In briefKnee3DVLM: A dual-sequence vision-language model for full knee MRI assessment
SummaryThis paper presents Knee3DVLM, a vision-language model that uses full-volume DESS and fluid-sensitive TSE knee MRI to predict 57 binary diagnostic targets based on the MOAKS score for structured reporting. On 1,074 held-out examinations, the fused DESS-TSE model outperformed the single-sequence configurations on accuracy, balanced accuracy, and ROC-AUC metrics. In a multiclass analysis across five pathology categories, it was numerically higher than the strongest reported 3DReasonKnee result.
Maryam Baizhigitova, Andrew Seohwan Yu, Po-Hao Chen et al.
In briefMetaLearnNCA: Few-shot learning using interacting neural cellular automata
SummaryThis paper proposes MetaLearnNCA, a few-shot meta-learning framework built from coupled Neural Cellular Automata that adapts to new tasks without computing analytical gradients at inference. It splits adaptation into an Active-NCA that performs task inference over a 2D spatial memory grid, and a Meta-NCA that updates this grid by diffusing spatial error residuals across local neighborhoods. The authors report 96.12% in-distribution accuracy on Omniglot and out-of-distribution gains on MNIST, KMNIST, and Fashion-MNIST across 1-, 5-, and 10-shot settings over 10 seeds.
In briefLearning PDE solution operators with variable starting states using Latent Dynamics Networks
SummaryThis paper extends the Latent Dynamics Network (LDNet), a data-driven surrogate for simulating physical systems described by partial differential equations, so it can handle varying initial conditions. The initial latent state is inferred from a few early observations, using either an auto-decoding approach or meta-learning. The authors report accurate results on advection-diffusion, fluid dynamics, and solid mechanics problems, with meta-learning speeding up inference and the decoder producing high-resolution fields even when trained on spatially subsampled data.
Stefano Maria Pizzamiglio, Stefano Pagani, Francesco Regazzoni
In briefUNREAL: One Model for Both Retrieval and Long-Context Inference
SummaryThe paper proposes UNREAL, a framework that uses a frozen LLM's internal representations to select evidence, both for searching a large corpus and for long-context inference. It adds fewer than 500K trainable parameters and leaves the backbone unchanged. On a Wikipedia index, the UNREAL backbones outperform state-of-the-art retriever-reranker systems in recall on HotpotQA, 2WikiMultiHopQA, and MuSiQue.
WhyA model-internal evidence selection unifies retrieval and long-context inference, adding under 500K parameters; large-scale Wikipedia and long-context experiments show broad impact.
In briefAgentic AutoRAG: Optimizing RAG Pipelines with Reasoning-Driven Agents
SummaryThis paper introduces Agentic AutoRAG, an LLM-agent optimizer that tunes retrieval-augmented generation pipeline settings for both accuracy and cost. After each trial, it attributes failed questions to either retrieval or generation and uses a knowledge base of model rankings and pricing to choose the next configuration. On three multi-hop QA benchmarks it achieves higher LLM-judge accuracy than all compared baselines, and on a real-world healthcare corpus in cost-aware mode it reaches 77% median exam accuracy at about 58% of the strongest baseline's cost per query.
Lasse B. Strand, Robert Jakob, Kevin O'Sullivan et al.
In briefRethinking cross-tokenizer on-policy distillation: from alignment coverage to supervision reliability
SummaryThis paper studies on-policy distillation, where a student learns from its own generations using teacher feedback, when teacher and student use different tokenizers. Across three teacher-student pairs on math reasoning and code generation, strict 1:1 alignment already covers most student tokens. Restricting reverse KL to a student-selected top-16 subset of the shared vocabulary at strictly aligned positions matches full shared-vocabulary performance. Adding span-level MSE supervision in mismatched groups increases coverage but reduces accuracy, and the authors argue that supervision reliability matters more than broader alignment coverage.
In briefAssemState: A zero-shot method for furniture assembly guided by manuals and physical state
SummaryAssemState is a zero-shot framework that guides furniture assembly using both assembly manuals and physical state. It splits manual pages into single-part operations to recover an assembly tree, then iteratively refines 6D poses using after-state feedback and simulation-based release tests. Compared with the strongest prior baseline, it improves assembly-tree recovery metrics and reduces mean Chamfer distance, though multimodal LLMs still struggle with spatial relationship reasoning.
In briefHuC-VideoMAE: Human-centric video masked autoencoding trained on synthetic data
SummaryThis paper studies self-supervised pretraining of video transformers on synthetic human-motion videos generated from motion-capture data, as an ethical alternative to web-crawled datasets like Kinetics-700 that lack subject consent. The authors find standard VideoMAE masking performs poorly on synthetic data and propose a human-centric masking scheme using body keypoints and person bounding boxes. On NTU RGB+D cross-view-subject evaluation, the method closes 49% of the gap to Kinetics pretraining without using any real frames during pretraining.
Ricardo Pizarro, Roberto Valle, Jos\'e M. Buenaposada et al.
In briefEMHO: Improving Embodied AI Agents by Refining Their Control Harness from Experience
SummaryThe paper proposes EMHO, a self-evolving framework that keeps the embodied model frozen and revises the agent's harness, which governs planning, context, and tool use, using execution trajectories and harness history. It also introduces EMHO-Merge to optimize one shared harness across multiple subtasks. Evaluated on EmbodiedBench navigation and manipulation tasks, EMHO improves task success for both Qwen 9B and 27B models.
In briefNeMo-DCR: Bit-exact delta-compressed weight sync for trillion-parameter agentic RL
SummaryThe paper presents NeMo-DCR, a method for syncing updated model weights between training and rollout clusters by sending only the changed values while keeping the received bits identical to a full transfer. At 3% and 5% change rates, it refits 30B to 1T parameter models 12 to 40 times faster than a full-checkpoint transport baseline. A 1T model refit at 3% via a relay tree drops from 87.5 minutes to 150 seconds.
In briefSymmetry-Aware Feature Learning: A Polynomial Sample-Efficiency Gap for Multi-Index Models
SummaryThis paper studies learning in multi-index models with high-dimensional Gaussian inputs whose teacher directions form a cyclic symmetry orbit. It compares symmetry-tied convolutional networks, untied networks trained with full-group data augmentation, and untied networks without symmetry access. The first two reach weak directional recovery with about d^(p-1) samples, while the symmetry-agnostic learner needs about r times d^(p-1), a polynomial gap.
In briefDeformable CT-ultrasound alignment that accounts for anatomy under probe-induced tissue deformation
SummaryThis paper presents a deformable registration framework that aligns ultrasound with preoperative CT, addressing tissue deformation caused by probe contact. Rigid alignment is first obtained with robot-assisted optical tracking, then a per-frame sinusoidal implicit neural representation estimates the deformation, regularized by CT-derived tissue stiffness and probe-contact and fan-geometry constraints. The method improves alignment by 17% over rigid initialisation and outperforms classical deformable baselines while keeping topological folding near zero.
