome signal, which is used to adjust updates on unsuccessful rollouts. The authors report that one training procedure works for shared-tokenizer, cross-tokenizer, and closed API teachers, and that BSD outperforms the base model and scenario-specific baselines.
In briefForagerSim: A Simulator for Testing Map-Based Reasoning in Large, Changing Environments
SummaryThe paper introduces ForagerSim, a simulation platform for embodied tasks that require map-based reasoning, such as robot search and rescue and animal foraging. Its difficulty can be adjusted along perception, scale, topology, dynamics, and task. Experiments on eight example worlds compare reinforcement learning, an engineered robotics stack, and a human player. Learned policies do well on simple 2D-sensing tasks but fall short of humans in larger or more complex environments, where map-based baselines perform well, suggesting map-based reasoning is the bottleneck.
Tom\'a\v{s} Musil, Martin Kot\'ik, Michal Vavre\v{c}ka et al.
In briefWhen Coding Agents Outpace Review: Evaluating PRD-Driven Development in Mature Enterprise Repositories
SummaryThe paper introduces PACE, a benchmark of 43 enterprise tasks with 750 expert-audited Code Points for measuring whether coding agents fully fulfill Product Requirement Documents. Across 11 model–scaffold configurations, most scores fall between 52.1 and 78.1, with incomplete implementation and missing integration identified as the main bottlenecks. The authors report that diagnosis-guided harness refinements raised mean performance from 59.4 to 63.0 using the same model and tasks.
In briefModernizing Colors in Long Legacy Videos at Scale
SummaryThe paper introduces long video color modernization, which updates the color and tonal look of legacy footage while keeping scene content and consistency across frames and shots. It proposes a training-free framework that uses a pretrained text-guided image editing model to modernize grids of keyframes, then propagates shot-specific color transforms through each shot. The method was evaluated on legacy videos spanning five decades, using human judgments of modernness and metrics for fidelity, stability, and quality, plus a new inter-shot consistency measure, SW-IGC. It achieves the highest perceived modernness while remaining competitive on other measures.
In briefAnswer Delivery Under a Shared Token Budget: Stopping and Finalization Study
SummaryThis paper studies a fixed-budget inference policy that combines first-valid-box stopping with a charged finalization cue. On 100 independently audited GSM8K test questions per model, it raised delivered endpoint accuracy over first-box stopping by 8.0 points for Qwen3-4B-Thinking-2507 and 7.0 points for DeepSeek-R1-Distill-Llama-8B. A finalization-only ablation and a fixed scalar reader are used to separate cue recovery from stopping and extraction effects.
In briefBOSS: Token-budgeted online sampling and selection for LLM reasoning
SummaryThis paper introduces BOSS, a training-free policy that decides, under a strict per-instance token cap, whether an LLM should stop, generate another candidate answer, or spend remaining tokens judging candidates already produced. It allocates generation and adjudication online from one shared token budget, checking worst-case costs before each call. Across two Llama-3.1 models and four reasoning benchmarks, BOSS achieves the highest accuracy in 22 of 24 model-task-budget settings.
In briefChoosing the top k with human raters: adaptively mixing ratings and pairwise comparisons
SummaryThis paper studies how to identify the k best items from a large pool, such as job applicants, LLMs, or product ideas, using a limited budget of human evaluations. It allows both ratings and pairwise comparisons, using ratings for a coarse early screen and comparisons to refine the top ranks. The authors propose an adaptive algorithm whose main allocation policy is forward-looking, and they prove it is asymptotically consistent. Late in the process it switches to a myopic policy. In simulations it outperforms the limited existing benchmarks.
In briefCTLite: A lightweight model for irregular multivariate time series forecasting using reusable summaries
SummaryThis paper introduces CTLite, a lightweight framework for forecasting irregular, multivariate time series, which are sparse and asynchronous. It separates historical summarisation, done by a parameter-free continuous-time encoder, from the learned prediction step, which combines local extrapolation with a low-rank cross-variable correction. On four real-world benchmarks, CTLite reports the best or second-best average MSE and MAE using only 88 to 2,537 trainable parameters. The authors also state that reusing historical summaries reduces inference latency and total runtime.
