Đang tải bảng tin…
Towards Data Science0 Hot0 bình luận10 phút đọc
The LLM Judge That Kept Agreeing With Itself
A production incident involving an LLM-as-judge system that approved a SQL query with a silently dropped filter clause reveals structural bias problems in using LLMs to evaluate other LLMs' outputs. The judge and generator agent shared the same underlying model, causing self-preference bias where the judge favored outputs written in a familiar style regardless of correctness. Other systematic biases identified include verbosity bias (favoring longer, more detailed outputs) and position bias (order of comparison affecting verdicts). Fixes included routing judgment to a model from a different family than the generator, rewriting rubrics to explicitly penalize unnecessary verbosity, and continuously calibrating judge decisions against human review on held-out samples, with categories showing weak judge-human agreement routed to humans by default.
Nguồn: https://towardsdatascience.com/the-llm-judge-that-kept-agreeing-with-itself. 8 Sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.
