Reproduce “Learning When and How to Intervene: A Hindsight-Distilled Sentinel for Coding 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
Jiangrui Zhao, Chenglong Li, Meng Zhang, Xiaoting Du
HiSentinel places a small model before a coding agent’s proposed action, choosing ALLOW, REDIRECT, or HARD-PAUSE. A future-aware teacher uses recorded outcomes during training to supervise a causal student that acts without future observations. With Qwen3-Coder, the 1.7B system raises SWE-bench Verified Mini resolution from 30% to 44%, but labels reflect one realized continuation, not a counterfactual comparison.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.