Language-driven dexterous grasp models like DextER fail systematically when instructions specify where not to grasp. We introduce an inference-time framework combining Sequential Monte Carlo with classifier-free guidance to handle such negations without extra training data. A frozen 3D grounding model localizes forbidden parts. We build the NegGrasp benchmark with paired positive/negative instructions and constraint-aware metrics. Our method reduces violation rate from 57.9% to 17.2% while improving success rates.

Researchers present a framework that allows quadruped robots to exploit their nonlinear dynamics to generate multiple natural gaits without dedicated controllers. By identifying six nonlinear normal modes (NNMs) on the eBert robot and using black-box optimization to set step length, they show how each NNM evolves into a distinct gait with different speeds. Experiments demonstrate that eBert can use its mechanics for task-specific movements, potentially paving the way for new agile and efficient robots leveraging embodied intelligence.

This research introduces a markerless interaction framework and a dedicated cross-device dataset to study cross-device alignment in gaze-based assistive manipulation. It proposes Graph-based Reference Selection and the Gaze-Surface Intersection Error (GSIE), showing that alignment methods ranked highly by conventional pose metrics are not always optimal in task-level performance.

Researchers developed WM-RMoE, a world model-based framework that enhances autonomous overtaking safety by evaluating cumulative trajectory risks via latent dynamics model rollouts, coordinating specialized safety experts with hierarchical gating, and preserving multi-modal maneuvers with Gaussian Mixture Models. It outperforms baselines in safety, stability, and generalization.

Researchers propose a method to automatically extract critical driving scenarios from egocentric sensor data using scene graphs and Linear Temporal Logic (LTL). It processes sensor data and HD maps to create scene graphs, uses LTL to specify scenarios, and applies a model checker to identify instances. Validated on Argoverse 2 datasets and dashcam videos, it aids scenario-based coverage analysis for ADS.

This paper combines algebraic and geometric approaches for kinematic analysis of generic 3R robots. The authors use an inverse kinematic model based on conformal geometric algebra, propose a generic framework to study conditions for a 3R robot to have four inverse kinematic solutions (IKS), and analyze the distribution of these solutions in the workspace. They provide necessary and sufficient conditions for a class of orthogonal robots to have four IKS.

This paper investigates the inverse kinematics of generic 3R robots using Conformal Geometric Algebra (CGA). By reformulating the problem as the intersection of two circles, the authors directly derive a univariate polynomial in θ₂ without eliminating θ₁ and θ₃. This approach provides deeper insights into kinematic properties and simplifies the solution process. The work is published in the proceedings of IMA 2025.

StreamScout is an adaptive inference framework for streaming video understanding that maintains a lightweight textual timeline and progressively augments it with up to three visual views of increasing detail, only escalating when needed. This reduces inference cost and token consumption while improving accuracy. On OVO-Bench, StreamScout-S improves Qwen3-VL-8B by 14.65 accuracy points while using 59% fewer tokens than uniform sampling and achieving an average response time of 1.04 seconds.

Most advances in keypoint descriptors address monomodal settings where changes are due to viewpoint, illumination, or contrast. Multimodal scenarios—such as multispectral imaging, RGB-depth, satellite, or medical imaging—create images from fundamentally different sensing processes, making the same structures appear differently. Common solutions either train descriptors for each modality pair (requiring retraining) or use large models (increasing runtime). CrossFeat proposes a framework that adapts an existing monomodal descriptor to work across modalities by learning a crossing function in descriptor space. It also introduces geometry-appearance disentanglement to preserve structural information, altering only appearance while keeping geometric properties. Experiments across multiple doma…

This paper introduces decision-level driving affordance prediction, and presents the CoLT-Drive benchmark with 3,536 counterfactual samples to evaluate models' ability to predict plausible driving actions when rare objects appear. It also proposes the KPA framework, combining structured perception-to-decision prompting, SLERP-based expert merging, and a RegMoE module, to adapt small VLMs while preserving open-world knowledge. Experiments show KPA achieves 60.8% pair accuracy on CoLT-Drive, outperforming the pretrained baseline (50.3%) and LoRA SFT (32.4%), while maintaining competitive in-domain accuracy.

Multimodal large language models (MLLMs) excel on OCR document understanding and text-rich visual reasoning benchmarks, but existing evaluations often assume all tasks are valid and answerable. In reality, questions may rely on illegible text, occluded evidence, nonexistent visual targets, contradictory premises, or missing variables. This paper introduces VeriOCRBench, a human-verified benchmark with 1,800 samples, including 1,600 trap-embedded invalid tasks and 200 trap-free controls, to evaluate task verification, root-cause diagnosis, and over-refusal. Evaluation of 15 leading MLLMs reveals persistent blind compliance, diagnosis failures, and prompt-induced over-refusal, highlighting a critical reliability gap in current OCR reasoning systems.

