Learned world models hold significant potential for robotic manipulation, as they can serve as simulator for real-world interactions. While extensive progress has been made in 2D video-based world models, these approaches often lack geometric and spatial reasoning, which is essential for capturing the physical structure of the 3D world. To address this limitation, we introduce iMoWM, a novel interactive world model designed to generate color images, depth maps, and robot arm masks in an autoregressive manner conditioned on actions. To overcome the high computational cost associated with three-dimensional information, we propose MMTokenizer, which unifies multi-modal inputs into a compact token representation. This design enables iMoWM to leverage large-scale pretrained VideoGPT models while maintaining high efficiency and incorporating richer physical information. With its multi-modal representation, iMoWM not only improves the visual quality of future predictions but also serves as an effective simulator for model-based reinforcement learning (MBRL) and facilitates real-world imitation learning. Extensive experiments demonstrate the superiority of iMoWM across these tasks, showcasing the advantages of multi-modal world modeling for robotic manipulation. Homepage: https://xingyoujun.github.io/imowm/

Accurate perception is critical for vehicle safety, with LiDAR as a key enabler in autonomous driving. To ensure robust performance across environments, sensor types, and weather conditions without costly re-annotation, domain generalization in LiDAR-based 3D semantic segmentation is essential. However, LiDAR annotations are often noisy due to sensor imperfections, occlusions, and human errors. Such noise degrades segmentation accuracy and is further amplified under domain shifts, threatening system reliability. While noisy-label learning is well-studied in images, its extension to 3D LiDAR segmentation under domain generalization remains largely unexplored, as the sparse and irregular structure of point clouds limits direct use of 2D methods. To address this gap, we introduce the novel task Domain Generalization for LiDAR Semantic Segmentation under Noisy Labels (DGLSS-NL) and establish the first benchmark by adapting three representative noisy-label learning strategies from image classification to 3D segmentation. However, we find that existing noisy-label learning approaches adapt poorly to LiDAR data. We therefore propose DuNe, a dual-view framework with strong and weak branches that enforce feature-level consistency and apply cross-entropy loss based on confidence-aware filtering of predictions. Our approach shows state-of-the-art performance by achieving 56.86% mIoU on SemanticKITTI, 42.28% on nuScenes, and 52.58% on SemanticPOSS under 10% symmetric label noise, with an overall Arithmetic Mean (AM) of 49.57% and Harmonic Mean (HM) of 48.50%, thereby demonstrating robust domain generalization in DGLSS-NL tasks. The code is available on our project page.

Ample empirical evidence in deep neural network training suggests that a variety of optimizers tend to find nearly global optima. In this article, we adopt the reversed perspective that convergence to an arbitrary point is assumed rather than proven, focusing on the consequences of this assumption. From this viewpoint, in line with recent advances on the edge-of-stability phenomenon, we argue that different optimizers effectively act as eigenvalue filters determined by their hyperparameters. Specifically, the standard gradient descent method inherently avoids the sharpest minima, whereas Sharpness-Aware Minimization (SAM) algorithms go even further by actively favoring wider basins. Inspired by these insights, we propose two novel algorithms that exhibit enhanced eigenvalue filtering, effectively promoting wider minima. Our theoretical analysis leverages a generalized Hadamard--Perron stable manifold theorem and applies to general semialgebraic $C^2$ functions, without requiring additional non-degeneracy conditions or global Lipschitz bound assumptions. We support our conclusions with numerical experiments on feed-forward neural networks.

Recent work suggests that large language models (LLMs) encode factuality signals in their internal representations, such as hidden states, attention weights, or token probabilities, implying that LLMs may "know what they don't know". However, LLMs can also produce factual errors by relying on shortcuts or spurious associations. These error are driven by the same training objective that encourage correct predictions, raising the question of whether internal computations can reliably distinguish between factual and hallucinated outputs. In this work, we conduct a mechanistic analysis of how LLMs internally process factual queries by comparing two types of hallucinations based on their reliance on subject information. We find that when hallucinations are associated with subject knowledge, LLMs employ the same internal recall process as for correct responses, leading to overlapping and indistinguishable hidden-state geometries. In contrast, hallucinations detached from subject knowledge produce distinct, clustered representations that make them detectable. These findings reveal a fundamental limitation: LLMs do not encode truthfulness in their internal states but only patterns of knowledge recall, demonstrating that "LLMs don't really know what they don't know".

The performance of Large Language Models (LLMs) is highly sensitive to the prompts they are given. Drawing inspiration from the field of prompt optimization, this study investigates the potential for enhancing Automated Essay Scoring (AES) by refining the scoring rubrics used by LLMs. Specifically, our approach prompts models to iteratively refine rubrics by reflecting on models' own scoring rationales and observed discrepancies with human scores on sample essays. Experiments on the TOEFL11 and ASAP datasets using GPT-4.1, Gemini-2.5-Pro, and Qwen-3-Next-80B-A3B-Instruct show Quadratic Weighted Kappa (QWK) improvements of up to 0.19 and 0.47, respectively. Notably, even with a simple initial rubric, our approach achieves comparable or better QWK than using detailed human-authored rubrics. Our findings highlight the importance of iterative rubric refinement in LLM-based AES to enhance alignment with human evaluations.