In briefMultiCard: Deep learning for estimating row counts across multiple attributes in databases
SummaryThis paper presents MultiCard, a learned model that estimates query result sizes (cardinality) for database optimization. It handles both string LIKE patterns and numerical range filters in one end-to-end architecture, combining a Transformer and LSTM for string patterns and a numerical encoder that uses catalog histograms. On mixed multi-attribute queries, the paper reports up to a 16-fold reduction in PostgreSQL's mean Q-error with inference latency below 5 ms.
Mobina Koochaknia, Maryam Ramezani, and Mohammad Hasan Shafaat
In briefAn Empirical Study of Parameter-Efficient Fine-Tuning for Mixture-of-Experts LLMs
SummaryThis paper systematically studies how to apply parameter-efficient fine-tuning (PEFT) to mixture-of-experts (MoE) large language models. Across three MoE backbones, it examines design choices such as which modules to adapt, the granularity of expert adaptation, the PEFT method, and the load-balancing objective. The authors find that adapting experts alone yields the highest downstream performance, that LoRA and DoRA together are strongest, and that no nonzero load-balancing coefficient Pareto-dominates the zero baseline.
In briefAutonomous AI for mining productivity: a multi-agent approach to operational value
SummaryThe paper proposes EI2ROC, a three-tier multi-agent architecture for mining remote operation centres, combining a large language model orchestrator with monitoring and production-capacity prediction agents. It reviews agentic evaluation along six dimensions and proposes a multi-dimensional benchmarking framework. The authors state which components are implemented in the prototype, which belong to the reference design, and which remain future work, and report preliminary cost and latency observations from an operating site.
N Guerin (CGS i3), A Dehoux (ERAMET), M Nakhla (CGS i3)
In briefLong-wave infrared polarimetric image fusion to improve vehicle detection in remote sensing
SummaryThis paper addresses blurred vehicle contours and low contrast in infrared remote sensing images by proposing a Mamba-based model that fuses long-wave infrared total intensity with degree of linear polarization. It uses bidirectional state space scanning and prior-guided losses to reduce noise while keeping background structure and polarization features. The authors also release AIR-LWIR-Pol, a dataset of 1,901 registered image pairs. Using YOLO11, vehicle detection reaches 55.0% mAP50 and 29.4% mAP50-95, and the method outperforms comparison methods on four fusion metrics.
In briefGAIDA: Generative AI Tool for Human-Reviewed Document Annotation
SummaryThis paper proposes a modular human-in-the-loop architecture that combines OCR extraction, LLM-based semantic pre-annotation, and human validation in one editable workflow, and presents GAIDA as its reference prototype. On the FUNSD and WildReceipt benchmarks, OCR assistance cut region-creation and transcription time, and reviewing LLM pre-annotations reduced semantic-labeling time, even though the pre-annotations were imperfect. The authors also used the same implementation to build and openly release AraReceipt, a dataset of 100 manually annotated Arabic receipt images.
Islam Mesabah, Ali Elhosary, Priyabanta Sandulu et al.
In briefPlanning for Embodied Agents When Cooperation Needs Are Hidden
SummaryThis paper introduces HiddenCoop, a benchmark for evaluating agent cooperation in tasks where the need to cooperate only becomes apparent after agents observe and interact with task-related targets. It proposes Discovery-Probability Graph Planning (DPGP), which uses a graph encoding a probability distribution over target numbers and locations to support information exchange among agents. A search algorithm then derives action plans that minimize expected cost. The authors report success rates of 99.3% on C-WAH and 58.3% on TDW-MAT, along with reduced execution time and LLM token usage.
