Reproduce “Calibrating LLM-Based Evaluator” on OpenResearch

About this paper

Yuxuan Liu, Tianchi Yang, Shaohan Huang, Zihan Zhang, Haizhen Huang, Furu Wei, Weiwei Deng, Feng Sun and 1 more

A gradient-free, multi-stage approach called AUTOCALIBRATE automatically derives and refines natural language scoring criteria for large language models, significantly improving their alignment with human judgments as reference-free evaluators for natural language generation tasks. The method boosts correlation with human evaluations by up to 30% across summarization, data-to-text, and hallucination detection, surpassing uncalibrated LLMs and other neural evaluators.

Key insights

• Explicit, natural language scoring criteria are crucial for aligning LLM evaluations with human judgments, effectively resolving the issue of "insufficient prompting."
• LLMs possess the ability to "self-calibrate" by inferring and refining evaluation rules from specific instances where their internal scoring mechanisms conflict with human preferences.
• A gradient-free approach allows for the effective calibration of powerful, closed-source LLMs, making these models practical for a wider range of real-world evaluation tasks without requiring access to their internal parameters.

Summary by alphaXiv.