Reproduce “Proactive Guidance of Multi-Turn Conversation in Industrial Search” on OpenResearch

About this paper

Xiaoyu Li, Xiao Li, Li Gao, Yiding Liu, Xiaoyang Wang, Shuaiqiang Wang, Junfeng Wang, Dawei Yin

Researchers from Baidu Inc. introduced a two-phase framework for proactive guidance in multi-turn conversational search, employing Goal-adaptive Supervised Fine-Tuning (G-SFT) for dynamic context management and Click-oriented Reinforcement Learning (C-RL) for optimizing user preferences. The developed system achieved a 149.06% relative increase in online Click-Through Rate and reduced inference latency to approximately 1 second for real-time industrial deployment.

Key insights

• Dynamic goal adaptation through a dedicated agent (GAA) is crucial for maintaining conversational coherence and relevance in multi-turn interactions, significantly improving guidance quality and user engagement.
• Scalable knowledge transfer from powerful LLMs to smaller, more compact models effectively reduces inference latency by 69.55% while preserving high guidance quality, enabling real-time industrial deployment.
• Optimizing guidance generation with click-oriented reinforcement learning, utilizing a generate-rank paradigm and diversity-aware sampling for DPO, directly enhances user engagement measured by Click-Through Rate.

Summary by alphaXiv.