In briefFly Brain Wiring Inspires a Compact Vision Model That Rivals ResNet18
SummaryThe paper introduces ConnectomeX and FlyVision, a vision architecture built on the Drosophila visual connectome, keeping parallel ON/OFF pathways, recurrence, and graph interaction. FlyVision is tested on MNIST, CIFAR-10, ImageNet-1K, skin-disease classification, and chest radiography, staying within three points of ResNet18 on ImageNet-1K and skin disease with far fewer parameters, and exceeding it on chest radiography. The authors also extend it to brain-age estimation from MRI (BrainAGE), reporting a mean absolute error of 5.98 years on 433 held-out scans.
In briefUnderstanding image bitstreams directly for privacy-friendly AIoT devices
SummaryThis paper introduces Image Bitstream Fine-grained Understanding (IBFU), which classifies images and generates detailed descriptions directly from encoded byte sequences without full pixel decoding. The authors propose the BFG foundation model, combining a bitstream semantic encoder and an autoregressive description generator. They also build CFU-D, a dataset with intact and corrupted bitstreams. On Stanford Dogs Caption, BFG's average CIDEr score changes only slightly (0.6339 to 0.6077) under corruption, while several vision-language models degrade severely.
In briefKnowing When Not to Answer: Generalizing Signals of Missing Information Across Domains and Dialogue Turns
SummaryThis paper examines whether large language models can recognize when a question cannot be answered from the information given, including during multi-turn dialogue. The authors introduce a turn-labeled benchmark of 423 conversations with 1,661 labeled turn-states and an evaluation setup with a simulated user, testing six datasets and six open-weight models. Probes transfer reliably between datasets sharing missing information, but probes for epistemic unknowns transfer poorly to math. Single-turn probes fail to detect when a conversation becomes answerable, and a gate based on the probe does not reliably outperform vanilla generation or prompted consolidation.
Jerzy Kami\'nski, Ilya Galyukshev, Artem Kuznetsov et al.
In briefSSR: A sparse segment reduction method that speeds up ternary matrix multiplication
SummaryThis paper proposes Sparse Segment Reduction (SSR), a ternary matrix multiplication method for accelerating inference in ternary large language models and general ternary weight networks. SSR uses a dedicated ternary data format and an algorithm that exploits sparsity through computation trees that scale with sparsity. According to the paper, SSR achieves 2.1-11.3x speedup over RSR++ on ternary GEMM at 45-95% sparsity, and 3.5-6.3x end-to-end speedup with 4.9% memory saving on Llama-3 1B inference.
Adeline Pittet, Shien Zhu, Val\'erie Verdan et al.
In briefVETTA: Jointly Assigning Credit at Turn and Token Levels for Multi-Turn LLM Agents
SummaryMulti-turn LLM agents receive sparse task feedback over several interactions and generate responses token by token, making it hard to tell which responses and generation decisions mattered. VETTA learns turn-level and token-level values through separate heads on a shared lightweight critic, then combines them for PPO updates. The critic keeps only early Transformer blocks from the actor's pretrained checkpoint to reduce value-learning cost. Experiments on four agent benchmarks report strong success rates with substantially lower critic-side computation.
WhyJointly learns turn- and token-level credit with a shared lightweight critic, a substantive combination; validated across four agent benchmarks with strong results.
In briefGeoPID: Splitting and Steering Visual Information in Vision-Language Models
SummaryThis paper presents GeoPID, a training-free framework that uses geometry to analyze how vision-language models combine visual and textual information, splitting it into redundant, modality-unique, and synergistic parts. Across 22 models and 14 benchmarks, the authors find that correct answers show stronger vision-unique components when questions strongly require visual grounding. Based on this, they propose amplifying visual representations along the vision-unique subspace during inference, which raises average relative accuracy by 7.63% without updating any model parameters.
In briefAtom-JEPA: a self-supervised model for predicting properties of 3D atomic structures
SummaryThis paper introduces Atom-JEPA, a self-supervised pretraining framework that learns latent representations from unlabeled 3D atomic structures. It uses complementary atom-level and substructure-level objectives inspired by joint-embedding predictive architectures. The model is pretrained on large molecular and crystalline datasets and fine-tuned on various property prediction tasks. It reports state-of-the-art results on molecular ADMET and quantum-chemical property prediction, and strong competitiveness on crystalline material properties.
Kasper Helverskov Petersen, Rasmus Hannibal Tirsgaard, Fran\c{c}ois R J Cornet et al.
In briefUP-MOPD: Update Projection for Multi-Teacher On-Policy Distillation
SummaryThis paper proposes UP-MOPD, a method for on-policy distillation from multiple teachers that combines domain expertise into one student model. It lets the mixed gradient update the optimizer state to form a candidate displacement, then projects only violating candidates to the nearest feasible update before committing it. In experiments combining medical and general domains, UP-MOPD improves late-training IFEval-loose accuracy by 2.96 points over vanilla M-OPD.
In briefForesight-over-Graph: A method for knowledge base question answering that looks beyond local graph horizons
SummaryThis paper addresses a limitation of LLM-guided graph reasoning, where hop-by-hop greedy or beam pruning can discard evidence that only becomes useful after deeper graph context is explored. It proposes Foresight-over-Graph (FoG), which uses far-to-near feedback to guide path exploration, iteratively builds a question-relevant evidence subgraph, and keeps a compact memory subgraph. The authors report state-of-the-art results on widely used KBQA benchmarks, including a 16.58% Hit improvement on CWQ, along with fewer LLM calls and lower token usage.
In briefDecision-Focused Learning for MDPs via Occupancy-Measure Linear Programs
SummaryThis paper studies decision-focused learning (DFL) for Markov decision processes (MDPs). Existing methods differentiate through the KKT conditions of the Bellman equation and solve a linear system over all state-action pairs, which limits scalability. The authors reformulate the MDP as an occupancy-measure linear program and derive a closed-form gradient by identifying active constraints with a pivoting algorithm. To handle gradient discontinuities and state-size scaling, they propose an augmented Lagrangian surrogate, random row sketching, and a learnable soft state-aggregation layer, reporting lower regret than KKT-based DFL and two-stage baselines at significantly lower computational cost.
Zihao Zhao, Ashwath K. Karunakaram, Ali Eshragh et al.
In briefUniCounting: instance-aware merging of proposals for counting multiple categories without image queries
SummaryThis paper studies counting multiple object categories from a single RGB image, with no exemplar or text query, using a fixed vocabulary. UniCounting uses frozen SAM 2.1 for masks and frozen DINOv2 and OpenCLIP for features, and trains only a 3,267-parameter relation head that predicts whether masks belong to the same instance. On COCO clean500 it gets lower point-estimate vector ℓ1 error and absent-class false mass than calibrated OWLv2-All80, with comparable micro presence F1.
In briefAI-driven optimization speeds up development of PLGA in situ forming depots for long-acting injectables.
SummaryCorbion and Intrepid combined Corbion's PURASORB polymer library with an AI algorithm called ANDROMEDA 1 to develop in situ forming depots for a therapeutic peptide. Over about 15 weeks, 181 formulations with 6-12% w/w drug loading were prepared and characterized. Four lead candidates at 6%, 9%, and 12% drug loading met the viscosity and injectability criteria and showed distinct 30-day in vitro release profiles. Polymers with intermediate molecular weights offered a favorable balance between sustained release and solution viscosity.
Pauric Bannigan, Siddarth Chandrasekaran, Brigitte A. G. Lamers et al.