In briefHGASurv: Hierarchical Multimodal Model for Patient Survival Prediction
SummaryHGASurv is a model that predicts patient survival risk by combining pathology whole-slide images, genomic profiles, and clinical variables. It keeps the spatial layout of image patches at several scales and uses clinical data to build patient-similarity graphs for integrating information. The authors report the strongest overall C-index across three TCGA cohorts, with ablation studies supporting both main components.
In briefPhyPro-FM: A Physics- and Prompt-Guided Graph Model for Power Grid Contingency Analysis
SummaryThis paper addresses the computational and scalability challenges of N-1/N-2 contingency analysis in power systems with high renewable penetration and changing topologies. It proposes PhyPro-FM, a graph foundation model combining graph neural networks, a learnable global query mechanism, and prompt-guided learning that uses historical snapshots as few-shot prompts. The training emphasizes branch power flow errors over nodal voltage errors, and inference includes residual-based correction and special handling of singular branches. The authors report that PhyPro-FM outperforms state-of-the-art methods in accuracy and generalization.
In briefTraining-time logit averaging sets a new state of the art for non-invasive speech BCI
SummaryThis paper systematically compares six strategies for aggregating repeated trials in non-invasive neural decoding. Training-time logit averaging, the simplest method proposed, achieves a new state of the art on the phoneme classification task of the 2025 PNPL competition. The authors also draw a theoretical link between logit averaging and joint ensemble training.
Miran \"Ozdogan, Gereon Elvers, Michael Bronstein et al.
In briefCARRS: A Real-Time Routing System for Emergency Vehicles That Respects Vehicle Constraints
SummaryThis paper presents CARRS, a software architecture that plans routes for emergency vehicles such as police cars, fire engines, and ambulances in changing urban traffic while respecting vehicle-specific constraints. It combines classical pathfinding algorithms (A*, Dijkstra, and Bidirectional A*) with Constraint Satisfaction Problem methods. The system was evaluated using OpenStreetMap data and simulated dynamic traffic, where it generated optimal routes for relevant subgraphs in under 35 seconds, rerouted in under 20 seconds, and adhered to vehicle-specific constraints more than 90% of the time.
In briefVibeTraining-Bench: A Benchmark for Co-Designing Data, Trainer, and Harness in Iterative Improvement
SummaryThe paper introduces VibeTraining-Bench, a benchmark testing whether research agents can jointly optimize training data, the trainer, and the inference harness across tasks like terminal operation, software engineering, tool use, and math reasoning. Agents start from the same initial system under a fixed budget and iteratively refine components using validation feedback before final test evaluation. The authors report that frontier agents substantially improve target systems, such as raising SWE-bench Verified performance from 32.9% to 53.7%, but cumulative gains remain an open challenge.
In briefA generic on-board AI chain for Earth observation satellites, from model deployment to mission orchestration.
SummaryThis paper presents a generic on-board AI processing chain for Earth observation satellites, using wildfire segmentation as the main example alongside ship detection and cloud segmentation. The AI model is tuned, quantized to 8-bit integers, and ported to an AMD Versal VCK190, and a reusable tiler-inference-mosaicer workflow converts tile-level results into scene-level outputs. Each step is packaged as a unikernel and orchestrated on board by the ORCHIDE cloud-native orchestrator, and the full chain is characterized end to end on a representative test bench.
Michel-Fran{\c c}ois Foulon (TAS), Liz Ramos Medina (TAS), Thomas Goudemant et al.
In briefLarge-Scale Biomedical RAG: How Corpus Size, Document Format, and Reranking Affect Accuracy and Cost
SummaryThis paper evaluates biomedical retrieval-augmented generation systems as the document collection grows, comparing abstract, flat full-text, and hierarchical full-text representations. It also measures retrieval latency and index storage to assess operational cost. Hierarchical representation matched or outperformed flat full-text retrieval in most end-to-end comparisons while requiring 67% fewer indexed units.
Adri\'an S\'anchez-Freire, Gonzalo P\'erez-G\'omez, Aitziber Atutxa et al.
