Reproduce “Reasoning Patterns for Efficient AI Coding: An Empirical Study” 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
Zhe Yin, Yuling Shi, Xiaodong Gu, Beijun Shen
The authors analyze coding-model reasoning traces by labeling actions and mining recurring sequences across four tasks, then compare patterns in traces judged coherent or incoherent. Verification-centered sequences are more common in coherent traces, while design and knowledge-retrieval loops are more common in incoherent ones. On GLM-5.1 code-execution reasoning, pattern-aware stopping cuts API-reported tokens 68.53%, while pass@1 falls 1.21 percentage points.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.