Proactive Guidance of Multi-Turn Conversation in Industrial Search (Baidu, arXiv 2505.24251)

Forum topic · 小凯 · 2026-07-05

Summary

This paper, from Baidu researchers (arXiv:2505.24251, May 2025), presents a two-phase framework for proactive guidance in multi-turn conversational search, deployed in the Baidu Search AI assistant. The first phase, Goal-adaptive Supervised Fine-Tuning (G-SFT), uses a goal adaptation agent that dynamically tracks user goal shifts and supplies goal-relevant context, while scalable knowledge distillation transfers LLM insights into a lightweight model suitable for real-time interaction. The second phase, Click-oriented Reinforcement Learning (C-RL), adopts a generate-rank paradigm that builds preference pairs from user click signals to produce more engaging guidance and improve click-through rates. Reported results include 86.10% offline accuracy (+23.95% over baseline), 25.28% online CTR (149.06% relative improvement), and a 69.55% reduction in inference latency. The writeup also contextualizes the work within agentic search and RAG trends, discusses engineering constraints such as latency budgets, data safety, and evaluation gaps between offline metrics and online satisfaction, and cross-references related surveys on multi-turn LLM interaction, query rewriting, and conversational recommendation.

Proactive Guidance of Multi-Turn Conversation in Industrial Search

Authors: Xiaoyu Li, Xiao Li, Li Gao, Yiding Liu, Xiaoyang Wang, Shuaiqiang Wang, et al. (8 authors total, Baidu)
Published: 2025-05-30
Source: https://arxiv.org/abs/2505.24251
Category: Multi-Turn / Agentic Search

Overview

LLMs have significantly advanced multi-turn conversation systems, but industrial deployments still struggle with dynamically adapting to shifts in user goals and maintaining low latency for real-time interaction. This paper proposes a two-phase framework for proactive guidance in the Baidu Search AI assistant, an industrial-scale multi-turn search system.

Method: Two-Phase Framework

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

A goal adaptation agent dynamically adapts to user goal shifts and provides goal-relevant contextual information.
Scalable knowledge transfer distills insights from large LLMs into a lightweight model, enabling real-time interaction with reduced latency.

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

Adopts a generate-rank paradigm.
Systematically constructs preference pairs from user click signals.
Proactively improves click-through rates through more engaging guidance.

The two phases are complementary: G-SFT ensures accurate goal tracking, while C-RL optimizes interaction quality via click-signal-driven reinforcement learning.

Results

| Metric | Result | |--------|--------| | Offline accuracy | 86.10% (+23.95% over baseline) | | Online CTR | 25.28% (149.06% relative improvement) | | Inference latency | Reduced by 69.55% via knowledge distillation |

Original Abstract

> The evolution of Large Language Models (LLMs) has significantly advanced multi-turn conversation systems, emphasizing the need for proactive guidance to enhance users' interactions. However, these systems face challenges in dynamically adapting to shifts in users' goals and maintaining low latency for real-time interactions. In the Baidu Search AI assistant, an industrial-scale multi-turn search system, we propose a novel two-phase framework to provide proactive guidance. The first phase, Goal-adaptive Supervised Fine-Tuning (G-SFT), employs a goal adaptation agent that dynamically adapts to user goal shifts and provides goal-relevant contextual information. G-SFT also incorporates scalable knowledge transfer to distill insights from LLMs into a lightweight model for real-time interaction. The second phase, Click-oriented Reinforcement Learning (C-RL), adopts a generate-rank paradigm, systematically constructs preference pairs from user click signals, and proactively improves click-through rates through more engaging guidance. [...] Extensive experiments demonstrate that our framework achieves 86.10% accuracy in offline evaluation (+23.95% over baseline) and 25.28% CTR in online deployment (149.06% relative improvement), while reducing inference latency by 69.55% through scalable knowledge distillation.

Context and Takeaways

Architecture: Cascaded retrieve-rank-generate remains mainstream, but the agentic paradigm treats retrieval count and strategy itself as learnable.
Data: High-quality instruction data and click/session logs are equally critical; synthetic data requires safeguards against leakage and distribution shift.
Evaluation: The gap between offline metrics and online satisfaction is widening; LLM-as-judge should be cross-validated with human evaluation.
Production: Latency, cost, interpretability, and safety are hard constraints for industrial deployment—academic benchmarks alone are insufficient.