In briefCIS-Transolver: A Layer-Crossing Neural Operator for Finer Car Airflow Prediction
SummaryThis paper proposes CIS-Transolver, a neural operator that rapidly predicts automotive aerodynamic flows and aims to better capture fine-scale flow changes near complex shapes. It extracts fine-scale information via local multiscale analysis, encodes it as geometric moment states, and propagates and updates these states across layers. Evaluated on ShapeNet-Car, AhmedML, and DrivAerNet++, it achieves the lowest mean relative L2 error on all five metrics, averaging an 8.72% reduction versus the best baseline for each metric.
In briefTyped decision models under label shift: calibration and training effects
SummaryThis paper examines whether typed decision models (TDMs), which output probabilities over caller-defined options, stay well calibrated when class frequencies change between development and deployment. The authors compare three TDMs, their base models and two trained classifiers on five public datasets under simulated prevalence shift. They find that decision training is not uniformly beneficial, and that the TDMs are not uniquely fragile under shift, with their main calibration weakness already present without shift.
Abdul Rahman Iqbal, Muhammad Fawad Nasim, Amir Saleem
In briefTriScore: Post-Hoc Out-of-Distribution Detection Combining Energy, Boundary Probes, and Transform Consistency
SummaryThe paper proposes TriScore, a post-hoc method for detecting out-of-distribution inputs using a frozen classifier. It combines three complementary signals from the same backbone: an energy prior, a boundary-fragility score, and a transform-consistency score, fused by an input-dependent gate. Under a non-adaptive adversarial-shift protocol, TriScore achieves the best mean AUROC among 15 baselines while remaining competitive on standard benchmarks.
WhyFuses energy, boundary-fragility, and transform-consistency signals via an ID-residual gate, achieving best mean AUROC over 15 baselines on four backbones.
In briefMCDA: Multi-view graph contrastive learning with dynamic topology augmentation for brain fMRI analysis
SummaryThis paper proposes MCDA, a framework for analyzing resting-state fMRI to diagnose brain disorders. It uses a module that iteratively refines augmented graph topology, along with self-aware and cross-sample augmentation, and a multi-view contrastive learning scheme that links the original network with its augmented views. The authors report that MCDA outperforms state-of-the-art methods on two public rs-fMRI datasets with 1,236 subjects. They also examine discriminative functional connectivity patterns as potential biomarkers.
In briefTesting whether a spiking MoveNet can run on future onboard satellite pose estimation
SummaryThis paper evaluates S-MoveNet-Lightning, a spiking neural network version of MoveNet-Lightning designed for autonomous spacecraft, using the RAPTOR dataset. It compares performance and computational cost against the standard non-spiking model. On average, the spiking version needs 2.2 times fewer multiply-accumulate operations than the non-spiking solution on FPGA platforms with the AMD DPU. The authors present these results as insight into the feasibility of event-driven embedded solutions for space applications.
Jonathan Courtois, Christophe Marabotto, Benoit Miramond et al.
In briefStrucChunk: Structure-Aware Retrieval for Improving Legal Document RAG
SummaryThis paper proposes StrucChunk, a chunking and retrieval method for hierarchically structured legal documents that treats them as graphs of provisions, preserves clause boundaries, and resolves cross-references before embedding. It combines BM25 with BGE-M3 using Reciprocal Rank Fusion. The authors evaluate on five legal corpora spanning three legal traditions and two languages, reporting recall gains over baselines. They release code and 5,256 evaluation queries.
WhyProposes structure-aware chunking and hybrid retrieval for legal documents, validated across five corpora and languages with strong baseline comparisons; moderate novelty, useful to legal RAG.