In briefEvolutionary One-Step Generators: fast, diverse sampling for discrete design
SummaryThe paper proposes EGO, a framework for training compact generators directly on discrete outputs, combining distribution matching, structural constraints, and optional diversity rewards via antithetic low-rank evolution strategies without criterion-specific differentiable surrogates. After training, the generator produces an entire graph in a single neural-network evaluation. The authors report over 50 times the valid-and-unique yield per estimated dense operation versus recent one-step flow-map baselines on molecular generation, and more distinct held-out elite architectures than relaxed gradient training on NAS-Bench-101.
Marcus Vukojevic, Erik Nielsen, Veronica Lachi et al.
In briefTesting GPT-5.5's Implicit Reading of Visual Encodings with Stevens's Power Law
SummaryThe paper adapts Stevens's power law to measure how AI models implicitly read visualizations, without legends or colormap names. GPT-5.5 first estimates the magnitude of a reference visual representation, then estimates the magnitude of each later image relative to that reference. The evaluation covers twelve visual variables, making model readings of visual encodings measurable, comparable with human perception, and more transparent to humans.
In briefTransect: An open-source tool for observing long-horizon LLM agent evaluations
SummaryThis paper presents Transect, an open-source package built on Inspect Scout that helps evaluators understand how long agent runs unfold and check model-generated behaviour labels against the transcript. It aligns events, token use, and sub-agent activity on a common timeline with navigable reports. The authors demonstrate it on an AI R&D evaluation that generated almost 13 million tokens, finding a focus on operational work and manuscript production with little sustained hypothesis generation.
Toby D. Pilditch, Konstantinos Voudouris, Alexandra Abbas et al.
In briefExplainable Failure Prediction and Prevention in Maritime Systems
SummaryThis chapter reviews the key AI technologies needed for explainable failure prediction and prevention in maritime systems. It presents a conceptual architecture that combines data acquisition, time-series forecasting, anomaly detection, risk assessment, decision-making, and explainable AI into a closed-loop framework supporting autonomous or human-in-the-loop corrective actions. The authors review relevant maritime studies and their methods, advantages, and limitations. They also discuss challenges such as uncertainty and robustness, model generalization, explainability, scarce maritime datasets, and operational deployment.
Dionisis Kalogeropoulos, Georgia Sovatzidi, Panagiotis G. Kalozoumis et al.
In briefTest-time adaptation for quantized vision transformers via single-pass recalibration
SummaryThis paper addresses the problem that quantized vision transformers (ViTs) lose accuracy under distribution shift. It proposes Quantizer-Aligned Recalibration (QuAR), a test-time adaptation method that uses a single forward pass without backpropagation or parameter updates, adjusting activations before a frozen quantizer toward the source calibration statistics. On ImageNet-C with ViT-B, QuAR reports the highest mean accuracy among backprop-free TTA methods across 3- to 8-bit precisions, with 46% lower latency and only 0.17 MB of memory overhead.
In briefUsing sensor layout as a prior for generating multi-channel brain signals
SummaryThe paper proposes building a graph from electrode positions on the head and using a graph-Matérn function of its Laplacian as the source covariance for flow-matching models, so generation starts from spatially coherent patterns rather than independent noise. Across eight EEG datasets and four flow-matching methods, this lowers the spectral discrepancy (PSD-KL) in five clinical bands on most datasets, and by up to 40% on PhysioNet-MI. The authors attribute the gain to the spatial eigenvectors of the sensor graph and report similar improvements on MEG, intracranial EEG, and a traffic-sensor network.
In briefUncertainty quantification is essential for reliable brain-network graph learning, with a case study
SummaryThis paper reviews uncertainty quantification methods for connectome-based graph learning and illustrates the risks of uncalibrated predictions. A temporal Graph Attention Network trained on dynamic functional connectivity matrices from the SUDMEX CONN dataset reached 80.0% accuracy (F1 = 0.794) for Cocaine Use Disorder. A Monte Carlo dropout audit, however, revealed severe overconfidence (ECE = 0.127), with some misclassified subjects receiving confidences up to 95%. The authors conclude that rigorous uncertainty quantification and calibration are necessary before graph-based biomarkers are deployed clinically.
In briefHow Much Planning Is Enough? Cutting Search and Compute in World-Model Planning
SummaryThis paper studies whether visual world models can reach near-full performance in goal-directed control with smaller action-search budgets. It proposes SufficientPlan, which includes Paired Sequential Budget Certification (PSBC) to find and verify a reduced budget, and Static-Context Reuse (SCR) to cache observation and goal representations across search iterations. Experiments across several world-model backbones and visual-control tasks report substantially lower search budgets and planning latency while keeping competitive control performance.
In briefNetwork intervention by repeatedly querying strategic agents
SummaryThe paper studies how a planner sets non-discriminatory prices in a network of strategic agents whose private information is unknown and who may misreport. The authors show that optimal prices decompose by centrality, which motivates Poll, an algorithm that samples one agent per round and updates the price from a local report. The authors report that Poll needs far fewer operations and communication than exact and other distributed methods, and that it converges to welfare-maximizing prices.
In briefPolarScale: A physics-based benchmark for estimating polarization (Stokes) from RGB images with consistent radiometric scale
SummaryThe paper introduces PolarScale, a benchmark that treats the radiometric scale, which is normally divided out when predicting polarization from RGB-like inputs, as an explicit prediction target. Models predict normalized Stokes components, AoLP/DoLP/DoCP, and a per-scene scale, which is compared against a constant-scale control. The authors report that the strongest restoration-based models estimate the scale with 3.6–4.3% mean relative error versus 5.7% for the constant control, and violate physical bounds on fewer than 0.25% of pixels.
In briefStatistical inference for sparse SVMs in high dimensions: confidence intervals and tests for features
SummaryThis paper develops a statistical inference framework for sparse support vector machines when the sample size and number of features grow proportionally. It handles the nonsmooth hinge loss by rewriting the L1-penalized SVM as a linear program and extracting the subgradient from dual variables, yielding a debiased estimator whose coordinates are asymptotically Gaussian. This supports confidence intervals, hypothesis tests, and false-discovery-rate-controlled variable selection, as shown in simulations and a breast cancer gene-expression analysis.
In briefDIPrune: Task-aware token pruning that speeds up multimodal AI models while keeping meaning
SummaryThis paper proposes DIPrune, a training-free method that removes less important tokens from multimodal large language models to reduce computation. The authors argue that existing pruning methods lose semantic information because they ignore the task or rely on unreliable attention estimates. DIPrune ranks tokens using two scores, one for static feature saliency within a layer and one for semantic change across layers. The authors report consistently state-of-the-art results on LLaVA and Qwen-VL.
In briefDigital-twin pipeline generates realistic driving images to cut real-world data collection for 3D perception
SummaryThe paper proposes DT-R2S2R, which reconstructs recorded driving clips in a georeferenced digital twin and uses aligned simulator renderings to condition a diffusion model that synthesizes photorealistic driving images. The generated data is used to train 3D detectors without target-region images. The authors report that DETR3D reaches 93.18% of the mAP of a target-region real-data oracle, and co-training with out-of-target real data exceeds that oracle.
