Reproduce “Agents' Last Doubt: Knowing When to Act and When to Step Back in 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

Qisai Liu, Qingyue Jiao, Bangwei Guo, Wenxi Dai, Changyi Yang, Jiahe Liu, Dimitris N. Metaxas, Yiyu Shi and 2 more

The authors introduce SURE-Act, which compares actions by cloning a decision-point state and running each option with the frozen backbone, then trains a controller to predict outcomes and rank choices. Ranking supervision adds 0.0309 utility over regression-only training. On 89 unseen families, the Qwen3-4B system’s 0.0575 Overall Score gain has a family-clustered interval including zero; repaired evaluations show more critical errors.

Summary by alphaXiv.

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