A new research paper highlights a significant issue with using Large Language Model (LLM) judges for evaluating open-ended responses, particularly in scenarios lacking a ground truth answer. The study found that these LLM judges tend to be overly lenient and can incorrectly credit wrong answers when no reference is provided. The research also demonstrated that providing reference answer information in the prompt can drastically alter the LLM judge's decisions, sometimes flipping them by as much as 85%, and that these changes generally align with human judgments. The paper emphasizes the necessity of calibrating LLM judges with reference-aware evaluations before deploying them in reference-free settings to ensure reliable assessment. AI
IMPACT Highlights potential unreliability in LLM-based evaluation, necessitating careful calibration for accurate assessment.
RANK_REASON Research paper published on arXiv detailing a flaw in LLM evaluation methods.
- alphaXiv
- arXiv
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- Kranti Chalamalasetti
- LLM judges
- ScienceCast
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