Surya Deep Singh, Ashutosh Tripathi, Gaurav Deep Singh
In briefCARNet: A Unified Network for Fine-Grained Multi-Drone Vehicle Detection
SummaryThis paper introduces CARNet, which jointly optimizes cross-view target association and recognition fusion through a shared target-pair representation, instead of a two-stage pipeline. It adds spatial cues for densely packed, visually similar targets and uses two causal-inspired intervention strategies to improve association and fusion robustness. The authors also present MDMT-CD, an extension of the MDMT dataset with fine-grained vehicle category labels, and report that CARNet outperforms existing methods.
In briefSelective Transfer Through Sequence Distillation: From Benchmark Knowledge to Backdoor Behaviour
SummaryThis paper studies which hidden properties of a teacher model survive knowledge distillation into a student model. It compares benchmark-associated knowledge with refusal backdoors across four distillation channels. Benchmark knowledge transfers even when students learn only from teacher-generated number sequences, whereas refusal backdoors almost fail to transfer through sequence distillation but transfer substantially through logit-based distillation. The authors conclude that distillation is selective, with the criterion being the structure of the teacher's property rather than the data the student sees.
Jonibek Mansurov, Sholpan Bolatzhanova, Filbert Aurelian Tjiaranta et al.
In briefThe Scheduler as a Hidden Retriever: Asynchrony in Search Agents' Evidence Gathering
SummaryThis paper shows that, in parallel asynchronous search by language-model agents, the timing of result delivery changes later search decisions and what evidence gets acquired, even when the initial evidence is identical. The authors propose Confluent Evidence Acquisition (CEA), which releases local reasoning once each subquestion's evidence is complete and allocates shared search independently of completion order. On three benchmarks, CEA brings acquisition disagreement to within 1.0 percentage point of repeat variation while matching the accuracy of complete-input controls.
WhyProposes confluent evidence acquisition with a joint invariance proof, showing scheduling timing affects search decisions across two model families and three benchmarks.
In briefCAST: Mold-Openable Generative Shape Design Using Paired Depth Images
SummaryThe paper presents CAST, which recasts 3D shape generation under a chosen mold-opening direction as 2D image generation using paired top and bottom depth images, so undercuts in that direction are excluded by construction. Diffusion models generate candidate shapes, and an FEA-guided optimizer refines their boundaries. On 211 SimJEB bracket designs, refinement reduced compliance in 193 of 200 generated candidates, with a median reduction of 36.8 percent in the optimization FEA model.
In briefAnswering while typing: streaming generation to cut wait time for answers
SummaryThis paper focuses on the time users wait from submitting a request to receiving a usable response, rather than on decoding speed itself. The authors propose streaming generation, in which the model starts producing a response from an incomplete request and revises it as the remaining input arrives, so little computation remains after submission. The regime is evaluated on autoregressive and diffusion language models, and the paper characterises when it reduces user-visible latency and analyses the mechanism behind it.
Nikita Dragunov, Matvey Mikhalchuk, Anton Razzhigaev
In briefRobot photography system using a vision-language-action model that plans camera moves for aesthetic photos
SummaryThis paper presents RoboPhoto-VLA, a unified vision-language-action framework that jointly handles perception, reasoning, and camera control for autonomous robotic photography. It uses PhotoCoT for structured action planning and Aesthetic Direct Preference Optimization (ADPO) to align camera behavior with human aesthetic preferences. The authors also built OpenPhotography, a robotic photography dataset with language-guided demonstrations. Real-world experiments report photos better aligned with human preferences and generalization to unseen subjects, viewpoints, poses, and complex instructions.
In briefToward Online Robust Zero-Sum Markov Games with Function Approximation
SummaryThis paper studies online robust zero-sum Markov games, where a learner must recover a worst-case minimax Bellman operator from nominal interaction data while solving the game. The authors propose an optimistic framework combining dual robust Bellman fitting, optimistic minimax planning, backward-validated confidence sets, and exploiter-mix data collection. They give conditional theorems for general and projected function approximation and closed results in structured settings. Small numerical experiments illustrate the predicted robustness and return-gap trade-off.
