LLM Judges Unreliable? New Framework Boosts Human Correlation by 30%
dair_ai · x · 2026-08-30
Building reliable LLM-as-a-Judge systems is challenging. A new study introduces a reference-full benchmark based on realistic human-to-human dialogues, featuring over 30,000 expert-generated turns and 36,000 human annotations. Results show that classical automatic metrics and reference-free LLM judges are unreliable against expert judgment. The proposed 'Mixture-of-Judges' framework combines multiple evaluative signals, recovering roughly 30% better correlation with human assessment.
More from Research
- A beginner-friendly introduction to causal inference kicks off, Part 1 — Nasereliver · 2026-10-16
- FBI method trains for Pass@K with one sample per update at 12.3% of the cost — EmtiyazKhan · 2026-10-16
- Berkeley researchers sample humanoid robot policies directly from weight space with diffusion — berkeley_ai · 2026-10-16
- OneWM: unified causal training shows world and action modeling improve each other — IAAI233 · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- 5.8M-param Speech Event Detector adds laughs, breaths and背景 sounds to Whisper transcripts — Ayush Kumar Bar · 2026-10-16