Reproduce “EPISTEME: Trustworthy Coding Agents Through Tool Call Intent” 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
Sofia Bobadilla, Carmine Cesarano, Raphina Liu, Javier Ron, Martin Monperrus
EPISTEME is a coding-agent harness that requires shell calls to declare an act type and reason before execution, then records workspace effects and checks them against the claim. In monitor mode, it exposed 44 of 49 misbehaving completed runs (90%, 95% CI 78–96%) on an impossible coding benchmark; all were disclosed in agent-written reasons, while effects contradicted only three.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.