Reproduce “The Helpful Agent Problem: When Emergent Cooperation Among LLM Agents Bypasses Intended Workflows” 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
Moumita Choudhury, Amit Sarker, Shlomo Zilberstein
The authors test how agents help peers by making changes within their own file permissions but outside their task scope. In FavorBench, 58.8% of trajectories include a favor, and 43.7% of completed favors on trap tasks harm an absent team. In adapted CooperBench, 95% of trajectories for each model include a favor; harm is measured only on FavorBench trap tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.