Refusal training on Large Language Models (LLMs) prevents harmful outputs, yet this defense remains vulnerable to both automated and human-crafted jailbreaks. We present a novel LLM-as-red-teamer approach in which a human jailbreaks a refusal-trained LLM to make it willing to jailbreak itself or other LLMs. We refer to the jailbroken LLMs as $J_2$ attackers, which can systematically evaluate target models using various red teaming strategies and improve its performance via in-context learning from the previous failures. Our experiments demonstrate that Sonnet 3.5 and Gemini 1.5 pro outperform other LLMs as $J_2$, achieving 93.0% and 91.0% attack success rates (ASRs) respectively against GPT-4o (and similar results across other capable LLMs) on Harmbench. Our work not only introduces a scalable approach to strategic red teaming, drawing inspiration from human red teamers, but also highlights jailbreaking-to-jailbreak as an overlooked failure mode of the safeguard. Specifically, an LLM can bypass its own safeguards by employing a jailbroken version of itself that is willing to assist in further jailbreaking. To prevent any direct misuse with $J_2$, while advancing research in AI safety, we publicly share our methodology while keeping specific prompting details private.

This paper introduces MobileH2R, a framework for learning generalizable vision-based human-to-mobile-robot (H2MR) handover skills. Unlike traditional fixed-base handovers, this task requires a mobile robot to reliably receive objects in a large workspace enabled by its mobility. Our key insight is that generalizable handover skills can be developed in simulators using high-quality synthetic data, without the need for real-world demonstrations. To achieve this, we propose a scalable pipeline for generating diverse synthetic full-body human motion data, an automated method for creating safe and imitation-friendly demonstrations, and an efficient 4D imitation learning method for distilling large-scale demonstrations into closed-loop policies with base-arm coordination. Experimental evaluations in both simulators and the real world show significant improvements (at least +15% success rate) over baseline methods in all cases. Experiments also validate that large-scale and diverse synthetic data greatly enhances robot learning, highlighting our scalable framework.

In light of the inherently complex and dynamic nature of real-world environments, incorporating risk measures is crucial for the robustness evaluation of deep learning models. In this work, we propose a Risk-Averse Certification framework for Bayesian neural networks called RAC-BNN. Our method leverages sampling and optimisation to compute a sound approximation of the output set of a BNN, represented using a set of template polytopes. To enhance robustness evaluation, we integrate a coherent distortion risk measure--Conditional Value at Risk (CVaR)--into the certification framework, providing probabilistic guarantees based on empirical distributions obtained through sampling. We validate RAC-BNN on a range of regression and classification benchmarks and compare its performance with a state-of-the-art method. The results show that RAC-BNN effectively quantifies robustness under worst-performing risky scenarios, and achieves tighter certified bounds and higher efficiency in complex tasks.

Due to the inherent lack of transparency in deep neural networks, it is challenging for deep reinforcement learning (DRL) agents to gain trust and acceptance from users, especially in safety-critical applications such as medical diagnosis and military operations. Existing methods for explaining an agent's decision either require to retrain the agent using models that support explanation generation or rely on perturbation-based techniques to reveal the significance of different input features in the decision making process. However, retraining the agent may compromise its integrity and performance, while perturbation-based methods have limited performance and lack knowledge accumulation or learning capabilities. Moreover, since each perturbation is performed independently, the joint state of the perturbed inputs may not be physically meaningful. To address these challenges, we introduce $\textbf{VisionMask}$, a standalone explanation model trained end-to-end to identify the most critical regions in the agent's visual input that can explain its actions. VisionMask is trained in a self-supervised manner without relying on human-generated labels. Importantly, its training does not alter the agent model, hence preserving the agent's performance and integrity. We evaluate VisionMask on Super Mario Bros (SMB) and three Atari games. Compared to existing methods, VisionMask achieves a 14.9% higher insertion accuracy and a 30.08% higher F1-Score in reproducing original actions from the selected visual explanations. We also present examples illustrating how VisionMask can be used for counterfactual analysis.

Inferring affordable (i.e., graspable) parts of arbitrary objects based on human specifications is essential for robots advancing toward open-vocabulary manipulation. Current grasp planners, however, are hindered by limited vision-language comprehension and time-consuming 3D radiance modeling, restricting real-time, open-vocabulary interactions with objects. To address these limitations, we propose GLOVER, a unified Generalizable Open-Vocabulary Affordance Reasoning framework, which fine-tunes the Large Language Models (LLMs) to predict visual affordance of graspable object parts within RGB feature space. We compile a dataset of over 10,000 images from human-object interactions, annotated with unified visual and linguistic affordance labels, to enable multi-modal fine-tuning. GLOVER inherits world knowledge and common-sense reasoning from LLMs, facilitating more fine-grained object understanding and sophisticated tool-use reasoning. To enable effective real-world deployment, we present Affordance-Aware Grasping Estimation (AGE), a non-parametric grasp planner that aligns the gripper pose with a superquadric surface derived from affordance data. In evaluations across 30 real-world scenes, GLOVER achieves success rates of 86.0% in part identification and 76.3% in grasping, with speeds approximately 330 times faster in affordance reasoning and 40 times faster in grasping pose estimation than the previous state-of-the-art.

