Reproduce “AgentLeak: Cloning Stronger LLM Agent Capabilities onto Weaker Agents Beyond Skill Stealing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaoting Lyu, Yuhong Wu, Yufei Han, Shichang Liu, Liang Zhang, Bin Wang, Bin Wang, Xiaobo Ma and 1 more
AgentLeak reconstructs attacker-side skills by comparing a stronger agent’s successful execution with a weaker agent’s failure and turning diagnosed procedural deficits into instructions. In a 20-scenario SkillsBench evaluation, it reached 72.80% macro-averaged IPR, versus 30.64% for raw skills. Its 87.12% CRR measures the recovered share of the victim’s performance gap, not parity; results apply to the tested setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.