Reproduce “LEANGUARD: Lean-Verified Guardrails for Tool-Using Agents” 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
Sreehari Sankar, Debargha Ganguly, Vikash Singh, Mohsen Hariri, Weicong Chen, Yanyan Zhang, Shouren Wang, Wang Yang and 12 more
LeanGuard checks proposed tool calls against a reviewed policy and trusted backend facts before dispatch, using a verified evaluator. Across five Qwen models on τ²-bench, at least 33 unguarded policy-violating calls executed and changed state; LeanGuard dispatched no denied calls. In DTap’s read-only deployment, direct attack success fell from 57.9% to 5.9%, while benign task success fell from 53.1% to 12.5%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.