Abstract Pseudo-relevance feedback (PRF) has been shown to improve the effectiveness of bag-of-words retrieval models. However, the incorporation of PRF into dense neural retrieval remains challenging. Direct updates to query embeddings can induce semantic drift, particularly when pseudo-relevance signals are noisy or when the initial ranking is already strong. In addition, many neural PRF approaches model the top-ranked documents as an unordered set. As a result, rank-order information is discarded, despite serving as an imperfect but potentially informative proxy for evidence reliability. In this article, we propose Dense-Chain Pseudo-Relevance Feedback (DC-PRF), which formulates dense PRF as sequential state refinement over an ordered curriculum of top-ranked documents. The proposed framework incrementally refines the query representation in embedding space. Each refinement step is explicitly conditioned on the evolving query state. To constrain query drift, we introduce a Safety Anchor that regularizes the refinement trajectory toward the original query. This is achieved through an anchor-based training objective and residual interpolation at inference time. We evaluate DC-PRF on five BEIR benchmarks (FiQA, NFCorpus, SciFact, Quora, and TREC-COVID). Experimental results indicate that DC-PRF tends to improve retrieval effectiveness on the evaluated datasets when early-ranked documents provide informative evidence. At the same time, the method exhibits largely baseline-preserving behavior in high-precision, saturated regimes (e. g. , Quora), reflecting a do-no-harm tendency under limited headroom. Under our latency evaluation protocol, DC-PRF yields a 12 × reduction in reranking latency relative to cross-encoders.
Khaled Albishre (Fri,) studied this question.