Reproduce “Beyond Passive Critical Thinking: Fostering Proactive Questioning to Enhance Human-AI Collaboration” on OpenResearch

About this paper

Ante Wang, Yujie Lin, Jingyao Liu, Suhang Wu, Hao Liu, Xinyan Xiao, Jinsong Su

Researchers from Xiamen University, Baidu Inc., and Shanghai Artificial Intelligence Laboratory introduced "proactive critical thinking" for large language models, enabling them to actively seek clarifications for imperfect user queries instead of passively rejecting them. This approach, evaluated on new mathematical reasoning benchmarks, substantially improved model accuracy on unanswerable questions while preserving general reasoning capabilities.

Key insights

• Proactive critical thinking is a trainable skill, with significant performance gains achievable through SFT and RL, demonstrating that LLMs can learn to intelligently ask clarifying questions.
• The "thinking mode" in LLMs, initially detrimental or neutral in vanilla models, becomes a beneficial asset for flaw analysis and effective question formulation after targeted RL training.
• High-quality, filtered training data is crucial for effectively instilling complex skills like proactive critical thinking, as unfiltered data leads to degraded model performance.

Summary by alphaXiv.