The Era of Agentic Organization: Learning to Organize with Language Models (arxiv.org)

The paper proposes "agentic organization," a new paradigm where language models orchestrate multiple cooperating agents to solve complex problems. Central to this is AsyncThink, an asynchronous thinking protocol that breaks internal reasoning into concurrently executable pieces: an organizer dynamically assigns sub-queries to worker agents, collects and merges intermediate results, and composes a final answer. Crucially, the organizer’s thinking graph—how tasks are split, scheduled, and merged—is not fixed but learned via reinforcement learning, enabling the system to discover efficient, task-specific organization strategies. AsyncThink delivers concrete benefits: it cuts inference latency by 28% versus a baseline parallel-thinking setup while boosting accuracy on mathematical reasoning benchmarks, and the learned asynchronous strategies transfer to unseen tasks without retraining. For the AI/ML community, this points to a scalable alternative to serial chain-of-thought: modular, concurrent workflows that can be optimized end-to-end. The work suggests practical pathways for building multi-agent LLM systems that are faster, more accurate, and more adaptable—opening avenues for generalizable orchestration layers that learn how best to decompose, parallelize, and merge reasoning across problems.