Reproduce “The Winner's Curse in LLM Self-Improvement Loops: Selection Noise, Lock-in, and Acceptance Rules” 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
Litao Hu, Yutong Tang
The authors study self-improving language-model loops that propose instruction edits and keep candidates scoring better on reused evaluation items. They compare proxy accuracy on selection items with performance on held-out items. In Qwen2.5 runs, proxy scores exceeded held-out accuracy by 13–20 percentage points with 16 items and 1–5 with 256; alternative acceptance rules did not improve runs over greedy acceptance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.