In briefMachine Learning Forecasts of German Redispatch Under Delayed Data and Shifting Patterns
SummaryThis paper uses published German transmission records to test how well probabilistic machine-learning models forecast grid congestion intervention volumes when data arrives with delays. It compares seasonal, ARX, quantile LightGBM, GRU, and Transformer models under a minimum seven-day latency constraint. Boosted-tree models with rolling calibration performed best overall, but interval coverage fell substantially during high-volume intervention events.
Faraz Shamim (1), Faris Shamim (2) ((1) KIST Medical College, Teaching Hospital et al.
In briefSC3BF: Shifted Collision-Cone Barrier Function for Avoiding Moving Obstacles
SummaryThis paper proposes SC3BF, a shifted collision-cone control barrier function for robots avoiding dynamic obstacles. It adds a state-dependent allowance to the collision cone condition, so the robot may approach obstacles at a rate that grows with distance and its own speed, rather than rejecting all velocities aimed at an obstacle. SC3BF is enforced through an ordinary quadratic program, and the paper proves a nonzero safety-preserving allowance always exists and derives it in closed form. In experiments with a kinematic bicycle among up to 100 moving obstacles, SC3BF reached the goal more often than three velocity-space baselines and modified the nominal input less than half as much.
In briefTransferable Spatiotemporal Coherence Attack on Black-Box Vision-Language Models for Driving
SummaryThis paper proposes Spatial Temporal Coherence Adversarial Attack (STCA), a method that perturbs video inputs to vision-language models used in autonomous driving. The approach combines caption-guided frame selection, a spatial perturbation stage, and a motion-guided mask that disrupts cross-frame temporal coherence. It operates under a black-box setting, relying on transferability from a white-box surrogate model. Experiments on BDD100K and nuScenes with three VLMs report that the spatial attack achieves an attack success rate while keeping high SSIM, suggesting these models remain vulnerable in driving scenarios.
In briefMoF: Preference-Aware Mixture Modeling for Personalizing Black-Box LLMs
SummaryThis paper proposes Mixture-of-Facets (MoF), a framework for personalizing black-box large language models. Instead of using dedicated parameters for each user, it models user preferences as combinations of shared latent preference facets, routed according to user history. The authors report stronger personalization performance and a more parameter-efficient design across diverse tasks, with strong generalization to users unseen during training.
In briefAI-assisted Lean 4 formalization of the Poincaré conjecture
SummaryThis paper describes an AI-assisted project that formalizes the Poincaré conjecture in the Lean 4 proof assistant. The project started with little reusable formal infrastructure for the geometric analysis behind the proof. The authors combined a mathematician-prepared proof blueprint with explicit milestone statements, which enabled parallel agent work and clear points for mathematical guidance. The paper also analyzes the human interventions and organizational choices in this workflow, and suggests the resulting infrastructure could eventually lower the cost of verifying geometric analysis results.
WhyFull Lean 4 formalization of the Poincaré conjecture, organized via milestone blueprints and AI assistance, offering substantial verified infrastructure.
In briefMedZERO: Self-improving AI agents for open-ended medical reasoning via controlled knowledge accumulation
SummaryThis paper presents MedZERO, a framework in which language models improve by generating their own medical questions and solving them through multi-turn reasoning with external knowledge tools. It separates temporary exploratory knowledge from curated persistent knowledge to support reliable continual learning. The authors evaluate 4B- and 8B-scale base models on five public medical reasoning benchmarks under open-ended evaluation. They report the highest average accuracy among compared methods at both model scales.
In briefLocating obstacles without communication, using the team's combined force and torque in leader-follower cargo transport
SummaryThis paper studies obstacle localization for robots that cooperatively carry a rigid payload without explicit communication. A leader robot directs the payload while follower robots react to obstacles they detect locally. The leader can measure the followers' combined force and torque but cannot separate their individual reactions. The authors design a follower control law, give sufficient conditions for exact obstacle recovery, and demonstrate the method in simulations.
Amin Kashiri, Aditya Mohan, and Yasin Yaz{\i}c{\i}o\u{g}lu
In briefStructure-aware graph abstention for reliable selective forecasting
SummaryThis paper studies selective forecasting, where a model abstains on high-risk test windows under a retained-coverage budget. Existing gates such as TEM score whole forecasts and may miss violated dependencies among variables in multivariate outputs. The authors propose a learned sparse graph and a structural energy measure to assess relational consistency. On seven long-horizon benchmarks and four backbones, structural gating often reduces selective MSE at matched coverage, though gains are not universal.
In briefSCOPE: Certified Theorem Proving Using a Language Model as a Planner
SummaryThis paper introduces SCOPE, a method for proving multi-step numeric propositions in Lean. The language model plans over a set of operators, a symbolic engine computes the numbers, and a compiler renders the proof. On a 218-problem suite, a 135M-parameter backbone certifies 191 problems (87.6%), compared with 18 for a 7B baseline model.
WhyDivides labor among model planning, symbolic numerics, and compiler rendering, greatly lifting a small model's certification rate, with a quantitative power-law analysis. Evaluation spans several base
In briefCatastrophic forgetting in thermal anti-drone detection from scale-dependent gradient imbalance
SummaryThis study examines catastrophic forgetting in thermal infrared anti-UAV detectors trained sequentially across three benchmarks of increasing difficulty. Naive fine-tuning causes about 90% capability loss, while knowledge distillation from a frozen teacher retains about 95% of prior performance, though without a no-distillation control. The authors suggest that collapse in large-target detection stems from scale-conditioned gradient imbalance rather than weight drift. A scale-stratified replay buffer roughly halves forgetting, but an ablation attributes the gain mainly to stratification, and the replay results come from a single seed and are preliminary.
Khac Duc Giang Nguyen, Seyed Sahand Mohammadi Ziabari, Ali Mohammed Mansoor Alsahag
In briefTesting Whether Tabular Foundation Models Mix Context Labels Together
SummaryThis paper examines whether pretrained tabular foundation models (TFMs) can be explained by a fixed-weight account of in-context learning, where features set the weight of each context label. The authors note that public TFMs standardize labels before prediction, so ordinary derivatives can make a model look nonlinear even when it matches a fixed-weight map. They propose two certificates that use only predictions at standardized labels to reject fixed-weight and independent nonlinear label-transformation explanations. Across five public TFMs, changing one context label alters how others influence predictions, a behavior they call joint processing, which emerges with training and is mostly carried by attention scores.
In briefCoDe-LoRA: Easing the orthogonality trade-off in continual learning for LLMs by consolidating and separating knowledge
SummaryThis paper addresses continual learning for large language models, arguing that orthogonal-projection LoRA methods such as O-LoRA block the sharing of representations among related tasks, an issue it calls the "Orthogonality Dilemma." The authors propose CoDe-LoRA, a replay-free method that separates universal knowledge consolidation from task-specific decoupling, using adaptive null space projection and semantic routing. The authors report the best average accuracy across four backbones and three continual learning benchmarks.
WhyExtends O-LoRA with adaptive null-space projection and semantic routing, validated across four backbones and three continual learning benchmarks.
