Institution profile

Beihang University

Academic institutionasia · cn
Official website
Research library2,419linked papers
Opportunities0open roles
Selected work

Representative Papers

GroupDebate: Enhancing the Efficiency of Multi-Agent Debate Using Group Discussion

Sep 21, 2024arXiv.org

To address the scalability bottleneck in multi-agent debate—specifically, the exponential growth in token consumption with increasing agent count and debate rounds—this paper proposes a *grouped multi-agent debate* architecture. Agents are partitioned into disjoint subgroups that conduct parallel internal debates; inter-group information exchange and a dynamic consensus mechanism then aggregate intermediate results efficiently. This approach breaks the traditional linear scaling constraint and represents the first systematic integration of grouping principles into multi-agent debate frameworks. Extensive experiments across multiple logical reasoning benchmarks demonstrate that our method reduces token consumption by up to 51.7% relative to baseline methods, while simultaneously improving accuracy by up to 25%. The architecture thus achieves a significant trade-off improvement between computational efficiency and reasoning performance.

37 citations2 influentialRead paper

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

An Empirical Study of Automated Vulnerability Localization with Large Language Models

Mar 30, 2024arXiv.org

This work systematically evaluates the effectiveness of large language models (LLMs) for line-level vulnerability localization (AVL)—a task lacking comprehensive empirical investigation. Experiments are conducted on BigVul (C/C++) and smart contract vulnerability datasets, covering over ten code-understanding LLMs (60M–16B parameters) spanning encoder-only, encoder-decoder, and decoder-only architectures, under zero-shot, one-shot, discriminative fine-tuning, and generative fine-tuning paradigms. Key contributions include: (1) the first empirical demonstration that discriminative fine-tuning substantially outperforms existing approaches; (2) the proposal of sliding-window context partitioning and right-forward embedding to mitigate context-length limitations; and (3) strong cross-CWE and cross-project generalization, yielding significant improvements in localization accuracy and surpassing state-of-the-art methods.

25 citations2 influentialRead paper

Your Group-Relative Advantage Is Biased

Jan 13, 2026

This work addresses a systematic bias in advantage estimation within population-based reinforcement learning, where difficult prompts are consistently underestimated while easy ones are overestimated, thereby disrupting the balance between exploration and exploitation. The study is the first to uncover the underlying mechanism of this bias and proposes a novel method—History-Aware Adaptive Difficulty Weighting (HA-DW)—which dynamically corrects advantage estimates by leveraging training dynamics and difficulty anchors. Through theoretical analysis grounded in GRPO and its variants, the effectiveness of HA-DW is empirically validated across five mathematical reasoning benchmarks, demonstrating significant performance improvements. These results underscore that correcting advantage bias is crucial for effective reinforcement learning from verifiable feedback (RLVR).

6 citationsRead paper

Toward Ultra-Long-Horizon Agentic Science: Cognitive Accumulation for Machine Learning Engineering

Jan 15, 2026

This work addresses the challenge of maintaining strategic coherence and iterative refinement in artificial intelligence systems over ultra-long scientific research cycles. To this end, it introduces the ML-Master 2.0 agent, which reconfigures context management as a cognitive accumulation process through a Hierarchical Cognitive Cache (HCC) architecture. Inspired by multi-level memory systems, HCC dynamically distills execution trajectories into stable knowledge, decoupling immediate actions from long-term strategy and thereby transcending the limitations of static context windows. Integrated with dynamic knowledge distillation, cross-task experience consolidation, and large language model–driven autonomous experiment planning, the proposed approach achieves a state-of-the-art medal rate of 56.44% on MLE-Bench under a 24-hour budget, demonstrating for the first time the feasibility of fully autonomous, ultra-long-horizon scientific discovery.

4 citationsRead paper
Recent publications

Latest Papers