LLM-as-a-judge systems default to high scores; checklist approach improves accuracy

A user discovered that their LLM-as-a-judge system, using Claude, was consistently assigning scores of 7 or 8 out of 10 to answers, with 68% of 1,140 responses falling into this range. This indicated the model was not accurately evaluating the answers but rather defaulting to a high score. Attempts to make the model "stricter" only shifted scores down slightly without improving ranking agreement. A significant improvement was achieved by using a 1-5 scale with detailed descriptions for each level, increasing agreement with human graders to 0.68, and further by employing a binary checklist summed in code, which reached an agreement of 0.79. AI

IMPACT LLM-as-a-judge systems require careful prompt engineering and scale definition to ensure accurate evaluations, with checklist formats showing promise.

RANK_REASON The item details research into the effectiveness and limitations of using LLMs as judges for evaluating responses, including experimental fixes. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• Claude
• LLM-as-a-Judge
• Preterview

• product
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →