Single-shot retrieval for multi-step search sessions

Develop single-shot text-based retrieval methods that can effectively handle multi-step user search sessions where the initial query does not immediately lead to engagement, thereby eliminating the need for user reformulation.

Background

The paper analyzes queries that required user reformulation (multi-step sessions) and notes that these are hard to construct because the initial query did not directly lead to engagement. It states that addressing such cases remains an open challenge for single-shot retrieval methods.

This highlights a gap between single-shot retrieval capabilities and real-world multi-step search behaviors, motivating future research on methods that can resolve complex sessions in a single generation.

References

\Cref{tab:search} further analyzes a subset of challenging queries that required user reformulation ("Multi-steps"); these sessions are difficult to construct because the initial query did not immediately lead to engagement and remain an open challenge for single-shot retrieval.

A Unified Language Model for Large Scale Search, Recommendation, and Reasoning  (2603.17533 - Nadai et al., 18 Mar 2026) in Appendix, Search by item type and query complexity