A new paper challenges the prevailing notion that object hallucination in multimodal large language models (MLLMs) is primarily driven by language priors, presenting evidence for an alternative cause: visual-origin hallucination, stemming from faulty visual feature extraction and image-text embedding misalignment. The authors introduce Adversarial Contrastive Fine-Tuning (ACFT), which requires only 0.9% of the COCO dataset and adds zero inference overhead, achieving state-of-the-art performance on hallucination benchmarks across models.

A new paper identifies a compositional safety blind spot in multimodal large language models used for video moderation: videos composed of individually benign components can communicate harmful meaning when interpreted as a whole. The authors coin the term Distributed Implicit Harm (DIH) and study two forms: temporally distributed harm across visual segments and cross-modal harm between audio and visual streams. They develop a multi-agent synthesis framework to create a dataset of over 9,000 DIH videos, and benchmark over 30 MLLMs, finding significant and consistent deficits in detecting both forms. Evaluations on real-world social media videos confirm the same failure mode, indicating a practical challenge for video moderation.

This paper presents an alternative derivation of explicit-time generative models like Flow Matching and Rectified Flow from a Lagrangian (particle-centric) perspective. By analyzing the local Taylor expansion of a continuous denoiser, the authors motivate a strict invariance condition (conservation of target identity) that leads to a quasi-linear advection PDE. Solving this PDE via the Method of Characteristics analytically yields the straight-line trajectories of Flow Matching, identifying the denoiser's Jacobian as the primary source of trajectory curvature and explaining why straight-line flows allow huge step sizes and why distillation is needed.

This paper introduces ZimaBlue, a scalable framework for learning generalizable World Action Models (WAMs) from large-scale video. It employs a three-stage training curriculum: causal embodied video pre-training on large-scale human and robot egocentric videos, video-action mid-training with unified action representation, and robot-specific specialization. Its asynchronous Slow-Fast dual-system architecture enables 30 Hz real-time action prediction on an NVIDIA RTX 4090. In real-robot zero-shot evaluations, scaling from target-robot data alone to over 120,000 hours of embodied video improves success from 36.1% to 77.8%, with particularly strong gains on unseen tasks.

Qwen-Drive-1.0 is a vision-language foundation model for autonomous driving that retains the architecture of a pretrained VLM and integrates 3D perception, visual question answering, and motion planning. An external bird's-eye-view perception head performs 3D detection, occupancy prediction, and map segmentation, while a Planning Expert generates future trajectories. Experiments show strong 3D perception and driving scene understanding with largely preserved general vision-language capability, and competitive motion planning across open, pseudo-closed, and closed-loop settings.

This paper proposes a cone-constrained bilinear decomposition for the total scaled-gradient variation (TSGV) regularizer, addressing computational challenges in image restoration. The method decouples the nonlinear weighted gradient, leading to an equivalent optimization with cone or sphere constraints. Using alternating minimization with a majorization-minimization strategy, it ensures monotonic decrease in energy and global convergence. Experiments on Gaussian denoising and NLOS imaging show superior or competitive PSNR and SSIM, especially at high noise levels.

A new arXiv paper introduces an open-source library, 'Scientific Agent Skills', containing 163 procedural skills across 16 practice areas to improve the defensibility of language model agents in scientific analysis. The library provides versioned, human-readable instruction files that agents load only when needed, along with reference materials and runnable scripts. The authors report no task-level evaluation and no host selection rate.

This preliminary small-scale study evaluates the accuracy of large language models (LLMs) in verifying medical causal claims and supporting them with peer-reviewed research. Assessing eight LLMs on 17 medical causal hypotheses, the study finds that while models perform well at retrieving relevant information (high recall), they often fail to provide valid scientific articles for support and to reject unsupported hypotheses. The work underscores the need for rigorous evaluation before using LLMs for search and retrieval in healthcare settings.

