Reproduce “Keeping LLM Agents on Procedure: Executable Finite-State Control for Long-Horizon Instruction Fidelity” 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

Yutao Sun, Haokun Li, Mingshuai Chen, Guanyan Li, Zhongyi Wang, Shengchao Qin, Jianwei Yin

The authors test whether an external runtime can keep an agent on procedure during long, unattended account reviews. It stores workflow state outside the conversation and gates out-of-order actions, transitions, and stopping. In a pre-registered study, Claude Haiku 4.5 averaged 139.2 faithful-and-correct decisions out of 200, 90.9 more than Remind-loop; enforcement does not ensure correct answers, and account-binding loopholes remained.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.