In briefLearning how SOCP solvers work step by step with graph neural networks
SummaryThis paper introduces SOCPNet, a recursive graph neural network for solving second-order cone programs, which are convex optimization problems. Instead of predicting the answer in one pass, it learns from the intermediate iterates of classical solvers and repeatedly refines its estimate over a variable number of steps. On multiple datasets, it consistently reduces final solution error compared with a SOCP-GNN baseline. The authors also show that trained models can warm-start an interior-point solver, cutting iterations without losing accuracy.
In briefTrust Your Targets: Is Target Propagation Block-Diagonal Gauss-Newton?
SummaryThis paper rederives target propagation, a backpropagation alternative that uses local targets to carry the learning signal, from a trust region under an implicit norm. It introduces CGTP, which solves layer-local subproblems in parallel with conjugate gradients, and shows KFAC arises as a special case. On MNIST and Fashion-MNIST MLPs, the authors report accuracy matching ADAM and KFAC with far fewer outer iterations, and stability across depth.
Giuseppe Castiglione, Christopher Buckley, Ivor Simpson
In briefFrom Cause to Effect: Generating Missing Events Linking Causes and Outcomes in Video
SummaryThis paper introduces Causal Chain Generation, a task where video-language models must produce the intermediate events connecting a cause and an effect. It proposes two metrics, SODA-Embedding and Causal-QA, and two test sets, SF20K-CC and CMD-CC. The authors find that current models often hallucinate or omit key events, and report substantial gains after fine-tuning with GRPO using a composite reward.
In briefWhy exact-arrival multi-agent path finding needs bidirectional cellular machines
SummaryThis paper studies exact-arrival multi-agent path finding on ranked directed acyclic graphs, where each cell must output its own bit of a valid joint path mask using only local communication. The authors show that on an explicit hard topology, no universally correct local-only architecture can give up either temporal orientation. With k one-bit global flags, the impossibility becomes a trade-off between rounds and bandwidth. The paper reports no benchmarks.
In briefTIM: A Model That Separates Time and Disease Features to Generate Longitudinal Radiology Reports
SummaryThis paper proposes TIM, a two-stage framework that generates radiology reports from longitudinal medical images by separating spatial disease features from temporal progression. The second stage iteratively corrects diagnostic inconsistencies between prior and current reports through error-sensitive mutual feedback. The authors report that TIM outperforms single-image and longitudinal baselines on the Longitudinal-MIMIC dataset across language and clinical metrics.
In briefExtracting Safety Relations from Aviation Accident Reports Under Sparse Conditions
SummaryThis paper studies how to extract safety-related relations between concepts from NTSB aviation accident narratives. Only 1.18% of about 21,000 candidate concept pairs are true result or contributing relations, making the task highly imbalanced. The authors compare a TF-IDF linear SVM with a RoBERTa classifier and find that performance drops sharply under exhaustive evaluation, with many false positives. They conclude that known concept endpoints do not guarantee reliable relation extraction, and controlling false positives is essential.
Aziida Nanyonga, Mustafa Tahsin Yilmaz, Saad Ezzini et al.
In briefA pitfall in evaluating models that predict gene responses to genetic perturbations
SummaryThis paper examines how evaluation choices affect gene perturbation models, comparing eight methods on six single-gene perturbation datasets under a unified protocol. It finds that repeating a single predicted expression profile across cell rows can inflate significance estimates, producing excessively large sets of predicted significant genes. After accounting for this artifact, the authors report that a random forest using Gene Ontology features achieves the best overall results.
George Panagopoulos, Saman Omidi, Tsz Pan Tong et al.
In briefVHDLBench: Benchmarking and Fine-Tuning Language Models for VHDL Chip Design
SummaryThis paper studies using large language models to help generate VHDL, a hardware description language used in digital circuit design. The authors built a dataset of over 18,500 syntax-verified VHDL examples and fine-tuned five 7B-parameter models with QLoRA. Models were checked through syntax and synthesis validation, functional simulation, and FPGA prototyping on Zynq SoCs. Fine-tuned models reached up to 92% syntax correctness and 73% functional accuracy, and generated designs ran on real hardware.