In briefLanguage Unalignability: Why Some Concepts Resist Cross-Cultural Benchmark Evaluation
SummaryThe paper argues that current multilingual LLM evaluation assumes semantic structures map across languages without loss, an assumption it says fails for pragmatic markers, honorifics and similar concepts. It formalizes this with a usage-cloud framework and reports evidence from behavioral benchmarks and a nine-model case study on Yami. The authors propose a multidimensional diagnostic profile for cultural competence.
In briefHow much memory context helps: a controlled test of deeper versus wider retrieval for chat assistants
SummaryThis paper studies how two settings in long-term conversational memory for language models affect answer accuracy: the structural depth of memory and the amount of context given to the answering model. Using the EverMemBench benchmark, the authors test depths D1 to D4 with context budgets of 1,024 to 4,096 tokens. Widening the context from 1K to 4K raised accuracy by 10.11 to 17.98 points, while deeper structures gave no consistent gain. The authors suggest studying large, coherent context blocks rather than deeper memory hierarchies.
In briefReducing Concept Drift in Network Quality Prediction for Autonomous Vehicle Teleoperation Using Past Data
SummaryThis paper uses a data measurement campaign to build a framework that predicts two network indicators for teleoperating autonomous vehicles: uplink data rate and round-trip latency. It adds historic data to the prediction pipeline to reduce the performance drop that machine-learning models suffer on unseen data due to concept drift. The authors also introduce a critical scenario detection metric to evaluate prediction performance specifically for teleoperation.
In briefOxiGen: A crystal generator that respects oxidation states
SummaryThis paper introduces OxiGen, a crystal diffusion model that explicitly represents oxidation states to generate inorganic crystals that are more realistic and experimentally feasible. It enforces global charge neutrality by construction, using a structured output layer with exact inference over a finite-state automaton. The authors report that OxiGen better reproduces observed oxidation-state distributions and yields the highest rate of stable, unique, and novel crystals among the compared methods, while keeping compositional validity high under property conditioning.
In briefWhere two groups of persistence diagrams differ: calibrated local inference with a fixed sample size
SummaryThis paper compares two samples of persistence diagrams, a topological summary of data, by testing mean differences simultaneously in several local neighborhoods of the birth-death plane. It uses a Gaussian multiplier bootstrap to build simultaneous confidence intervals that allow unequal group covariances. Simulations report simultaneous coverage close to the nominal level. An exploratory analysis on the MUTAG benchmark is also included.
In briefScalable extraction and visualization of multi-attribute logical and functional dependencies in tabular data
SummaryThis paper proposes LDTool and HLDTool, two tools for extracting and visualizing multi-attribute logical and functional dependencies from tabular data. LDTool extends dependency discovery beyond pairwise relationships, while HLDTool uses hypergraph-guided search-space reduction to scale to larger datasets. Experiments on three simulated and eleven real-world datasets show that LDTool recovers the same functional dependencies as existing methods with lower runtime in high-dimensional spaces, and HLDTool handles datasets with hundreds of features.
Chaithra Umesh (1), Arvind Lomrore (4), Neethu D (4) et al.
In briefDetecting key events in table tennis videos from 2D keypoints
SummaryThis paper presents EventNet, a two-stage method that automatically detects frame-accurate key events, such as ball-racket contacts, in table tennis videos. It extracts 2D keypoints of both players' upper bodies, table corners, and the ball center, then uses a transformer encoder to predict time-to-event values for each frame. The approach achieves an F1 score of 91.16% on Latte-MV and 73.08% on TTHQ.
Rainer Lienhart, Daniel Kienzle, Shin'ichi Satoh et al.
In briefWhose Face Is It Anyway? A Multi-Model Audit of Child Facial Expression Recognition
SummaryThe authors audit five expression recognition models, pretrained on adult faces, across child image and video datasets using one shared evaluation harness. All models perform worse on children, with the largest degradation for open-mouth faces and South-Asian children, and the bias tracks expression morphology rather than skin tone. A linear probe on frozen features suggests the gap lies mainly in the classifier head, and retraining only the head on a small amount of target data improves accuracy on the two largest child datasets.
WhyUnified multi-model audit shows child expression gaps lie in the classifier head rather than representations, supported by controlled comparisons across five models and datasets.
Tobias Hallmen, Robin-Nico Kampa, Elisabeth Andr\'e
In briefTwo-Sample Testing via Generative Processes
SummaryThis paper addresses whether two samples come from the same distribution by building a stochastic interpolant between them. Under a symmetric schedule, the interpolant's law is time-reflection invariant when the distributions match, so the test measures the Jensen–Shannon divergence between marginals at times t and 1−t. Permutation calibration gives exact finite-sample level, and the authors report that the test matches or outperforms state-of-the-art kernel two-sample tests.
In briefPerformative prediction when only some labels are observed; a robust method keeps models near a stable solution.
SummaryThe paper studies machine learning systems whose deployment changes population behavior, a setting called performative prediction. It shows that retraining only on observed labels, which exist for accepted individuals only, can misguide retraining and undermine convergence to a stable solution. The authors propose a worst-case objective based on confidence intervals for the probability of a positive label, and apply repeated risk minimization to stay within a bounded distance of the true stable point. In a lending experiment with fairness regularization, the approach performs closely to repeated risk minimization with complete label access.
In briefReinforcement learning with segment-level reward feedback under linear approximation
SummaryThis paper studies reinforcement learning when rewards are given only as segment-level feedback on trajectories, a middle ground between per-step and whole-trajectory rewards. The authors design algorithms for known and unknown transitions under linear function approximation and establish regret bounds that nearly match lower bounds. They find that with binary feedback, more segments sharply reduce regret, while with sum feedback the segment granularity has little effect.
WhyProposes RL algorithms under segment reward feedback with linear approximation, giving near-optimal regret bounds matched by lower bounds, and revealing distinct effects of segment granularity for bin
In briefDySCo: Dynamic sharding and depth-synchronized batching for edge-cloud LLM inference
SummaryThe paper presents DySCo, a collaborative runtime that lets resource-constrained edge devices compute part of a large language model's layers while keeping KV caches local. To address idle gaps and the inability to batch requests at different model depths, it introduces dyForward, a layer-range executor, and depth-synchronized batching (DSB). In experiments at an average concurrency of eight, DSB improves throughput by 275% over FIFO and 48% over exact-match batching.
Jingpo Xu, Paul Joe Maliakel, Ivona Brandic et al.
In briefzkLLMPoT: Zero-knowledge proof that a private LLM was trained to a claimed result
SummaryThis paper presents zkLLMPoT, a zero-knowledge framework that lets an auditor check properties of a trained large language model without seeing its weights or private training data. The auditor picks challenge sequences after the model weights are committed, and the trainer proves the objective value the model achieves on them, so the cost does not depend on the number of training iterations. Benchmarks across four model families report proving times of 41 to 59 seconds for 1.1 to 1.5 billion parameter models and 131 seconds at 13B for the covered operators, with verification under half a second at sequence length 512.
In briefMASC: Multi-agent self-calibration for consistent client role-play in counseling
SummaryThis paper presents MASC, a framework in which multiple agents generate and self-calibrate simulated counseling clients so they stay psychologically consistent over long dialogues. The authors also introduce CRPC-Bench, which includes 38 motivational interviewing client profiles annotated with Big-Five personality and emotion data. Experiments report that MASC outperforms existing methods on profile, personality, receptivity, and turn-level consistency, with the heterogeneous configuration performing best overall.