Data contamination undermines the reliable evaluation of large language models (LLMs) in mathematical problem solving. While rewriting-based evaluation mitigates memorization, existing methods lack guarantees of problem validity and answer correctness. We propose Proof-Verified Benchmark Rewriting (RePro), the first framework to integrate Lean-oriented neural automated theorem provers (ATPs) into benchmark rewriting, which rewrites problems and regenerates answers with correctness ensured by Lean-verified proofs. Experiments on GSM8K and MATH show that RePro's retained rewritten instances achieve 100% well-definedness, feasibility, and answer correctness, while existing methods still produce invalid or incorrect instances. Moreover, several models exhibit accuracy drops on proof-verified…

CUDA-Harness is a new framework for generating and optimizing high-performance CUDA kernels from natural language. It introduces Intermediate-Structured Generation to connect high-level semantics with low-level kernel generation, uses Synthesis-Based Verification to mitigate reward hacking, and proposes Feedback-Adaptive Evolution to prioritize correctness while optimizing performance. Experiments show its effectiveness and generalization across LLMs, hardware, and C-to-CUDA transpilation.

Social media user behavior is shaped not only by content and interactions but also by the moral value signals they express. Yet existing user representation methods largely miss this dimension. To address this, researchers propose ValueGraph, a graph pre-training framework that uses automatically inferred moral-value signals as noisy auxiliary signals. It learns semantic and structural representations from post-reply graphs and aligns users via contrastive and clustering objectives based on relative value similarity. Experiments on stance detection and Twitter bot detection show consistent gains over strong text-based, graph-based, and text-only LLM baselines, highlighting value-signal guidance as a useful inductive bias for socially informed user modeling. The study was submitted to arXi…

Self-supervised respiratory encoders lack clinical semantic grounding for zero-shot inference. This research proposes a framework aligning these encoders with medical terminology in a shared latent space, turning them into zero-shot-capable foundation models. Using a medical LLM to synthesize structured reports from metadata addresses data scarcity. Across 9 tasks on 6 datasets, the method achieves a 61.3% mean zero-shot AUC, surpassing CLAP (51.4%) and Qwen2-Audio (54.9%), and reaches the highest linear probing AUC (71.6%) with only 43% of data used by full-scale baselines.

A new study uncovers the internal circuits behind LLM safety behavior, identifying three stages: harmful detection heads, safety neurons, and refusal heads. By using circuit-guided weight scaling, they improved safety rates under adversarial attacks by 26.5% across six LLMs, with only a 1.7% accuracy drop. Accepted to EMNLP 2026 Findings.

GUI world models are often tested as one-step predictors, but their real purpose is to support multi-step agent interactions. This paper introduces GUI-CC, a benchmark that evaluates these models as environments, with offline and online tasks. Results show that generating plausible screens doesn't guarantee reliable simulation; models often fail to maintain task-relevant context or support executable multi-step rollouts.

This paper reveals a critical blind spot in outcome-only evaluation of LLM agents: it cannot detect agents that reach the correct answer through wrong methods. In a controlled environment with fault injection, the outcome-only judge catches 84% of loud faults but only 45% of silent ones, while flagging 33% of correct trajectories. A step-rubric judge achieves 77% silent recall with zero false alarms at 3x the cost. Moreover, no judge catches an invented promise appended to a perfect trajectory, and self-consistency triples cost without improvement. The authors advocate for stratifying recall by outcome survival and release the environment and analysis pipeline.

A new paper proposes a framework for extracting user profiles from authentic anonymized social media posts, capturing nuanced preferences better than stereotyped synthetic profiles. Evaluated in supervised fine-tuning and test-time multi-perspective reasoning, behaviorally grounded profiles consistently outperform synthetic profile baselines on complex recommendation and open-ended query benchmarks, offering a foundation for next-generation personalized language systems.

The paper introduces a hypothesis-testing framework based on Weight of Evidence (WoE) to evaluate feature importance methods (FIMs). It quantifies how strongly observed evidence supports a given hypothesis on feature importance, enabling principled evaluation of alignment and stability. Experiments with LIME and SHAP demonstrate its utility.

A study compares nine variants from five foundation model families in zero-shot mode against two state-of-the-art electricity price forecasting benchmarks in Germany, Poland, and Spain from 2021 to 2025. Only TabPFN models consistently outperform the benchmarks statistically, but their economic advantage is limited: they perform best under unlimited bids and riskier quantile-based strategies, while a Distributional Deep Neural Network benchmark is more profitable when risk tolerance is lower. Thus, foundation models cannot universally replace market-specific models; their value depends on architecture and the decision problem.

This paper introduces a recursive formula to efficiently compute the stochastic complexity (shortest code length) of vectors with cluster structure under the Normalized Maximum Likelihood (NML) model. The proposed approach reduces the time complexity from polynomial to linear, benefiting data clustering based on the Minimum Description Length (MDL) principle.