Reproduce “Agent Behavior Maps: Locating Agents by What They Would Do, Without Running Them” 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
Shota Fujikawa, Takuya Tamura, Kosuke Akimoto, Daichi Haraguchi, Masafumi Oyamada
The authors introduce Agent Behavior Maps to compare LLM agents by scoring recorded actions, without collecting new rollouts. For each candidate, its harness reconstructs recorded histories and the model assigns likelihoods to next actions; centered scores form coordinates. Across eight agents, map distances track trajectory-KL ordering but overstate its magnitude, supporting nearby-agent identification rather than reliable distances for every pair.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.