Importance aware weighting improves rubric based ai evaluation scores
Not All Criteria Matter: Importance-Aware Criteria Weighting for Rubric-Based LLM Evaluation
Computation and Language
Summary
Evaluating large AI language models often involves checking several criteria and giving each equal importance, but not all criteria matter the same. The authors propose a new method that asks the AI to compare criteria two at a time to figure out which are more important, making the overall evaluation fairer. This method fixes inconsistencies and uses math to combine the comparisons into weights for each criterion. Their tests show this approach works better than treating all criteria equally or other weighting methods. It also helps improve AI models that learn from these evaluations.
What this means in practice
- •For machine learning engineers: Create more accurate evaluation metrics for AI language models by weighting evaluation criteria based on pairwise importance judgments.
- •For product teams building ai assistants: Improve downstream AI assistance by using better weighted evaluation scores to train stronger policy models via reinforcement learning.
Authors
Haocheng Yang, Licheng Pan, Yiming Dong, Xiaoxi Li, Yuan Lu, Zhenhao Chen, Jiajun Fan, Shuai Shao, Chunyuan Zheng, Hao Wang
Abstract
Rubric-based evaluation assesses the responses of large language models (LLMs) by aggregating the verification results of explicit criteria, yet not all criteria matter equally. However, existing methods commonly aggregate the verification results with uniform weighting, which assigns minor and overlapping criteria the same weight as essential ones and thus produces misleading evaluation scores. Although some methods prompt an LLM to directly assign weights, direct weighting requires a holistic judgment over all criteria that LLMs struggle to make reliably. To address these limitations, we propose Importance-Aware Criteria Weighting (ICW), which derives the weights of criteria from pairwise importance comparisons made by an LLM. Specifically, ICW transforms the holistic judgment over all criteria into a series of pairwise importance comparisons, checks and revises inconsistent comparisons, and derives the weights from the principal eigenvector of the comparison matrix. Extensive experiments demonstrate that ICW consistently outperforms uniform weighting and competitive weighting methods, and translates into better policy models in downstream reinforcement learning.