Study Warns Top LLM Leaderboard Scores May Be Inflated by Winner's Curse
A new arXiv paper argues that reported performance of top-ranked models on public LLM leaderboards can be systematically overstated due to statistical selection effects known as the winner's curse.
Read the full story at arXiv stat.ML ↗
Timeline · 1 report
- 2026-11-18 05:00 · arXiv stat.ML
SOTA or Luck? The Winner's Curse in LLM Leaderboards