In briefExploration via Activation-Space Steering for Verifiable-Reward RL
SummaryThis paper addresses a limitation of reinforcement learning with verifiable rewards (RLVR), where repeated sampling from a policy tends to concentrate on similar reasoning paths. The authors propose Activation-Space Exploration (ASE), which clusters activation differences between correct and incorrect responses to build a bank of latent exploration vectors applied to later rollouts. On Qwen3-1.7B-Base and Qwen3-8B-Base trained on DAPO-Math-17k, ASE improves average Mean@8 over GRPO across six math reasoning benchmarks by 4.87 and 1.87 percentage points, respectively, under the same rollout budget.
In briefDynamics-Aware Sparse Attention: A Paper on Faster Autoregressive Video Generation
SummaryThis paper proposes Dynamics-Aware Sparse Attention (DASA), a framework that makes autoregressive diffusion transformers for long video generation more efficient. It separates video content into dynamic and static regions, focusing computation on motion-driven areas and compressing the historical KV cache to keep only dynamic features. Experiments on VBench report a 2.16x speedup in generation while preserving high visual fidelity.
In briefREFLEX: RNA Ensemble Generation Using a Flexibility-Calibrated Stochastic Bridge
SummaryThis paper introduces REFLEX, a model that generates ensembles of RNA backbone structures from the input sequence alone. It uses experimental B-factors to calibrate residue-wise noise widths and a learned flexibility condition to guide recovery of target conformers. On a multi-conformer RNA benchmark, REFLEX reports the best precision-recall trade-off for ensemble coverage across evaluation thresholds.
In briefAudioScorer: Learning which audio tokens matter for a given question in large audio models
SummaryLarge audio language models process long audio through dense token sequences, which is computationally costly. Existing pruning methods rely on heuristics such as attention or temporal sampling rather than learning which tokens matter for a specific question. The paper proposes AudioScorer, a lightweight model that scores audio tokens per question, trained by distilling answer-gradient saliency rankings from a frozen model plus answer supervision. At 10% token retention, it achieves the best or tied-best result on 7 of 10 tasks.
In briefLOCU: Löwdin-orthogonalized constraint updates for multi-constraint policy optimization
SummaryThis paper studies policy optimization with multiple constraints in constrained Markov decision processes, where interactions between constraint gradients can produce ill-conditioned, numerically unstable update directions. The authors propose LOCU, which applies symmetric Löwdin orthogonalization to the natural gradients of the reward and constraints under the Fisher-Rao metric, combined with Fisher-geometric screening and near-collinear compression. Experiments reported in the paper show LOCU stays stable under high constraint coupling, narrow feasible regions, and infeasible starts, and generalizes across different cost critic architectures.
In briefXGLASS: Using X-ray, Gaze, and Language to Rank CT Slices by Diagnostic Relevance
SummaryThis paper presents XGLASS, a method that transfers radiologist expertise from 2D X-rays to 3D CT scans to prioritize diagnostically relevant slices for each disease. It aligns gaze-based importance maps from X-rays with coronal CT projections and predicts an ordering of axial slices. The approach outperforms 2D and 3D-native baselines in budgeted retrieval and can serve as a lightweight preselector for volumetric CT analysis.
In briefSEAM: Spectral embedding alignment for low-resource speech synthesis
SummaryThis paper introduces SEAM, a method that builds embeddings for unseen tokens from known subtokens without needing audio, to expand the vocabulary of conditional text-to-speech models. SEAM uses a spectral filter to keep phonetic order and closed-form ridge regression to align new embeddings with the pretrained conditioning space. The authors report that across three languages and two TTS architectures, it approaches or matches baseline performance with up to four times less training data.
Walaa Othman, Baher Mohammad, Stamatios Lefkimmiatis et al.
