AIPO: : Learning to Reason from Active Interaction

Short summary

AIPO is a reinforcement learning framework that enhances LLM reasoning by enabling models to consult specialized collaborative agents when encountering reasoning bottlenecks. Unlike methods requiring complete trajectory guidance, AIPO uses fine-grained agent feedback to efficiently expand capability boundaries. Experiments across multiple reasoning benchmarks show consistent improvements with robust generalization.

• •Multi-agent interaction framework (Verify, Knowledge, Reasoning agents) guides LLM exploration
• •Improves reasoning capability boundary without sample-inefficient trajectory-level supervision
• •Validated across AIME, MATH500, GPQA-Diamond, LiveCodeBench with consistent gains

Generated with AI, which can make mistakes.