GroUSE: A Benchmark to Evaluate Evaluators in Grounded Question Answering

2024-09-10

arXiv.org

AI Summary

Existing RAG evaluation methods rely on LLM-as-a-Judge (e.g., GPT-4) but overlook its systematic deficiencies in calibration and discriminative power, leading to inadequate identification of generator failure modes. Method: We propose GroUSE—the first meta-evaluation benchmark for judge models—comprising 144 unit tests covering seven canonical RAG failure modes. We innovatively distill GPT-4’s reasoning traces to model judge behavior and fine-tune Llama-3 to enhance calibration. Contribution/Results: Experiments reveal poor generalization of mainstream open-source judge models. Fine-tuned Llama-3 achieves significantly improved agreement with GPT-4 (+28.6% Kendall τ) and superior calibration. GroUSE precisely identifies evaluation blind spots, offering an interpretable, reproducible, and principled evaluation paradigm for RAG judge models.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Metareasoning and Metaheuristics
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Retrieval-Augmented Generation (RAG) has emerged as a common paradigm to use Large Language Models (LLMs) alongside private and up-to-date knowledge bases. In this work, we address the challenges of using LLM-as-a-Judge when evaluating grounded answers generated by RAG systems. To assess the calibration and discrimination capabilities of judge models, we identify 7 generator failure modes and introduce GroUSE (Grounded QA Unitary Scoring of Evaluators), a meta-evaluation benchmark of 144 unit tests. This benchmark reveals that existing automated RAG evaluation frameworks often overlook important failure modes, even when using GPT-4 as a judge. To improve on the current design of automated RAG evaluation frameworks, we propose a novel pipeline and find that while closed models perform well on GroUSE, state-of-the-art open-source judges do not generalize to our proposed criteria, despite strong correlation with GPT-4's judgement. Our findings suggest that correlation with GPT-4 is an incomplete proxy for the practical performance of judge models and should be supplemented with evaluations on unit tests for precise failure mode detection. We further show that finetuning Llama-3 on GPT-4's reasoning traces significantly boosts its evaluation capabilities, improving upon both correlation with GPT-4's evaluations and calibration on reference situations.

Problem

• RAG Evaluation
• Large Language Models
• Accuracy and Coverage

Innovation

• GroUSE Framework
• Enhanced Automated Evaluation
• Llama-3 Fine-tuning