Generalized Pseudo-Relevance Feedback

TL;DR

Generalized Pseudo-Relevance Feedback addresses two core limitations of prior query rewriting approaches—the relevance assumption and the model assumption—by grounding natural-language rewrites in retrieved evidence and optimizing them through a utility-driven training pipeline. The framework comprises retrieval-augmented rejection sampling, cold-start supervised fine-tuning, and reinforcement learning with a multi-view reward to align rewrites with downstream retrieval performance. Empirical results on MS-MARCO and BEIR demonstrate that GPRF consistently surpasses PRF, VPRF, and zero-shot GRF baselines across sparse and dense retrievers, including cross-model settings, with substantial improvements in $NDCG@10$ and Recall. These findings suggest that GPRF provides a robust, transferable solution for query rewriting, enabling more reliable and domain-general IR systems without heavy reliance on task-specific prompts or model internals.

Abstract

Query rewriting is a fundamental technique in information retrieval (IR). It typically employs the retrieval result as relevance feedback to refine the query and thereby addresses the vocabulary mismatch between user queries and relevant documents. Traditional pseudo-relevance feedback (PRF) and its vector-based extension (VPRF) improve retrieval performance by leveraging top-retrieved documents as relevance feedback. However, they are constructed based on two major hypotheses: the relevance assumption (top documents are relevant) and the model assumption (rewriting methods need to be designed specifically for particular model architectures). While recent large language models (LLMs)-based generative relevance feedback (GRF) enables model-free query reformulation, it either suffers from severe LLM hallucination or, again, relies on the relevance assumption to guarantee the effectiveness of rewriting quality. To overcome these limitations, we introduce an assumption-relaxed framework: \textit{Generalized Pseudo Relevance Feedback} (GPRF), which performs model-free, natural language rewriting based on retrieved documents, not only eliminating the model assumption but also reducing dependence on the relevance assumption. Specifically, we design a utility-oriented training pipeline with reinforcement learning to ensure robustness against noisy feedback. Extensive experiments across multiple benchmarks and retrievers demonstrate that GPRF consistently outperforms strong baselines, establishing it as an effective and generalizable framework for query rewriting.

Figures (5)

• Figure 1: The comparison between PRF, VPRF, GRF, and our proposed GPRF. The pseudo-documents or rewritten queries produced by GRF and GPRF can be processed in various ways, such as directly concatenating them with the original query, integrating them into PRF or VPRF systems, or performing retrieval separately for each, then post-processing the results.
• Figure 2: Overview of the Utility-oriented Training Pipeline. The high-utility reformulations obtained via rejection sampling in Stage 1 are directly utilized as training labels for supervised fine-tuning (SFT) in Stage 2. In Stage 3, we directly use the performance on downstream retrieval tasks as the reward signal of reinforcement learning (RL).
• Figure 3: Bucket-based analysis on MS dev. Queries are grouped into buckets based on their baseline BM25 performance, and the NDCG@10 improvement of three feedback-based methods, RM3, Lamer, and GPRF, is evaluated within each group. From left to right, the relevance of top-retrieved feedback documents in each group increases.
• Figure 4: A case study shows that our training framework can effectively alleviate the noisy feedback problem.
• Figure 5: Cross-model experimental results on DL19 and DL20 with Llama. The results of providing different retrievers with various feedback are reported. It can be observed that using varying feedback consistently improves the performance of different retrievers.