How should we evaluate the robustness of language model defenses? Current defenses against jailbreaks and prompt injections (which aim to prevent an attacker from eliciting harmful knowledge or remotely triggering malicious actions, respectively) are typically evaluated either against a static set of harmful attack strings, or against computationally weak optimization methods that were not designed with the defense in mind. We argue that this evaluation process is flawed. Instead, we should evaluate defenses against adaptive attackers who explicitly modify their attack strategy to counter a defense's design while spending considerable resources to optimize their objective. By systematically tuning and scaling general optimization techniques-gradient descent, reinforcement learning, random search, and human-guided exploration-we bypass 12 recent defenses (based on a diverse set of techniques) with attack success rate above 90% for most; importantly, the majority of defenses originally reported near-zero attack success rates. We believe that future defense work must consider stronger attacks, such as the ones we describe, in order to make reliable and convincing claims of robustness.

State-of-the-art (SOTA) LLMs have progressed from struggling on proof-based Olympiad problems to solving most of the IMO 2025 problems, with leading systems reportedly handling 5 of 6 problems. Given this progress, we assess how well these models can grade proofs: detecting errors, judging their severity, and assigning fair scores beyond binary correctness. We study proof-analysis capabilities using a corpus of 90 Gemini 2.5 Pro-generated solutions that we grade on a 1-4 scale with detailed error annotations, and on MathArena solution sets for IMO/USAMO 2025 scored on a 0-7 scale. Our analysis shows that models can reliably flag incorrect (including subtly incorrect) solutions but exhibit calibration gaps in how partial credit is assigned. To address this, we introduce agentic workflows that extract and analyze reference solutions and automatically derive problem-specific rubrics for a multi-step grading process. We instantiate and compare different design choices for the grading workflows, and evaluate their trade-offs. Across our annotated corpus and MathArena, our proposed workflows achieve higher agreement with human grades and more consistent handling of partial credit across metrics. We release all code, data, and prompts/logs to facilitate future research.

De novo ligand design is a fundamental task that seeks to generate protein or molecule candidates that can effectively dock with protein receptors and achieve strong binding affinity entirely from scratch. It holds paramount significance for a wide spectrum of biomedical applications. However, most existing studies are constrained by the \textbf{Pseudo De Novo}, \textbf{Limited Docking Modeling}, and \textbf{Inflexible Ligand Type}. To address these issues, we propose MagicDock, a forward-looking framework grounded in the progressive pipeline and differentiable surface modeling. (1) We adopt a well-designed gradient inversion framework. To begin with, general docking knowledge of receptors and ligands is incorporated into the backbone model. Subsequently, the docking knowledge is instantiated as reverse gradient flows by binding prediction, which iteratively guide the de novo generation of ligands. (2) We emphasize differentiable surface modeling in the docking process, leveraging learnable 3D point-cloud representations to precisely capture binding details, thereby ensuring that the generated ligands preserve docking validity through direct and interpretable spatial fingerprints. (3) We introduce customized designs for different ligand types and integrate them into a unified gradient inversion framework with flexible triggers, thereby ensuring broad applicability. Moreover, we provide rigorous theoretical guarantees for each component of MagicDock. Extensive experiments across 9 scenarios demonstrate that MagicDock achieves average improvements of 27.1\% and 11.7\% over SOTA baselines specialized for protein or molecule ligand design, respectively.

Large vision-language models (LVLMs), which integrate a vision encoder (VE) with a large language model, have achieved remarkable success across various tasks. However, there are still crucial challenges in LVLMs such as object hallucination, generating descriptions of objects that are not in the input image. Here, we argue that uncertain visual tokens within the VE is a key factor that contributes to object hallucination. Our statistical analysis found that there are positive correlations between visual tokens with high epistemic uncertainty and the occurrence of hallucinations. Furthermore, we show theoretically and empirically that visual tokens in early VE layers that exhibit large representation deviations under small adversarial perturbations indicate high epistemic uncertainty. Based on these findings, we propose a simple yet effective strategy to mitigate object hallucination by modifying the VE only. Our method comprises a proxy method with adversarial perturbations for identifying uncertain visual tokens efficiently and a method to mask these uncertain visual tokens during the self-attention process in the middle layers of the VE, suppressing their influence on visual encoding and thus alleviating hallucinations. Extensive experiments show that our method significantly reduces object hallucinations in LVLMs and can synergistically work with other prior arts.

Large language models (LLMs) exhibit remarkable generative capabilities but raise ethical and security concerns by memorizing sensitive data, reinforcing biases, and producing harmful content. These risks have spurred interest in LLM unlearning, the task of removing knowledge associated with undesirable data from pre-trained models. However, most existing methods assume access to clean, well-defined forget data samples, whereas real-world forget data could often be low-quality, synthetically rewritten, or watermarked, casting doubt on the reliability of unlearning. This work presents the first study of unlearning under perturbed or low-fidelity forget data, referred to as noisy forget sets. By systematically benchmarking state-of-the-art LLM unlearning methods, RMU and NPO, on such noisy forget sets, we find that unlearning remains surprisingly robust to perturbations, provided that core semantic signals are preserved. To explain this robustness, we propose a saliency-based interpretation: key semantic components that drive forgetting remain consistently influential despite substantial variation in surface form. This suggests that unlearning algorithms are primarily guided by deep semantic cues rather than shallow lexical patterns.