In briefTask-Sufficient Contraction: Reducing Sources for Machine Information Exchange
SummaryThe paper studies when a task can justify reducing a source before later encoder, rate, or distortion choices are made, while still preserving the full family of downstream problems. For a machine with fixed actions and loss, states are merged only when every available action has the same regret in both, yielding a consumer-specific source. For finite action sets this preserves the complete one-step rate-regret curve, and for quadratic loss on affine feasible-action sets an exact characterization is given as projection onto the directions where feasible actions can differ. The framework is positioned relative to Information Bottleneck, semantic rate-distortion, and goal-oriented quantization results.
In briefLeanPlan: Optimal planning using LLM-written heuristics with machine-checked admissibility proofs
SummaryThis paper introduces LeanPlan, a planning system that uses large language models to generate search heuristics and proofs that those heuristics are admissible, with the proofs checked by machine in Lean 4. An agentic loop refines a reusable domain-specific heuristic, its proof, and needed domain assumptions using feedback from a planner. The authors evaluate it on ten International Planning Competition domains and three new ones, and report that it generated heuristics and proofs for all of them. They state that LeanPlan usually expands fewer states than the Scorpion planner and solves more tasks overall.
WhyFirst machine-checked admissible LLM heuristics for optimal planning, fewer expansions than Scorpion across 13 domains; evidence fairly solid.
Andr\'{e} G. Pereira, Augusto B. Corr\^ea, Felipe Meneguzzi et al.
In briefSensor-Language-Action Models: A unified framework linking sensor data, language, and actions
SummaryThis paper introduces Sensor-Language-Action (SLA) modeling, which connects multimodal sensor observations, natural language, and actions within one model, using language as the semantic link between sensing and acting. The authors build a large benchmark spanning over 116,000 individuals, 79 sensor modalities, and 60 action groups, and present OpenSLA for hierarchical action prediction, state understanding, and action explanation. The paper reports superior performance over state-of-the-art methods on clinical prediction, operating room, and metabolic health tasks, along with language-guided evidence grounding and zero-shot generalization to unseen actions and cohorts.
In briefOSFP4: Jointly Optimizing Diagonal Smoothing and Block Scales for NVFP4 Quantization
SummaryThis paper introduces OSFP4, a quantization method for NVFP4, a compact number format used in large language model inference. For each linear layer, it optimizes a diagonal smoothing matrix together with block scales to reduce matrix-product quantization error under NVFP4. The authors report the highest average accuracy among the compared methods and about 94-97% of vendor NVFP4 prefill throughput on the measured workloads.
In briefTSRN-RTVD: A real-time system that removes motion blur from video.
SummaryThis paper presents TSRN-RTVD, an efficient video deblurring system that reconstructs the camera trajectory during exposure and uses the recovered motion to guide restoration. It runs at 30 FPS on a single consumer GPU and reaches 30.08 dB PSNR on the GoPro dataset. The authors also demonstrate the system on consumer devices with interactive visualization.
Nikita Alutis, Danila Evsyukov, Egor Chistov et al.
In briefConfidence reversal in neural decoding when contextual priors are applied
SummaryThis paper studies how contextual priors affect confidence in neural-signal speech retrieval, using MEG-MASC and MOUS data with additive shallow fusion and the top-two margin as confidence. Among initially incorrect predictions, a larger margin makes repair more likely when the correct candidate starts near the top of the local ranking, but less likely when it starts lower. Changing only the fusion weight shifts this reversal to deeper ranks as predicted. Reading local and prior scores separately improves selective decoding, answering on 74.5% of windows instead of 56.7%.
In briefHow many independent samples a satellite image really holds, and the limits on learning from spatially correlated pixels
SummaryThe paper argues that neighboring pixels in satellite images are correlated, so treating every pixel as independent overstates the sample size for classifier evaluation. For an n×n image with spatial correlation over r pixels, it shows the effective sample size is Θ(n²/r²), proving a finite-sample upper bound and a matching lower bound. The theory is validated on synthetic data and image tiles from Landsat 8, Sentinel-2, and Sentinel-1.
In briefVOMMI: A Framework Using Portable Demonstrations for Mobile Manipulation Learning
SummaryVOMMI is a framework that collects mobile-manipulation demonstrations with portable RGB devices, without robots, and uses them to post-train vision-language-action policies. It refines trajectories offline with sparse geometric anchors (R2-VO) and conditions the policy online through a residual adapter. A policy trained only on portable demonstrations achieved 18.2% lower base-velocity error than one trained on robot-collected data, and improved mean success rate by 8.3 points over OpenPI 0.5 on three real-robot tasks.
In briefLearn2Play Bench: How Well LLM Agents Learn from Experience in Unfamiliar Games
SummaryThis paper introduces Learn2Play Bench, a benchmark of newly designed text-based games with novel or counterintuitive rules that require agents to learn through interaction rather than rely on pretrained knowledge. The games give reproducible feedback and automatic scoring, and varied game instances test whether learned knowledge transfers to new situations. The authors report three findings: retaining full action and feedback records supports learning better than summarizing them into rules, top human players reach higher peak scores than the evaluated agents, and changing the agent harness can improve performance while lowering estimated inference cost.
In briefProof Assistants for Teaching Logic: The LogiKEy Teaching Approach
SummaryThis paper describes a method for teaching logic to mixed classes of computer science, mathematics, and philosophy students, based on the LogiKEy methodology used in courses for over a decade. It uses classical higher-order logic as a metalogic, encoding other logics via semantic embeddings so that one proof assistant such as Isabelle/HOL serves as a common environment for learning and comparing them. The authors present a graded sequence of classroom examples, from puzzles in modal and epistemic logic to Gödel's ontological argument, and respond to the objection that this approach is monist rather than pluralist.
Christoph Benzm\"uller, David Fuenmayor, Luca Pasetto
In briefRACE-FPP: AI-assisted calibration method for more robust fringe projection 3D measurement
SummaryFringe Projection Profilometry needs precise system calibration, but checkerboard corner detection can degrade under lens blur and unfavourable target orientations. This paper presents a full FPP calibration pipeline that adds deep-learning corner detection to camera calibration and calibrates the projector by sampling phase values at white-square centres. On a mixed dataset of clean and degraded images, camera reprojection error dropped from 1.237 to 0.259 pixels, and projector reprojection error fell by about 50%. Reconstructed artefacts also showed improved geometric accuracy.
Osman Ali (1), Xiangjun Kong (1), Tibebe Yalew (1) et al.
In briefA sequential test for telling two simulated models apart at any stopping time
SummaryThis paper studies hypothesis testing between two model distributions, P0 and P1, when only i.i.d. simulated samples from each are available rather than analytic densities. The authors construct an e-test martingale that yields a sequential test. The test controls type-I error at any stopping time and has approximate growth optimality, geometrically decaying type-II error bounds, and asymptotic power one.
In briefTrust-region optimization of smooth particle-interaction energies in Wasserstein space
SummaryThis paper studies trust-region optimization of smooth potential-interaction energies over probability measures with finite second moment in Wasserstein space. The method combines a quadratic model along pushforward curves, an L^2 step radius, a Steihaug-Toint subsolver, and a ratio test for acceptance and radius updates. Under a lower energy bound and globally bounded Hessians, the authors prove the objective is nonincreasing and an epsilon-stationary iterate is reached within O(epsilon^-2) total outer trials. Numerical experiments cover a soft-particle energy, MMD minimization, and scaling studies.