For safety reasons, large language models (LLMs) are trained to refuse harmful user instructions, such as assisting dangerous activities. We study an open question in this work: does the desired safety refusal, typically enforced in chat contexts, generalize to non-chat and agentic use cases? Unlike chatbots, LLM agents equipped with general-purpose tools, such as web browsers and mobile devices, can directly influence the real world, making it even more crucial to refuse harmful instructions. In this work, we primarily focus on red-teaming browser agents, LLMs that manipulate information via web browsers. To this end, we introduce Browser Agent Red teaming Toolkit (BrowserART), a comprehensive test suite designed specifically for red-teaming browser agents. BrowserART is consist of 100 diverse browser-related harmful behaviors (including original behaviors and ones sourced from HarmBench [Mazeika et al., 2024] and AirBench 2024 [Zeng et al., 2024b]) across both synthetic and real websites. Our empirical study on state-of-the-art browser agents reveals that, while the backbone LLM refuses harmful instructions as a chatbot, the corresponding agent does not. Moreover, attack methods designed to jailbreak refusal-trained LLMs in the chat settings transfer effectively to browser agents. With human rewrites, GPT-4o and o1-preview-based browser agents attempted 98 and 63 harmful behaviors (out of 100), respectively. We publicly release BrowserART and call on LLM developers, policymakers, and agent developers to collaborate on improving agent safety

Diffusion models demonstrate superior performance in capturing complex distributions from large-scale datasets, providing a promising solution for quadrupedal locomotion control. However, offline policy is sensitive to Out-of-Distribution (OOD) states due to the limited state coverage in the datasets. In this work, we propose a two-stage learning framework combining offline learning and online preference alignment for legged locomotion control. Through the offline stage, the diffusion planner learns the joint distribution of state-action sequences from expert datasets without using reward labels. Subsequently, we perform the online interaction in the simulation environment based on the trained offline planer, which significantly addresses the OOD issues and improves the robustness. Specifically, we propose a novel weak preference labeling method without the ground-truth reward or human preferences. The proposed method exhibits superior stability and velocity tracking accuracy in pacing, trotting, and bounding gait under both slow- and high-speed scenarios and can perform zero-shot transfer to the real Unitree Go1 robots. The project website for this paper is at https://shangjaven.github.io/preference-aligned-diffusion-legged/.

In real-world scenarios, the impacts of decisions may not manifest immediately. Taking these delays into account facilitates accurate assessment and management of risk in real-world environments, thereby ensuring the efficacy of strategies. In this paper, we investigate risk-averse learning using Conditional Value at Risk (CVaR) as risk measure, while incorporating delayed feedback with unknown but bounded delays. We develop two risk-averse learning algorithms that rely on one-point and two-point zeroth-order optimization approaches, respectively. The regret achieved by the algorithms is analyzed in terms of the cumulative delay and the number of total samplings. The results suggest that the two-point risk-averse learning achieves a smaller regret bound than the one-point algorithm. Furthermore, the one-point risk-averse learning algorithm attains sublinear regret under certain delay conditions, and the two-point risk-averse learning algorithm achieves sublinear regret with minimal restrictions on the delay. We provide numerical experiments on a dynamic pricing problem to demonstrate the performance of the proposed algorithms.

Mechanistic interpretability aims to reverse engineer the computation performed by a neural network in terms of its internal components. Although there is a growing body of research on mechanistic interpretation of neural networks, the notion of a mechanistic interpretation itself is often ad-hoc. Inspired by the notion of abstract interpretation from the program analysis literature that aims to develop approximate semantics for programs, we give a set of axioms that formally characterize a mechanistic interpretation as a description that approximately captures the semantics of the neural network under analysis in a compositional manner. We use these axioms to guide the mechanistic interpretability analysis of a Transformer-based model trained to solve the well-known 2-SAT problem. We are able to reverse engineer the algorithm learned by the model -- the model first parses the input formulas and then evaluates their satisfiability via enumeration of different possible valuations of the Boolean input variables. We also present evidence to support that the mechanistic interpretation of the analyzed model indeed satisfies the stated axioms.

Learning generalizable visual dynamic representation across different embodied environments is crucial for real-world robotic manipulation. As the scale and diversity of robot demonstration data are limited, recent works have turned to large-scale pre-training using human data. However, the morphological differences between humans and robots introduce a significant human-robot domain discrepancy, challenging the generalization of these human-data pre-trained models to downstream manipulation tasks. To address this, we propose a novel adaptation paradigm that utilizes readily available paired human-robot video data to bridge the discrepancy. Following this paradigm, our method exploits a human-robot contrastive alignment loss to align the semantics of human and robot videos, adapting pre-trained models to the robotic domain in a parameter-efficient manner. The experiments demonstrate significant improvements on 25 tasks across three different benchmarks, where the single-task, language-conditioned multi-task settings are covered, and two different pre-trained models are evaluated. On the large RLBench benchmark, our adaptation method achieves an average improvement of $8.9\%$ in success rate over the pre-trained R3M model across multiple tasks. We will release the code and models upon acceptance.