A retrofit of a custom series elastic element to a black-box actuator improved force control bandwidth from 10.32 Hz to 30.32 Hz (2.93x), outperforming a commercial sensor by 7.63% at a cost of 25 GBP.

Researchers propose a new anisotropic diffusion method for ergodic search in multi-robot systems, overcoming the uniform error propagation of traditional isotropic diffusion, using Perona-Malik diffusion gradient to guide robot motion for more flexible coverage.

A new visual navigation method called MASt3R-Nav uses pixel-relative connectivity to build geometrically accurate maps without requiring global consistency, enabling more capable navigation than traditional topological graphs.

RED is a real-time scheduling framework for multi-task deep neural network workloads on resource-constrained robotic platforms. It adapts to runtime environmental changes by assigning intermediate sub-deadlines, leveraging MIMONet weight sharing, and reconstructing computation graphs. Implemented on NVIDIA Jetson and Apple M-series platforms, RED consistently outperforms existing methods in throughput, deadline satisfaction, robustness, adaptability, and overhead.

Coronary microvascular dysfunction (CMVD) affects approximately 40%-60% of patients with ischemia and non-obstructive coronary arteries, yet diagnosis remains challenging due to reliance on invasive functional testing or subjective Thrombolysis In Myocardial Infarction (TIMI) flow grade. The TIMI Myocardial Perfusion Frame Count (TMPFC) offers an objective, angiography-based quantitative measure of CMVD, but its clinical translation is hindered by cumbersome manual calculation and insufficient validation. This study aims to develop and validate a deep learning-powered TMPFC calculation (DL-TMPFC), enabling integration into clinical workflows. In a cohort of 655 patients from three independent institutions, DL-TMPFC showed excellent agreement with expert manual measurements (bias: -0.93 fr…

ActQuant is an action-guided mixed-precision post-training quantization framework for Vision-Language-Action (VLA) models, enabling sub-4-bit weight quantization through a two-stage approach that maintains high success rates on the LIBERO benchmark and a real UR3 robotic arm, significantly reducing memory footprint.

A study compares deep learning models with linear interpolation for imputing missing satellite data due to cloud cover in aquatic monitoring. CNN-based models, especially CNN, outperformed linear interpolation across four lakes. The imputed data improved the reliability of algal bloom detection using PlanetScope SuperDove imagery.

Researchers propose a brain-to-image system that decodes visual stimuli from EEG signals recorded during natural image viewing. It handles EEG-to-image retrieval (86.30% Top-1 accuracy among 200 candidates) and EEG-to-image reconstruction (CLIP score 0.903). The method uses multi-level blurring, EVNet features, InfoNCE loss, and multi-modal CLIP alignment with SDXL-Turbo generation, demonstrating feasibility of decoding rich visual representations from EEG.

This systematic review analyzes 75 studies on self-supervised learning (SSL) for medical imaging, categorizing methods into contrastive, predictive, generative, and hybrid paradigms. It finds no universal optimal SSL strategy; performance depends on alignment between pretext task, imaging modality, and target task. Contrastive methods suit classification but may miss subtle pathology; generative approaches preserve local anatomy for segmentation; hybrid methods offer balanced performance. Practical design guidelines and open challenges are outlined.

Digital avatar watermarking faces unique challenges: avatars are routinely post-processed with background replacement, reframing, and format conversion before deployment. This paper introduces the RAW benchmark with 50 synthetic avatar videos from 5 providers and 6 attacks simulating real-world workflows. Evaluation of 7 existing methods reveals that avatar-specific attacks degrade watermark recovery. The proposed WALT method embeds watermarks in UV texture space via 3D face reconstruction, achieving 92.4% robustness to zoom and 95.6% on background removal. The benchmark is released to facilitate research.

Nano World Models is a minimalist codebase for future video prediction centered on diffusion forcing. It provides a unified interface for generative objectives, model scales, action-conditioning mechanisms, latent observation spaces, datasets, evaluation protocols, and long-horizon rollouts, enabling controlled studies of world-modeling components. Experiments across control environments, games, and real-robot data validate its effectiveness. Code, configs, and pretrained checkpoints are released for open, reproducible research.

GazeWorld is a medical imaging world model that treats the image as the world and radiologist fixation sequences as trajectories. It autoregressively predicts latent representations of fixated patches while using a spatial-completion branch for unvisited regions. At inference, it generates patch representations from the image alone without real gaze data. Frozen GazeWorld features achieve state-of-the-art diagnostic accuracy on all nine supervised settings across CheXpert, RSNA Pneumonia, and SIIM-ACR Pneumothorax, and highest zero-shot accuracy on all three benchmarks. On GazeSearch, a generic decoder trained on the same frozen features outperforms the purpose-built LogitGaze-Med by over 16% in ScanMatch and 22% in SED. The work demonstrates that modeling how experts read, not just their…

Audio large language models fail systematically when transcribing code-switching speech (English-Mandarin). Three failure modes are identified: language omission, translation-instead-of-transcription, and hallucination. Direct Preference Optimization (DPO) is applied with 100K preference pairs (570 hours), achieving MER reductions up to 89.6% in-distribution and 20.0% out-of-distribution.

AERIC is a lightweight safety monitor that reads hidden states during decoding to anticipate implicit harmful content without extra forward passes. With only 387 trainable parameters, it outperforms larger models on multiple benchmarks and increases latency by just 2.34%.

