Institution profile

Shanghai AI Laboratory

Academic institutionasia · cn
Official website
Research library2,222linked papers
Opportunities0open roles
Selected work

Representative Papers

Derail Yourself: Multi-turn LLM Jailbreak Attack through Self-discovered Clues

Oct 14, 2024arXiv.org

This paper identifies a novel security vulnerability in large language models (LLMs): in multi-turn dialogues, LLMs exhibit heightened sensitivity to “natural distribution shifts”—semantically related yet superficially benign prompts—that evade existing alignment mechanisms. To exploit this, we propose ActorAttack, the first multi-turn jailbreaking framework grounded in actor-network theory; it leverages LLM-driven role-association modeling to automatically discover covert, diverse attack paths while preserving intent invisibility. Our contributions are threefold: (1) we introduce SafeMTData, the first open-source multi-turn adversarial dataset; (2) ActorAttack significantly outperforms state-of-the-art single- and multi-turn baselines across aligned models including GPT-4o, Claude, and Qwen; and (3) fine-tuning on SafeMTData substantially enhances model robustness against multi-turn attacks.

26 citations7 influentialRead paper

DAG-Plan: Generating Directed Acyclic Dependency Graphs for Dual-Arm Cooperative Planning

Jun 14, 2024arXiv.org

Long-horizon collaborative tasks for dual robotic arms face challenges including complex spatiotemporal dependencies among subtasks, difficulty in dynamic action allocation, and limited expressiveness of linear programming formulations. This paper proposes the first LLM-driven DAG-structured task decomposition framework, which automatically parses high-level instructions into directed acyclic graphs (DAGs) encoding dependency constraints, and integrates environment perception to enable real-time, dynamic action allocation and parallel adaptive execution across both arms. The method breaks away from predefined operational paradigms, supporting end-to-end, interpretable, and generalizable collaborative planning. Evaluated on the Dual-Arm Kitchen benchmark, it achieves a 52.8% efficiency gain over single-arm systems, improves success rate by 48% and reduces LLM query count by 84.1% compared to conventional dual-arm planners, significantly enhancing robustness and scalability in complex scenarios.

8 citationsRead paper

On Finding Small Hyper-Gradients in Bilevel Optimization: Hardness Results and Improved Analysis

Jan 02, 2023Annual Conference Computational Learning Theory

This work investigates the theoretical hardness and algorithmic efficiency of finding stationary points of the hyperobjective in nonconvex–convex and nonconvex–nonconvex bilevel optimization, under the Polyak–Łojasiewicz (PL) condition—rather than strong convexity—on the lower-level objective. We first establish an impossibility result: for zero-respecting algorithms, computing a hyperstationary point is fundamentally intractable in the nonconvex–convex setting. Under the PL condition, we break the reliance on strong convexity and derive tighter hypergradient convergence complexity bounds. We propose a novel analytical framework unifying implicit function differentiation, hypergradient estimation, and first-order optimization. This yields complexity guarantees of $ ilde{mathcal{O}}(varepsilon^{-2})$, $ ilde{mathcal{O}}(varepsilon^{-4})$, and $ ilde{mathcal{O}}(varepsilon^{-6})$ for deterministic, partially stochastic, and fully stochastic settings, respectively—substantially improving upon existing nonconvex bilevel optimization methods.

7 citations3 influentialRead paper

Data Scaling Laws in Imitation Learning for Robotic Manipulation

Oct 24, 2024arXiv.org

This study investigates data scaling laws in robotic imitation learning, addressing whether a single policy can achieve zero-shot generalization to unseen environments and objects of the same category using limited yet high-quality demonstration data. We propose a “diversity-first, demonstration-count-threshold-driven” data collection paradigm for efficient real-world robot data acquisition. We empirically uncover, for the first time, a power-law relationship between the number of training environments/objects and zero-shot performance. Evaluation is conducted under a rigorous, realistic cross-environment and cross-object benchmark using behavior cloning. With only half-a-day of real-robot demonstrations collected by four human operators—totaling ~4 hours—we achieve approximately 90% success rates on two distinct manipulation tasks across unseen environments and unseen objects. This demonstrates substantial improvements in data efficiency and generalization robustness, establishing scalable, practical foundations for deploying imitation learning in diverse real-world settings.

7 citationsRead paper
Recent publications

Latest Papers