Why Cohen’s kappa can overstate alignment between LLM judges and humans
IanArawjo · x · 2026-07-21
A discussion of why Cohen’s kappa can be a misleading proxy for whether an LLM judge is “aligned with humans.” The attached chart shows that even when weighted kappa looks high, the uncorrected false-positive rate can still vary a lot depending on whether the disagreement is random noise or a systematic bias.
Key point
- These agreement metrics are much more sensitive to random disagreement than to consistent bias.
- An LLM judge that is consistently 1 point above or below human raters can look “aligned” under inter-rater metrics while still being materially miscalibrated.
- The plot contrasts uncorrected vs PP-corrected rates across several kappa bands to illustrate the issue.
Related event: Study Warns: Unchecked LLM Judges Yield High False Positives(7 posts)→
More from Research
- Datology webinar: Curation Studio matched top data baselines with under 20% compute — josh_wills · 2026-10-27
- KANLab: an interactive browser playground for Kolmogorov-Arnold Networks — Lukas Niehaus · 2026-10-27
- CMU's LLM2Cert synthesizes stability certificates verified via SMT and Lean — CarnegieMellonU · 2026-10-27
- NVIDIA open-sources FLORA, a robotics data engine ~125x cheaper than GPT-6 Astra — nvidia · 2026-10-27
- RSNA 2026 Deep Learning Labs return: eight hands-on AI classes over four days in Chicago — PRouzrokh · 2026-10-27
- Steering two orthogonal concept neurons makes a model compose "gnome" — Sauers_ · 2026-10-27