Reproduce “Risk-Triggered Search: Behavior-Preserving Safe Control” 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

Xinyu Deng, Bohao Lan, Kunpeng Li, Mingkai Yuan, Weihang Pan, Binbin Lin

RTS adds risk-aware search to a frozen visual controller, predicting process-violation risk along candidate action sequences and searching only when a nominal plan raises an alarm. A risk gate filters candidates, while amplitude-preserving selection discourages stopping. On Reacher, RTS-AP reaches safely in 50.17% of episodes versus 8.50% for Direct, despite similar goal success; results are task-specific, not a safety guarantee.

Summary by alphaXiv.

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