Judge as A Judge: Improving the Evaluation of Retrieval-Augmented Generation through the Judge-Consistency of Large Language Models
Year: 2025
Type: conference-paper
Abstract: Retrieval-Augmented Generation (RAG) has proven its effectiveness in alleviating hallucinations for Large Language Models (LLMs).However, existing automated evaluation metrics cannot fairly evaluate the outputs generated by RAG models during training and evaluation.LLM-based judgment models provide the potential to produce high-quality judgments, but they are highly sensitive to evaluation prompts, leading to inconsistencies when judging the output of RAG models.This paper introduces the Judge-Consistency (ConsJudge) method, which aims to enhance LLMs to generate more accurate evaluations for RAG models.Specifically, ConsJudge prompts LLMs to generate different judgments based on various combinations of judgment dimensions, utilizes the judge-consistency to evaluate these judgments, and selects the chosen and rejected judgments for DPO training.Our experiments show that ConsJudge can effectively provide more accurate judgments for optimizing RAG models across various RAG models and datasets.Further analysis reveals that judgments generated by ConsJudge have a high agreement with the superior LLM.All codes are available at https://github.com/OpenBMB/ConsJudge.Case 1: What are the virulence factors of anthrax?Ground Truth: Bacillus anthracis.Choice A: The virulence factors of anthrax are a group of proteins produced by the Bacillus anthracis bacterium that contribute to its ability to cause disease in humans and animals.Choice B: anthracis.Choice C: lethal factor, edema factor and antiphagocytic factor.Choice D: lethal factor,antiphagocytic factor and other factors.Raw Metric (ROUGE-L): Choice B(0.667) > Choice A(0.143) > Choice C(0) = Choice D(0) Zero-Shot: ...choice C and D are more complete as they specifically mention the names of the virulence factor... Best answer :C.SFT:... from the perspective of Coherence... Best answer: C ... .ConsJudge: from the perspective of semantic consistency, completeness, and coherence... Best answer: A ... .Case 2: How much is 55% of 40 is greater than 4/5 of 25? Options: (A) 29.(B) 776.(C) 2. (D) 12. (E) 99.Ground Truth: (C) Choice A: According to the information of question, 55% of 40 is 22, which is 2 greater than 4/5 of 25, which equals 20.Choice B: (C), 55% of 40 is greater than 4/5 of 25 by 1. Choice C: (C).55% of 44 is smaller than 4/5 of 25 (which is 20) by 2. Choice D: 55% of 40 is smaller than 4/5 of 25.Raw Metric (Accuracy): Choice A(0), Choice B(1), Choice C(1), Choice D(0) Zero-Shot: ...there is no hallucination in any of the choices.... Choice C is the best answer ... .SFT: ... from the perspective of Completeness ...
Authors: Shuliang Liu, Xinze Li, Zhenghao Liu, Yukun Yan, Cheng Yang, Zheni Zeng, Zhiyuan Liu, Maosong Sun, Ge Yu
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Awards: G2602313477, G3078953789, G4466500611
OpenAlex ID: W4412888671
