Multi-stage Large Language Model Pipelines Can Outperform GPT-4o in Relevance Assessment

TL;DR

The paper presents a modular two-stage relevance judgment pipeline that first performs binary relevance filtering and then a three-level relevance classification, using a mix of models and prompts to balance accuracy and cost. On the TREC-DL23 dataset, this approach achieves substantial α gains over GPT-4o-based baselines at low cost (around $0.20 per million tokens), and certain configurations even outperform the GPT-4o baseline while maintaining much lower expenses. The results demonstrate that multi-model, multi-stage designs can improve both the quality and scalability of relevance labeling, with meaningful implications for large-scale retrieval evaluation and dataset annotation. The work also discusses limitations (such as duplicate labels) and avenues for future improvement, including prompt optimization and broader dataset validation.

Abstract

The effectiveness of search systems is evaluated using relevance labels that indicate the usefulness of documents for specific queries and users. While obtaining these relevance labels from real users is ideal, scaling such data collection is challenging. Consequently, third-party annotators are employed, but their inconsistent accuracy demands costly auditing, training, and monitoring. We propose an LLM-based modular classification pipeline that divides the relevance assessment task into multiple stages, each utilising different prompts and models of varying sizes and capabilities. Applied to TREC Deep Learning (TREC-DL), one of our approaches showed an 18.4% Krippendorff's $α$ accuracy increase over OpenAI's GPT-4o mini while maintaining a cost of about 0.2 USD per million input tokens, offering a more efficient and scalable solution for relevance assessment. This approach beats the baseline performance of GPT-4o (5 USD). With a pipeline approach, even the accuracy of the GPT-4o flagship model, measured in $α$, could be improved by 9.7%.

Figures (8)

• Figure 1: Visual overview of the pipeline approach where different models can judge at different stages of the relevance judgement labelling.
• Figure 2: Reproduction of baseline (UMBRELA) with GPT-4o and GPT-4o mini.
• Figure 3: Cost vs Cohen's Kappa ($\kappa$) and Krippendorff's Alpha ($\alpha$).
• Figure 4: GPT-4o prediction vs GPT-4o mini predictions on TREC-DL 23.
• Figure 5: GPT-4o mini / GPT-4o mini Bin/R predictions on TREC-DL 23.