Are LLMs Reliable Judges? 8 Months of Lessons Learned
larabyeol · reddit · 2026-07-07
A team processing roughly 50,000 LLM outputs weekly shared 8 months of practical experience using GPT-4o for automated evaluation. It proved fast and effective at catching obvious regressions, formatting errors, and hallucinated entities. However, it suffered from poor cross-run consistency (scoring divergence exceeded 30% within two weeks without locked prompts and model versions), positional bias (preferring the first option in A/B tests), and leniency toward its own outputs. They recommend locking evaluation prompts and model versions, as well as swapping positions across two runs to mitigate these issues.
More from Research
- HarnessMM cuts agent tokens per task by up to 88% while raising success rates — omarsar0 · 2026-10-26
- Ben Recht: The world needs a truly open language model, and it's easier to build than you think — beenwrekt · 2026-10-26
- New Blog Breaks Black-Box Distillation Into Five Practical Steps — chrisrohlf · 2026-10-26
- DetectorScope turns black-box AI text detectors into auditable ones with sparse autoencoders — _akhaliq · 2026-10-26
- OPST paper shows LLMs self-improve with zero human data, guided only by topic descriptions — _akhaliq · 2026-10-26
- Tencent's HiSTEVE 1.0: hierarchical VLA agent plays Minecraft with native keyboard and mouse, hits 85% SOTA — _akhaliq · 2026-10-26