Relevance to Utility: Process-Supervised Rewrite for RAG

2025-09-19

AI Summary

In retrieval-augmented generation (RAG) systems, a mismatch exists between retrieval relevance and generative utility: retrieved documents are often topically relevant yet lack critical information required for reasoning. Method: We propose a process-supervised rewriting optimization framework that—uniquely—leverages intermediate reasoning signals generated during the LLM’s answer production as explicit supervision to model document utility more accurately. We further design an LLM distillation pipeline to transfer large-model judgment capabilities on rewriting quality to a lightweight rewriting model. Contribution/Results: Our method jointly optimizes retrieval, rewriting, and answer generation in an end-to-end manner. It achieves significant improvements over strong baselines across multiple open-domain question answering benchmarks, demonstrating both the effectiveness and generalizability of process supervision in enhancing generative utility.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Retrieval-Augmented Generation systems often suffer from a gap between optimizing retrieval relevance and generative utility: retrieved documents may be topically relevant but still lack the content needed for effective reasoning during generation. While existing "bridge" modules attempt to rewrite the retrieved text for better generation, we show how they fail to capture true document utility. In this work, we propose R2U, with a key distinction of directly optimizing to maximize the probability of generating a correct answer through process supervision. As such direct observation is expensive, we also propose approximating an efficient distillation pipeline by scaling the supervision from LLMs, which helps the smaller rewriter model generalize better. We evaluate our method across multiple open-domain question-answering benchmarks. The empirical results demonstrate consistent improvements over strong bridging baselines.

Problem

• Bridging relevance-utility gap in RAG systems
• Optimizing retrieval for generative answer correctness
• Improving document utility through process supervision

Innovation

• Process-supervised rewrite for RAG
• Optimizing generation correctness probability directly
• Efficient distillation pipeline with LLM supervision