AIBRIEF
A paper accepted at SIGIR 2026, the leading information retrieval research conference, posted to arXiv (2607.18626), adapts pseudo relevance feedback — a technique that reuses top results to expand the query — to multi vector AI search.
A SIGIR 2026 paper, PLAID-PRF, posts up to 4.3% nDCG@10 over PLAID by reusing a retriever's own centroid vectors as query-expansion tokens instead of re-clustering documents at query time.
PLAID is a "late-interaction" retriever: it represents both the query and each document as many small vectors rather than one big one. That is more accurate than single-vector search, but expensive to refine. Earlier pseudo-relevance feedback (PRF) methods paid that cost by re-clustering documents for every query. PLAID-PRF skips the step. It treats PLAID's internal centroid vectors as centroid-like expansion tokens, picks a small, diverse set, appends them to the original query, and reruns PLAID.
On the in-domain MSMARCO benchmark and four out-of-domain BEIR datasets, the authors report up to 4.3% nDCG@10 and 7.3% MRR@10 over PLAID alone, with substantially less computation than prior PRF methods. A bar chart in the arXiv PDF places PLAID-PRF at the top of the average nDCG@10 band (~0.76), above LLM-VPRF and the ColBERTv2/CWPRF/RepLLaMA cluster.
The evaluation is on standard academic benchmarks, not user tasks. No vendor ships the method, no user study is included, and the cost-saving claim is relative to prior PRF, not to PLAID's baseline. PRF also assumes the top results are good enough to expand from, which can fail on adversarial or tail queries.