In briefSTRUCTURALCOST: A Controlled Reading Time Dataset for Modeling Human Sentence Processing Difficulty
SummaryThis paper introduces STRUCTURALCOST, a self-paced reading dataset with 475 participants and 40,800 observations that isolates the processing cost of long-distance subject-verb dependency resolution. It replicates a previously low-powered psycholinguistic finding at larger scale: human reading times at the main verb increase with dependency length, driven by syntactic embedding rather than linear distance. Various language models partially mirror this difficulty pattern but underestimate the integration cost humans incur, a gap that persists across architectures and model sizes.
Nina Nusbaumer, Iria de-Dios-Flores, Corentin Bel et al.
In briefTool-calling retrieval vs. vector RAG for a small Greek–English knowledge base
SummaryThe paper builds KyGround, a 198-question benchmark from published records of a Greek–English agricultural platform in Kythera, and compares tool-calling retrieval with vector retrieval-augmented generation. On canonical Greek questions, vector RAG answered 95.3% correctly versus 71.6% for the tool agent. The tool agent's failures came from literal searches that returned nothing when query terms did not match records verbatim, such as Greeklish or unaccented input. The authors conclude that tool interfaces for community knowledge bases need search tolerant of how users type.
Nikolaos D. Tantaroudas, Ilias Karachalios, Andrew J. McCracken
In briefWhere LLMs Retrieve Network Information: Locating the Responsible Attention Heads and Neurons
SummaryThis study examines how large language models retrieve relationships between network entities such as domain names and IP addresses, seeking the internal attention heads or neurons that causally enable this retrieval. Across five models from three architecture families, the authors identify heads through causal ablation and build detectors evaluated without retraining on an independent reverse-DNS dataset. A small set of causally validated heads achieved 99.5–100% held-out accuracy, while neuron-level results varied by model and single-neuron detectors fell below chance in three of five models on CAIDA data.
Md Abdul Kadir (1, 2), Md Mohasin Hossain (1 et al.
In briefMacJEPA: Audio-visual recognition that stays robust when sensors drop out in untrimmed egocentric videos
SummaryThis paper proposes MacJEPA, a model that recognizes visual actions and acoustic events in egocentric videos from interval queries. It treats missing modalities as temporally localized sensor outages within long, untrimmed videos and simulates them during training with window-local modality dropout. According to the paper, a single checkpoint stays competitive with full input on Epic-Kitchens-100 and Epic-Sounds and outperforms published missing-modality baselines when either stream is removed.
In briefPIE-PS: Photometric Stereo from Physical Irradiance Event Streams
SummaryThis paper presents PIE-PS, a learning-based framework that reconstructs dense surface normals from raw event-camera streams captured under known moving lighting. It pairs adjacent events at each pixel with their light directions to form Physical Irradiance Events, using a signed event rate that does not need the unknown contrast threshold. A graph network shares context across nearby events, and a reliability attention module down-weights unreliable observations. Experiments on synthetic and real data report better results than prior event-based photometric stereo methods and a direct physics-only solver.
In briefSmall robot policies need fine-grained visual representations
SummaryThis paper examines what drives performance in multi-task robot manipulation policies. The authors build CoRP, a compact 48.9M-parameter policy with a representation extractor and flow-matching action generator, reaching 97.0% on LIBERO and 75.78%/73.36% on RoboTwin 2.0 Clean/Randomized. Experiments indicate that pretrained visual initialization and visual token compression matter most, while language conditioning helps mainly when the goal is ambiguous.
In briefView Matters: keyframe-guided text-driven editing of 3D Gaussian scenes
SummaryThis paper addresses text-driven editing of 3D Gaussian scenes, noting that rendered views differ in how reliably they guide an edit, yet are usually treated equally. The authors propose View Matters, which selects reliable keyframes using geometric visibility, semantic distinctiveness, and edit relevance, propagates their editing signals to other views, and keeps this reliability preference during optimization. Across 23 scene-prompt pairs, it achieves the highest average CLIP text-image and directional similarity among the evaluated methods, with an editing time of about four minutes.
In briefLFHE: Bounded Local Topology Search for Decentralized Learning on Non-IID Data
SummaryThe paper studies how communication topology affects decentralized learning when data is non-IID. It proposes Local-First Heuristic Evolution (LFHE), which rewires the network using only ego-neighborhood and friend-of-a-friend information under bounded sparse degree. Across four image, speech, and text benchmarks, LFHE achieves competitive performance, and matched controls identify its structural term as the main topology-selection signal.
Yin-Kuan Liang (Durham University), Yan Gao (University of Cambridge), Yang Long (Durham University)
In briefBeyond the Leaderboard: Multi-Dimensional Evaluation of Dense and MoE Models for Program Repair
SummaryThis paper argues that automated program repair is usually judged only by whether patches pass tests, which can hide differences in maintainability, security, and computational cost. The authors propose a Weighted Quality Index inspired by ISO/IEC 25010, combining correctness, maintainability, security, and generation efficiency. Comparing three dense Qwen2.5-Coder models with a Mixture-of-Experts DeepSeek-Coder-V2-Lite model on identical hardware, they find model rankings shift with the weighting, and the MoE model shows no significant correctness difference from the 7B and 14B dense models while using 3 to 6 times fewer active parameters.
In briefVisual Orchestration Tax: Auditing and Certifying Visual Evidence Reuse in Agentic VLM Pipelines
SummaryThe paper observes that agentic vision-language model pipelines repeatedly pass the same static image through multiple agents and tools, creating redundant processing. The authors define audit metrics to count repeated visual-evidence touches and introduce SharedVisCache, a reuse hook keyed by image content, preprocessing fingerprint, and encoder assumptions. Reported experiments show that certified reuse reduces vision-encoder calls while keeping outputs unchanged.
In briefInferring alloy composition and process parameters from optimized microstructure and texture
SummaryThis paper asks whether the recipe of a magnesium alloy, meaning its composition and extrusion process settings, can be inferred backwards from its microstructure and texture. Using an in-house dataset of 107 extrusion conditions across 14 alloys, the authors compare conventional grain and texture statistics, a pretrained image encoder embedding, and a graph neural network on the grain network. Conventional descriptors identified the correct alloy for 65% of held-out conditions, versus 17% for always guessing the most common alloy, while learned embeddings stayed below 30%. Process parameters were recoverable but noisier, with microstructure roughly halving the temperature error compared with composition alone.
In briefAlign, Then Correct: A Training-Free Two-Stage Low-Rank Fix for Extremely Quantized LLMs
SummaryThis paper addresses the accuracy loss of large language models when their weights are quantized to extremely low bit-widths. It proposes a training-free, two-stage closed-form method that adds low-rank adapters: the first stage aligns layer outputs with the full-precision model using a Fisher-weighted objective, and the second stage applies a rank-constrained natural-gradient step on the compensated model. At 2 bits under QuIP#, WikiText-2 perplexity drops from 12.43 to 10.26 on Qwen3-8B and from 21.11 to 13.22 on Qwen3-4B.
