Proactive Guidance of Multi-Turn Conversation in Industrial Search

Xiaoyu Li, Xiaoli Li, Li Gao, Yiding Liu, Xiaoyang Wang, Shuaiqiang Wang, Junfeng Wang, Dawei Yin2025 · DOI

Abstract

Xiaoyu Li, Xiao Li, Li Gao, Yiding Liu, Xiaoyang Wang, Shuaiqiang Wang, Junfeng Wang, Dawei Yin. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

2025.

In the evolving landscape of AI search assistants, providing accurate answers is only the first step. The greater challenge lies in anticipating user needs through proactive guidance. This paper presents a novel framework developed by Baidu researchers to steer multi-turn search conversations. By implementing a dual-phase architecture consisting of Goal-adaptive Supervised Fine-Tuning (G-SFT) and Click-oriented Reinforcement Learning (C-RL), the system effectively handles dynamic goal shifts while significantly reducing inference latency in industrial settings.

Introduction

Modern search assistants frequently struggle with "proactive guidance"—the ability to suggest relevant follow-up questions that help users refine their search goals. Two primary obstacles hinder current systems:

  1. Dynamic Goal Shifts: Users often pivot their intent mid-conversation. Standard models can become anchored to initial, now-irrelevant context.
  2. Latency-Quality Trade-offs: While Large Language Models (LLMs) offer high-quality reasoning, their inference speed often fails to meet the sub-second requirements of industrial search.

Methodology

The proposed architecture addresses these challenges through a two-phase approach designed for both accuracy and real-time performance.

Phase 1: Goal-adaptive Supervised Fine-Tuning (G-SFT)

To manage shifting user intents, the framework utilizes a Goal Adaptation Agent (GAA). Instead of processing raw, exhaustive chat histories, the GAA generates three focused outputs: an explicit goal analysis, a shift detection signal, and a goal-relevant summary. This distillation ensures the model remains focused on the user's current intent. To maintain high quality at scale, knowledge was distilled from a "teacher" model (GPT-4o) into a more efficient 21B-parameter model (ERNIE Speed).

Phase 2: Click-oriented Reinforcement Learning (C-RL)

To align the model with actual user behavior, the system employs a generate-rank paradigm. Diverse Beam Search (DBS) generates candidate suggestions, which are then evaluated by a Click Estimator trained on real-world interaction data. Finally, Direct Preference Optimization (DPO) is applied to prioritize suggestions that are engaging, diverse, and useful to the end-user.

The dual-phase architecture of the proactive guidance framework, featuring G-SFT and C-RL.

Experimental Results

The framework was evaluated using proprietary industrial-scale data from the Baidu Search AI assistant. Offline evaluations utilized a dataset of over 6,000 high-quality QA pairs, while the C-RL phase was trained on 12,000 preference pairs derived from user click signals.

Performance and Engagement

  • Accuracy: The system achieved 86.10% in offline evaluations, marking a 23.95% improvement over the baseline.
  • User Engagement: Online testing showed a Click-Through Rate (CTR) of 25.28%, a 149.06% relative improvement over the previous production system.
  • Efficiency: By replacing raw history with concise summaries, inference latency was reduced by 69.55% (from ~2.89s to ~0.88s).
Ablation study showing the incremental performance gains from individual components including Summary (S), Shift Detection (SD), and the Goal Adaptation Agent (GAA).

Conclusion

Baidu's proactive guidance framework demonstrates that combining intent-aware summarization with click-driven reinforcement learning can bridge the gap between LLM reasoning and industrial latency requirements. While the current system effectively handles goal shifts, future work will focus on more sophisticated state-tracking to manage temporary pivots where users return to previous topics after a brief diversion.

Summary updated

BibTeX
@misc{li2025proactive,
  author = {Xiaoyu Li and Xiaoli Li and Li Gao and Yiding Liu and Xiaoyang Wang and Shuaiqiang Wang and Junfeng Wang and Dawei Yin},
  title = {{Proactive Guidance of Multi-Turn Conversation in Industrial Search}},
  year = {2025},
  doi = {10.18653/v1/2025.acl-industry.50},
}
Figure