LLM Judges Exhibit Score Range Bias
xennygrimmato_ · x · 2026-07-15
This ACL paper discusses a new bias in LLM-as-a-Judge: score range bias.
Core Findings
- When applying an equivalent shift to the scoring range of the judge model (e.g., changing 1-5 to 2-6), the model's correlation with human judgments systematically changes.
- This shift is independent of the evaluated content, indicating a previously under-documented evaluation distortion in judge models.
- The phenomenon is observable across different model families, including Llama-3 and Qwen2.5, and holds true across various parameter scales.
Paper's Approach
The authors propose using contrastive decoding to mitigate this bias.
This post relays the release of a PatronusAI paper at ACL 2026, pointing to reliability issues inherent in evaluation methodologies rather than a specific model's leaderboard score.
More from Research
- Researchers identify "Forking": train loss falls while validation loss turns up — ZimingLiu11 · 2026-10-26
- Google DeepMind pilots world's first double-blind evaluations for frontier models — scychan_brains · 2026-10-26
- Rewriting scaling law study: optimal synthetic data share rises from 10% to 32% with compute — heghbalz · 2026-10-26
- FLASH: GPU-native deformable sim learns cloth folding in minutes on one RTX 5090 — ChongZzZhang · 2026-10-26
- LEMON: zero-shot LLM misbehavior monitor flags hallucinations from attention maps — yoavartzi · 2026-10-26
- Behind the EMNLP keynote: LMs as simulated learners to generate child language acquisition hypotheses — najoungkim · 2026-10-26