Learning Effective Agentic Organization Policies

Develop algorithms that learn effective organization policies—the strategies for coordinating multiple large language model agents to collaborate concurrently—so that optimal thinking structures are discovered automatically across diverse queries without manual design for each task.

Background

The paper envisions agentic organization, where multiple agents based on LLMs collaborate concurrently to solve complex problems. While parallel thinking methods typically rely on fixed, manually designed workflows, the authors emphasize that adaptivity and dynamism are essential for such systems, and designing optimal structures for every query is intractable.

AsyncThink is introduced as a protocol and training framework (including reinforcement learning) to organize internal thinking via Fork-Join actions. However, the broader challenge of learning effective organization policies in general-purpose settings is explicitly stated as an open problem.