CTR-Guided Generative Query Suggestion in Conversational Search (EMNLP 2025 Industry Track)

Forum topic · 小凯 · 2026-07-05

Summary

This entry indexes an industry-track paper titled "CTR-Guided Generative Query Suggestion in Conversational Search", published in the EMNLP 2025 proceedings by the ACL (Association for Computational Linguistics), available in the ACL Anthology at https://aclanthology.org/2025.emnlp-industry.178/. The paper addresses conversational search systems, focusing on generative query suggestion guided by click-through rate (CTR) signals. The approach sits at the intersection of information retrieval, recommendation, and large language model (LLM) generation, where suggestion of follow-up queries in multi-turn search dialogs must balance semantic relevance with commercial engagement metrics such as CTR. The surrounding forum entry places the work within a broader taxonomy of conversational search research, alongside related efforts on generative query reformulation (e.g., ConvGQR), collaborative search agents, and retrieval-augmented conversational systems. The forum notes that exact quantitative results should be verified against the original PDF, and frames the paper within engineering concerns common to industrial search stacks: latency budgets, online A/B validation of offline gains, hallucination control, and deployment cost. This page serves as a citation-friendly, English-language summary and cross-reference hub for readers tracking the conversational search and generative information retrieval literature.

Overview

CTR-Guided Generative Query Suggestion in Conversational Search is an industry-track paper published at EMNLP 2025 (Empirical Methods in Natural Language Processing), part of the ACL (Association for Computational Linguistics) anthology.

Publication source: https://aclanthology.org/2025.emnlp-industry.178/
Track: EMNLP 2025 Industry Track
Topic area: Conversational Search / Generative Information Retrieval

What the Paper Is About

Based on the title and venue, the work targets generative query suggestion in conversational search, guided by CTR (click-through rate) signals. In multi-turn conversational search experiences, the system proactively suggests follow-up queries to users. Using generative models (typically LLM-based) for this task must reconcile two objectives:

1. Semantic/intent relevance — suggestions should align with the user's conversational information need; 2. Engagement signals — CTR feedback from logs reflects which suggestions users actually click, providing a real-world supervision or reward signal.

This places the paper in the emerging area where conversational query reformulation, generative retrieval, and industrial recommendation/CTR modeling intersect.

Context in the Literature

The forum entry positions this work among related conversational search research, including:

*ConvGQR: Generative Query Reformulation for Conversational Search* (arXiv:2305.15645)
*A Survey of Conversational Search* (ACM, Sep 2025)
*ChatRetriever: Adapting Large Language Models for Generalized and Robust Retrieval* (arXiv:2404.13556)
*CoSearchAgent: A Lightweight Collaborative Search Agent* (arXiv:2402.06360)

Broader themes discussed in the forum entry's analysis of this research area:

The evolution from BM25 and dense retrieval toward generative retrieval and LLM-based agents, where the number and strategy of retrieval steps themselves become learnable;
The gap between offline metrics (nDCG, MRR) and online satisfaction/CTR, requiring interleaving experiments and human evaluation;
Industrial deployment constraints: latency (p99 budgets), token/GPU cost per query, safety of open retrieval, and hallucination mitigation.

Practical Notes for Practitioners

When evaluating papers of this type for production use, the forum recommends checking:

Whether reported offline gains translate to online CTR/satisfaction in interleaved A/B tests;
Index update frequency and embedding version compatibility;
Failure modes such as empty retrieval, wrong tool calls, or over-generation;
Cost per query (token usage, GPU occupancy) and routing to smaller distilled models.

Caveats

The original forum post summarizes the paper primarily from its title, venue, and public metadata; specific experimental numbers, author affiliations, and methodological details should be confirmed by consulting the official PDF.