The authors model each keep-if-better decision using correlated measurement errors, accounting for candidates that share both an incumbent and evaluation items.
Reused evaluation sets exaggerate gains in LLM self-improvement loops
Experiments with prompt-rewriting systems show that repeatedly selecting candidates on small evaluation sets creates large gaps between reported and held-out performance.
Big Tech
Litao Hu · Yutong Tang
Meta · Microsoft
Research Digest··3 min read
Hu and Tang study the winner’s curse in LLM self-improvement: selecting the best-looking rewrite from several candidates whose measured performance contains noise.
Why this paper
From Microsoft and Meta
In one line
Reusing small evaluation sets in LLM self-improvement loops inflates apparent gains, while held-out evaluation reveals smaller gains and often harmful accepted changes.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§