Kate Moussouris confirms that the 'jailbreak' which got Claude Fable 5 banned under export control was actually its ability to fix code. Experts warn that preventing AI from fixing bugs weakens defense, and non-technical decision-makers may ban models that help secure code based on misunderstanding.

LV-Calib is a calibration framework for LiDAR-camera extrinsic estimation and LiDAR boundary-response calibration using a printable planar target. It leverages visual fiducials and circular reflectivity boundaries to extract features, and iteratively refines LiDAR 3D points to handle transition band distortions caused by finite beam footprint and mixed-intensity returns. Experiments demonstrate sub-pixel reprojection accuracy, millimeter-level LiDAR feature consistency, and improved odometry performance. Code and data will be released.

ViTaL framework uses bi-level optimization combining vision and touch to steer generative robot policies at inference time, significantly improving success in contact-rich manipulation tasks.

A new study reveals that operators maintain task performance in contact-rich human-robot interaction under varying environmental conditions by increasing physiological effort, particularly autonomic workload, to suppress thermal discomfort. The findings motivate physiology-aware control architectures.

This paper proposes DynaHMRC, a decentralized framework where each robot acts as a role-aware LLM agent, enabling flexible collaboration through a four-stage closed-loop process: self-description, task allocation with leadership bidding, leader election, and reflective execution. It addresses scalability issues of centralized LLM schedulers, insufficient handling of dynamic tasks, and domain-specific data scarcity. Experiments show higher success rates and efficiency compared to strong baselines, with promising scalability.

VANDERER is a novel exploration framework for mobile agents that uses only monocular camera data, leveraging a Visual Curiosity Module to guide pre-trained diffusion policies. It outperforms NoMaD by 13.4% in explored area across simulated environments and reveals a correlation between visual and geometric curiosity in outdoor settings.

TacStyle proposes a novel approach that learns a structured latent representation of user preferences and uses a foundation model to interpret this space to select desired behaviors, enabling fine-grained control over robot behavior with significantly fewer preference labels. Experiments in simulation and real-world demonstrate more precise adaptation.

Researchers propose a pipeline that uses synthetic data to train a Transformer-based model for safe UAV landing zone detection, bridging the sim-to-real gap and eliminating manual annotation.

A new framework called Steady-Forcing addresses the trade-off between spatial stability and motion continuity in long-horizon fixed-camera nature video generation. It uses components like visual anchors (V-Sink), motion memory (EMA-Sink), and temporal encoding to improve background consistency and fluid dynamics over multi-minute autoregressive rollouts.

This paper proposes BBR-Net, a boundary-balanced replay network that selects samples using boundary-aware priority and class balance to preserve anatomical structure. Evaluated on CAMUS and CardiacNet, forward setting retains source performance and reduces catastrophic forgetting, while reverse setting reveals failure when initial representations are noisy. Structural perturbation analysis shows replay effectiveness depends on stored structural reliability.

Vision-language models are crucial in robotics and other domains, but output uncertainty quantification remains essential. FUSE proposes a probabilistic framework combining aleatoric (data-ambiguity) and epistemic (model-diversity) uncertainties via Bayesian fusion into a scalar measure that predicts output correctness, achieving state-of-the-art calibration.

This paper argues that convolution and self-attention can be unified within a k-nearest neighbor aggregation framework, and introduces ConvNN, a unified framework that exactly recovers standard convolution and self-attention, enabling exploration of the continuous spectrum between local and global aggregation.

This study enhances maritime security using AI and computer vision, comparing six deep learning architectures on 6,468 images. The Vision Transformer achieved 100% accuracy with the lowest error rates and fastest processing, demonstrating AI's potential for surveillance, border protection, and autonomous navigation.

RAMS is a lightweight runtime controller for object detection on embedded edge devices, dynamically selecting among YOLOv8 tiers (NANO/SMALL/MEDIUM) based on resource pressure and detection results. On Jetson Orin, the safety2 policy achieves 3.41 ms mean latency (5.6x faster than fixed MEDIUM) while retaining 74% proxy accuracy. Detection-conditioned switching improves SWAS by up to 47.3% under heavy load.

ReportQA is a clinical and flexible radiology report evaluation framework using QA pairs and LLM-as-judge to compute QAScore, addressing limitations of existing NLG and CE metrics. Experiments show QAScore aligns better with radiologist judgments, and question-driven inference outperforms report-based paradigms.

This study evaluates LSTM, TCN, and Transformer models on the WESAD dataset for multimodal emotion recognition using wrist and chest sensors. Transformer achieved highest accuracy in multimodal settings, TCN best on wrist-only. Ensemble late-fusion achieved 98.91% accuracy and 98.56% macro-F1.

This study re-evaluates online augmentation modules for web agents under a fixed inference budget, finding that AWM, ASI, and ReasoningBank do not provide significant advantages over a token-matched vanilla baseline that uses the same budget for additional actor steps. The baseline matches or surpasses all augmentation methods in success rate while often using fewer total tokens. The effect extends to enterprise knowledge-work tasks, and run-to-run variance is highlighted as a crucial evaluation metric.

Nemotron 3 Ultra is a 550B total/55B active parameter hybrid Mamba-Transformer MoE model by NVIDIA. Pre-trained on 20T tokens with 1M context, it achieves 6x higher throughput than open LLMs with on-par accuracy. Open-sourced on HuggingFace.

Introduces CoRA, a GRPO-based reinforcement learning framework that aligns model confidence with generated rationales, reducing misleading high-confidence answers. On datasets like MedQA, it reduces alignment error by up to 26.51% while maintaining accuracy and improving calibration.

Causal Transformers, while efficient for left-to-right decoding, struggle with arbitrary conditionals (e.g., text conditioned on both past and future tokens). The proposed AC-GPT introduces a simple modification to standard causal Transformers, enabling evaluation and sampling from arbitrary conditionals—including past, future, and mixed contexts—in a single forward pass. It preserves left-to-right ordering and next-token prediction, allowing fine-tuning of existing LLMs. Empirical results show it outperforms baselines on arbitrary conditional modeling without degrading standard performance.

Researchers propose 'Telegraph English,' a readable symbolic format for compressing context in multi-hop question answering. It outperforms matched-budget baselines, including coherent summaries, by 13-20 F1 points. The advantage does not increase with reasoning depth.

This paper presents the first study on the robustness of proof autoformalization models in Lean 4. It introduces global and local perturbations to test model faithfulness, finding that all seven evaluated models are sensitive to changes and mostly fail to reflect local perturbations faithfully.