SIRA: Superintelligent Retrieval Agent Compresses Multi-Turn Search into a Single Discriminative BM25 Query

Forum topic · 小凯 · 2026-05-10

Summary

SIRA (SuperIntelligent Retrieval Agent), a paper by Zeyu Yang, Qi Ma, Jason Chen, and Anshumali Shrivastava (arXiv:2605.06647, May 2026), rethinks retrieval-augmented agents by compressing multi-round exploratory search into a single, corpus-discriminative retrieval action. Instead of asking which terms merely match a query, SIRA asks which terms are likely to separate desired evidence from corpus-level distractors. On the corpus side, an LLM offline enriches each document with missing search vocabulary; on the query side, it predicts evidence vocabulary the query omits; document frequency statistics are used as tool calls to filter missing, overly common, or unlikely terms. The final retrieval step is a single weighted BM25 call combining the original query with verified expansions. Across ten BEIR benchmarks and downstream QA tasks, SIRA significantly outperforms dense retrievers and state-of-the-art multi-turn agent baselines, showing that a well-formed lexical query guided by LLM cognition and lightweight corpus statistics can substantially outperform costlier multi-turn search while remaining interpretable, training-free, and efficient.

Paper Overview

Research area: Information Retrieval (IR) Authors: Zeyu Yang, Qi Ma, Jason Chen, Anshumali Shrivastava Published: 2026-05-07 arXiv: 2605.06647

Abstract

Retrieval-augmented agents are increasingly becoming the interface to large organizations' knowledge bases, but most still treat retrieval as a black box: they issue exploratory queries, inspect returned snippets, and iteratively reformulate until useful evidence emerges. This resembles a novice searching an unfamiliar database, rather than an expert navigating with strong priors about terminology and likely evidence — leading to unnecessary retrieval rounds, increased latency, and poor recall.

The authors introduce the Superintelligent Retrieval Agent (SIRA), which defines superintelligence in retrieval as the ability to compress multi-turn exploratory search into a single corpus-discriminative retrieval action. Rather than asking which terms are relevant to the query, SIRA asks which terms are likely to separate the desired evidence from corpus-level distractors.

Corpus side: an LLM offline enriches each document with missing search vocabulary.
Query side: the LLM predicts evidence vocabulary the query omits.
Filtering: document frequency statistics are invoked as tool calls to filter out missing, overly common, or unlikely-to-discriminate candidate terms.
Final retrieval: a single weighted BM25 call combining the original query with the verified expansions.

Results

On ten BEIR benchmarks and downstream question-answering tasks, SIRA significantly outperforms dense retrievers and state-of-the-art multi-turn agent baselines. This shows that a well-formed lexical query — guided by LLM cognition and lightweight corpus statistics — can substantially outperform far more expensive multi-turn search, while remaining interpretable, training-free, and efficient.

*Automatically collected on 2026-05-10.*