Large language models (LLMs) are increasingly used as automatic judges for summarization and dialogue evaluation. Prior work has documented biases such as position, verbosity, and style preferences, but largely focuses on outcomes. This paper asks whether LLM judges are cue-invariant, introducing a causal framework with interventions and metrics to test stability of rankings and explanations under non-evidential cue perturbations. Results show substantial cue-anchored rationalization, effectively mitigated by the PROOF-BEFORE-PREFERENCE method.

A tri-validation framework that performs explicit validation at three stages of automatic optimization modeling: semantic specification, mathematical formulation, and code generation, with a new benchmark NL4COP for combinatorial problems.

This study develops a large language model-based framework to extract segment disclosures directly from Form 10-K filings, preserving both reportable and nested segment information. A retrieval augmented system is designed to incorporate information across multiple filings to support comparability. The framework is demonstrated in longitudinal analysis within firms and cross-firm alignment of geographic segments, showing accurate extraction and effective handling of cross-period queries.

This paper presents the Multi-Persona Debate System (MPDS), a literature-grounded framework that automates scientific hypothesis generation by combining retrieval, long-context LLM reasoning, corpus-driven persona induction, and structured multi-agent debate. Evaluated on battery materials design, MPDS constructs literature snapshots of up to 500 papers, conducts three-round citation-aware debate, and produces mechanistically explicit proposals. It outperforms baselines in cross-perspective integration and shows promise as a diagnostic aid for identifying workflow bottlenecks.

Raon-Speech is a 9B-parameter speech language model for English and Korean, achieving top performance on speech understanding and generation while preserving text capabilities. Its full-duplex extension Raon-SpeechChat enables natural real-time conversation. The models are open-sourced.

This systematic review of 139 studies proposes a unified framework and meta-analysis. Results show multimodal fusion improves accuracy by 5.28% on average, multiview fusion boosts accuracy by 4.67% and F1 by 3.08%, but only a minority of studies used statistical tests, raising reproducibility concerns.

This paper studies truthful online preference aggregation for LLM fine-tuning in mobile crowdsourcing, addressing strategic misreporting by workers. It proposes a dynamic Bayesian game model and an online weighted aggregation mechanism that dynamically adjusts worker weights based on feedback accuracy, ensuring truthful feedback and achieving sublinear regret O(√T). Experiments on real-world datasets show significant performance gains.

This paper reframes proposer selection in LLM ensembles as a combinatorial selection problem akin to feature selection, emphasizing complementarity over accuracy or diversity. It explores computationally feasible greedy algorithms that assess complementarity using a small labeled set, validating complementarity as a guiding principle and identifying methods with the best performance-cost trade-offs.

Protein-ligand modeling underpins computational drug discovery. Existing benchmarks typically evaluate whether a protein and ligand interact and how strongly they bind, but provide limited evidence of whether models can localize binding sites or identify non-covalent interactions. To address this, we introduce InteractBind, a large-scale dataset of ~100k protein-ligand pairs with a benchmark for fine-grained evaluation. The core task is binding-site localization using interaction maps of six non-covalent interaction types. Evaluating eight existing models reveals limited binding-site localization despite strong binary binding prediction, with marked variation across interaction types. InteractBind encourages development of more interpretable and physically grounded models.

Proposes LLM-AutoSciLab, a closed-loop framework that combines hypothesis generation with hypothesis-conditioned experiment selection and mechanism refinement. It iteratively proposes plausible hypotheses, selects informative experiments, and updates state based on evidence. Introduces ActiveSciBench with 57 enzyme-kinetics tasks and 45 gene-regulatory-network tasks. Achieves 67.6% symbolic accuracy on NewtonBench, 35.1% on ActiveSciBench-Chem, and 31.1% exact graph recovery on ActiveSciBench-GRN, with 2-5x sample efficiency over baselines.

A study of 1,536 Gaussian release covariances finds none achieve both moderate utility and privacy against adaptive retrieval attackers, leading to a theoretical Fisher-ball lower bound and a proposal for architecture co-design.

Neural operators suffer from spectral bias, struggling to resolve high-frequency details. IRNO augments pre-trained operators with a learned refinement module via fixed-point iteration, achieving up to 56.05% error reduction on turbulent flow and substantial high-frequency error suppression.

This paper introduces a framework called Verifiable Transformers that converts task-localized Transformer circuits into bounded, solver-checkable claims. It employs direct verification for exactly encodable operators and surrogate-mediated verification for complex ones, demonstrating exhaustive verification on small symbolic tasks and applicability at GPT-2 scale, aiming to formalize mechanistic circuit explanations into provable or refutable propositions.

This paper introduces CAFD, a learning-based approach that integrates model-based signals, distance features, and a novel Concept Failure Ratio (CFR) feature extracted via Vision-Language Models to achieve superior fault detection performance while maintaining efficiency, with an average 18.3% FDR improvement over state-of-the-art baselines.

This paper proposes a novel framework called MODIAD for multimodal online distributed industrial anomaly detection. It formulates a Multi-class Intelligent Scheduling (MIS) problem and designs a Sequential Marginal Gain Greedy (SMG) algorithm to solve it efficiently. Furthermore, a Resource Efficient Class-Wise Low Rank Adaptation (REC-LoRA) strategy reduces training overhead. Experiments on MVTec 3D-AD and Eyecandies datasets demonstrate superior performance and efficiency.

The paper introduces 'algometrics', a framework for time series forecasting when predictive models influence the data they aim to predict. It distinguishes historical risk from deployment risk and proves three key results: deployment risk is not identifiable from passive historical data alone; historical model rankings can invert due to crowding; and randomized actions can identify short-horizon linear feedback with a finite-sample bound. The findings suggest that benchmarks in algorithmic markets should report feedback sensitivity alongside predictive accuracy.