Causal explanation of ReAct-BM25’s performance gain

Determine whether the additional performance gain of ReAct-BM25 is caused by its feedback loop, in which retrieved articles guide subsequent query reformulation, rather than by other properties of the iterative retrieval procedure.

Background

ReAct-BM25 iteratively combines BM25 retrieval with LLM-based planning and observation: the observer retains useful retrieved articles, and the planner uses those articles to generate a new query in subsequent rounds. The system improves over one-shot reformulation, particularly on several ranking metrics. The paper does not establish why this additional gain occurs and offers the feedback-loop explanation only as a conjecture.

References

We conjecture that this additional gain comes from the feedback loop: unlike one-shot reformulation, ReAct-BM25 reviews the retrieved articles and subsequently uses them to guide the next query.

— GreekBarRetrieval: A Benchmark for Greek Statutory Retrieval  (2608.18752 - Beta et al., 19 Aug 2026) in Section 5.3, “ReAct-BM25 Results”