Reproduce “Self-Reflective Generation at Test Time” 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
Jian Mu, Qixin Zhang, Zhiyong Wang, Menglin Yang, Shuang Qiu, Chengwei Qin, Zhongxiang Dai, Yao Shu
SRGen monitors next-token entropy and pauses when it exceeds a rolling local baseline. It optimizes a temporary hidden-representation adjustment to make the next-token distribution more decisive while preserving compatibility with the generated prefix, then discards it. This requires no retraining but uses inference-time gradients. On AIME2024, it scores 61.3% versus Self-Refine’s 52.0% for DeepSeek-R1-Distill-Qwen-7B, not establishing universal superiority.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.