PIR: Teach Reasoning Models to Ask Proactively and Bid Farewell to Blind Self-Thinking

PIR (Proactive Interactive Reasoning) is a new paradigm that transforms reasoning models from passive solvers into active inquirers through uncertainty detection and reinforcement learning, significantly reducing over-reasoning and hallucination issues.

The original authors are the Chen-X666 team (Xin Chen, Feng Jiang, Yiqian Zhang, etc.), the project was released in May 2026, the paper was accepted by ACL 2026 (arXiv:2601.22139), GitHub link: https://github.com/Chen-X666/Proactive-Interactive-R1.

Current mainstream reasoning models (such as GPT-o1, DeepSeek-R1) have significant flaws when facing complex problems: even if key information is missing or user intent is ambiguous, they still continue internal reasoning, eventually leading to overthinking, hallucinations, or conclusions that do not match user needs. This phenomenon is called 'blind self-thinking'.

Analogy: A student who, when encountering an unfamiliar question in an exam, writes randomly based on feeling instead of asking for clarification.

The PIR framework forms a 'think-ask-respond' loop through two stages:

1. Interaction Capability Activation: Use prediction entropy to detect uncertainty in reasoning; when the entropy value exceeds a threshold, actively ask questions. Train the model to learn when to ask through supervised fine-tuning (SFT);
2. User Intent Alignment: Introduce the US-GRPO (User Simulator-augmented Group Relative Policy Optimization) algorithm, where the reward function considers three dimensions: output accuracy, reasoning efficiency, and question usefulness.

The paper reports that PIR performs excellently in multiple benchmark tests:

• Accuracy increased by 32.70%;
• Pass rate increased by 22.90%;
• BLEU score increased by 41.36;
• Reasoning tokens and interaction rounds reduced by nearly half.

It not only improves performance but also reduces computational costs and interaction burdens.

The project is based on the LLaMA-Factory, verl, and verl-tool frameworks. The training process has two steps:

1. Cold-start SFT: Fine-tune the base model using 'reasoning-asking' data;
2. US-GRPO Reinforcement Learning: Optimize the strategy in a user simulator environment.

Pre-trained models (such as Proactive-Interactive-R1-Math-7B) are available for download from Hugging Face, supporting vLLM deployment for interactive demos.

PIR redefines human-machine interaction: from one-way user question → model answer to two-way dialogue, where the model actively clarifies needs, similar to how human experts ask for details.

Applicable fields expand to mathematical reasoning, code generation, medical diagnosis, legal consultation, etc. In the future, more intelligent assistants that ask proactively are expected to emerge, bidding farewell to blind guessing.