Distilling LLM Reasoning into an Interpretable Policy Tree for Human-AI Collaboration

Researchers introduce Collaboration Policy Tree (Co-pi-tree), a method that distills large language model reasoning into interpretable, executable policy trees for human-AI collaboration. The approach achieves 35% performance improvement while reducing LLM queries by 78% and latency by 97%, addressing key limitations of black-box reinforcement learning and costly real-time LLM querying.

Co-pi-tree represents a meaningful advancement in making AI systems both more interpretable and computationally efficient for collaborative tasks. The core innovation lies in converting LLM reasoning—typically opaque and expensive to execute at scale—into structured policy trees that can be understood and validated by humans. This addresses a critical tension in modern AI development: the trade-off between capability and interpretability.

The research emerges from growing concerns about deploying black-box reinforcement learning systems in human-facing applications. MARL policies offer strong performance but provide no visibility into decision-making logic, creating safety and trust issues. Conversely, querying LLMs at every decision point delivers better interpretability but becomes prohibitively expensive and slow in production environments. Co-pi-tree bridges this gap through a closed-loop refinement process where initial LLM reasoning is distilled into executable code, validated through interaction, and iteratively improved using natural language feedback.

For the broader AI ecosystem, this approach signals growing maturity in AI safety and human-AI collaboration research. The 77.7% reduction in LLM queries has direct economic implications, reducing inference costs while maintaining reasoning quality. The 97.1% latency improvement makes real-time collaborative AI systems practically viable.

The experimental validation in Overcooked-AI demonstrates the method's effectiveness in cooperative multi-agent scenarios, but scaling to more complex real-world domains remains an open question. Future work should explore applicability to physical robotics, autonomous systems, and enterprise workflows where both interpretability and efficiency determine adoption feasibility.

• →Co-pi-tree reduces LLM query costs by 78% while maintaining 35% performance improvement over baselines through policy tree distillation.
• →The method converts opaque language model reasoning into human-interpretable policy trees, addressing safety and transparency concerns in AI systems.
• →Latency improvements of 97% enable real-time human-AI collaboration previously impossible with querying-at-each-step LLM approaches.
• →Closed-loop refinement using natural language feedback allows iterative policy improvement without retraining from scratch.
• →The approach bridges the interpretability-efficiency gap critical for deploying AI systems in safety-sensitive collaborative environments.