Reasoning While Asking: Transforming Reasoning Large Language Models from Passive Solvers to Proactive Inquirers

Researchers introduce Proactive Interactive Reasoning (PIR), a new paradigm that enables large language models to ask clarifying questions during problem-solving rather than operating blindly with incomplete information. The approach combines supervised fine-tuning and policy optimization to achieve significant improvements in mathematical reasoning, code generation, and document editing tasks while reducing computational overhead.

The research addresses a fundamental limitation in current reasoning-focused LLMs: their tendency to proceed with internal problem-solving even when premises are unclear or information is ambiguous. Traditional Chain-of-Thought approaches generate extensive reasoning chains regardless of data quality, leading to inefficient computation and potentially flawed outputs. PIR reimagines the reasoning process as an interactive dialogue where models identify uncertainties and proactively request clarification before committing to solutions.

This work builds on years of LLM capability development where researchers discovered that more sophisticated reasoning emerged through prompting techniques and process-oriented training. However, previous approaches predominantly focused on external knowledge retrieval or tool use to address information gaps. PIR's innovation lies in targeting premise-level and intent-level uncertainty through direct user interaction, treating clarification as an integral reasoning component rather than an auxiliary feature.

The practical implications span multiple domains. In mathematical problem-solving, models achieving 32.70% accuracy improvements demonstrate tangible gains in correctness. The 22.90% pass rate increase in code generation suggests real utility for software development workflows. Critically, reducing computational requirements by nearly fifty percent addresses efficiency concerns that plague enterprise LLM deployments. The approach also cuts unnecessary interaction turns, preserving user experience quality.

Looking ahead, PIR's success in factual knowledge and question-answering tasks suggests broader applicability beyond the initially tested domains. The public release of model and code enables rapid community adoption and refinement. Future developments may explore how this paradigm scales to more complex multi-turn interactions and domain-specific reasoning tasks, potentially influencing how enterprises design LLM-powered systems.

• →Proactive Interactive Reasoning enables LLMs to request clarification on ambiguous premises, improving accuracy by up to 32.70% across multiple task domains.
• →The approach reduces computational reasoning requirements by approximately 50% while minimizing unnecessary user interactions.
• →PIR demonstrates strong generalization across mathematical reasoning, code generation, document editing, and factual knowledge tasks.
• →The method targets premise and intent uncertainty through direct user dialogue rather than external tool queries.
• →Publicly available implementation at GitHub accelerates potential industry adoption and further research development.