In briefThe Failure Is in the Readout: Emotion Benchmarks Measure Elicitation
SummaryThis paper examines the EmoNet-Face-HQ benchmark, which concluded that vision-language models score poorly on fine-grained facial emotion recognition and need a dedicated fine-tuned model. The authors instead read each category's answer as a binary query from model logits, keeping the same images, taxonomy and expert ratings. Under this readout, off-the-shelf models match or exceed the fine-tuned model, which the authors attribute to the elicitation method rather than perception.
Tobias Hallmen, Fabian Deuser, Robin-Nico Kampa et al.
In briefBenchmarking AI Clinical Note Generation: Comparing Ambient Documentation Systems
SummaryThis paper introduces MedConv, a multilingual dataset of 300 clinical encounters in English, Danish, and German. It uses this dataset with the ACI-BENCH benchmark to compare Corti, a clinical AI platform, against two leading general-purpose AI scribe applications. The evaluation combines entailment metrics with LLM-judged pairwise comparisons across eight PDSQI-9 dimensions, and finds Corti's text generation on par with or better than the commercial scribes.
Daniel Varab, Victor Petr\'en Bach Hansen, Asbj{\o}rn W. Helge et al.
In briefMaking COMET Scores Comparable Across Writing Systems in Indic Translation Evaluation
SummaryThe paper finds that the COMET translation metric is biased by writing system: re-encoding the target into Latin script shifts scores, with script identity explaining 22.9% of native-script score variance and agreement with annotators dropping in all five languages. The authors trace this to the tokeniser and propose COMET-QN, which maps each language-script score distribution onto a shared reference, raising pooled annotator agreement from 0.300 to 0.399. They recommend publishing normalised scores, diagnostics, and tokeniser identity alongside results.
WhyProposes script invariance and diagnoses tokeniser-induced COMET bias, with distribution mapping raising annotator agreement; well-supported and broadly relevant to MT evaluation.
G. L. John Salvin (1), Swapnil Hingmire (1) ((1) Indian Institute of Technology Palakkad)
In briefToken-efficient multi-agent collaboration via System One-guided division of labor (S1-MAS)
SummaryThis paper proposes S1-MAS, a multi-agent framework in which lightweight "System One" models handle bounded coordination decisions such as task selection, role assignment, and termination, while capable LLMs handle open-ended reasoning. Experiments on seven benchmarks compare it with AgentVerse, DyLAN, and SelfOrg. The authors report that S1-MAS achieves higher accuracy while cutting GPT-4o token consumption by 44.9%-97.2% and end-to-end latency by 37.8%-93.0%.
In briefPenalty-Framed Multiple-Choice Questions With No Correct Option: How LLMs Abstain
SummaryThis paper studies whether large language models can abstain when none of the listed multiple-choice options is correct. Using the mathematics subset of MMLU-Pro, the authors remove the correct option, allow an ABSTAIN choice, and penalize invalid forced selections. Even with explicit instructions and penalty-based scoring, models still give invalid forced answers on some questions they originally answered correctly, showing that standard accuracy does not capture this reliability aspect.
In briefWhy Lean checks of AI-translated proofs may not confirm the original argument
SummaryThe paper examines using AI to translate natural-language mathematical proofs into Lean for mechanical verification. The authors argue that resolving ambiguities faithfully is extremely hard, so a passing Lean check does not guarantee the original argument is correct. They give examples of AI mistranslations, including OpenAI's announced Navier-Stokes blow-up proof, and show its Lean formalisation does not match the natural-language proof.
WhyArgues faithful autoformalisation requires ambiguity resolution at SCI infinity, and shows with AI mistranslation examples that Lean verification does not guarantee NL proof correctness.
Alexander Bastounis, Fabian Circelli, Anders C. Hansen
In briefZero-shot Coordination Under Partial Observability via Partner Intention Prediction
SummaryThis paper proposes PIP (Predicting Intention of Partner), a method for zero-shot coordination when a partner is intermittently out of view. It uses a Joint-view VAE to build a partner representation from local observations and Partner-state Belief networks to infer the partner's hidden location and behavior from interaction history. PIP is evaluated on Burrito-PO, Overcooked-PO, and a Melting Pot substrate, with a human evaluation on Burrito-PO, and it achieves the highest mean performance among compared methods on all three benchmarks.
In briefTest-time agent evolution for multi-stage legal reasoning (letting legal AI agents adapt during use, without retraining)
SummaryThis paper studies how legal AI agents can adapt during deployment, using signals from earlier cases and ongoing interactions, without updating model parameters. The proposed framework combines Test-Time Memory Evolution, which retrieves, adapts, and consolidates past experience, with Rubric-Aligned Collaboration, which verifies and revises role-specific actions. Experiments on J1-EVAL and LegalWorld across five backbone models show consistent gains over representative baselines, with the two components providing complementary benefits.
In briefVLA-ACL: Pruning Visual Tokens for Faster Robot Control Models
SummaryVision-language-action models control robots well but are slow because they process long token sequences at every step. This paper introduces VLA-ACL, which trains a lightweight visual token pruning policy with action-level supervision while keeping the base model frozen. The pruned model's actions are trained to match those of a full-context teacher. Experiments on LIBERO and real-world tasks report up to 87.5% token pruning, up to 75% computation reduction, and about 1.5x inference speedup.[](https://github.com/du-owen/VLA-ACL)
In briefBeyond Single-Column Checks: Distributed Joint-Distribution Tests for Detecting Data Drift in Large E-Commerce Systems
SummaryThe paper compares five multi-column two-sample drift detectors across three environments, including a synthetic-injection benchmark built on a 137.5-million-row Trendyol ranking feature table. A distributed Maximum Mean Discrepancy method using Random Fourier Features on Apache Spark performed robustly under strong drift, with a Pearson correlation of 0.940, an 80.4% true positive rate, and a 3.2% false positive rate. The per-dimension Kolmogorov-Smirnov test failed because its statistic saturated on ID-like columns under asymmetric sampling. Weak drift configurations were hard for all detectors to discriminate.
Cagdas Pullu, Mahmut Emir Arslan, Bugra Balkac et al.
In briefMu-DisCoCat: A Quantum Circuit Pipeline for Combining Learned Concepts
SummaryThis paper proposes Mu-DisCoCat, a multimodal variational quantum learning framework that aims to let models understand new situations by recombining learned concepts. It first learns object representations from single-object image-text pairs, then fixes them and learns relations among multiple objects. The authors report higher relational out-of-distribution accuracy than a CLIP baseline in classical simulations, and run the model on noisy emulators and an IBM quantum processor, where hardware results correlated with simulated fidelities.
Mina Abbaszadeh, Matilda Karabina Moore, Raem Haq et al.
In briefDo LLMs act on what they know? Partner representations and cooperative actions
SummaryThe study examines whether LLMs can adapt to unfamiliar partners' communication conventions in a controlled, Hanabi-derived environment with frozen weights. Linear probes recover intent conventions more accurately than target conventions across eight LLMs, yet receiving choices do not consistently match the sender's convention. Presenting conventions as externally computed action recommendations produces larger cooperation gains on average than presenting them as general rules.
