Research map / Large language models / Reasoning and chain of thought

The Era of Agentic Organization: Learning to Organize with Language Models

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AsyncThink, an organizer-worker reasoning protocol where an LLM forks and joins concurrent sub-queries, trained with SFT cold start then GRPO-style RL with a thinking concurrency reward.

Problem

Parallel thinking methods aggregate independent traces with fixed workflows and incur latency from the slowest trace and aggregation; learning adaptive concurrent reasoning structures for LLMs remains open.

Approach

An LLM acts as both organizer and worker via text-level Fork/Join/Think/Answer actions, so it learns its own asynchronous thinking structure through RL with accuracy, format, and concurrency rewards, and latency is measured via a critical-path DP.

Key results

• On multi-solution countdown, AsyncThink reaches 89.0% All-Correct accuracy versus 68.6% and 70.5% for baselines.
• On AIME-24 and AMC-23, AsyncThink matches or exceeds parallel thinking at 38.7 and 73.3 accuracy with about 28% lower critical-path latency (1468 vs 2048 on AIME-24).
• Trained only on countdown, AsyncThink scores 89.4% on 4x4 Sudoku versus 84.2% for parallel thinking, at lower latency (2853 vs 3695).

Surprising findings

• Removing format SFT leaves the model stuck at thinking concurrency of 1/c despite successful Forks.
• A proposed leverage reward caused training instability and collapse to near-sequential behavior.

Limitations

• Latency metric is a critical-path abstraction, not measured wall-clock time.
• Experiments use a single 4B model and a small fixed agent pool capacity (c=2 or 4).
• Format SFT data relies on GPT-4o synthesis and rule filtering.

Experimental setup

Datasets: DeepScaleR Enigmata synthetic multi-solution countdown data GPT-4o-synthesized cold-start traces

Benchmarks: multi-solution countdown AIME-24 AMC-23 4x4 Sudoku (Enigmata) MMLU-Pro

Metrics: final-answer accuracy critical-path latency thinking concurrency ratio

Baselines: Sequential-Thinking Parallel-Thinking Parallel-R1 APR

Builds on: GRPO DeepSeek-R1 parallel thinking APR Parallel-R1

Closest papers by meaning

• Real-Time Reasoning Agents in Evolving Environments (ICLR 2026 · similarity 0.87)
• Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models (ICLR 2025 · similarity 0.87)
• Graph-enhanced Large Language Models in Asynchronous Plan Reasoning (ICML 2024 · similarity 0.87)
• AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.87)
• Multi-Agent Collaboration via Evolving Orchestration (NeurIPS 2025 · similarity 0.86)
• Scaling Large Language Model-based Multi-Agent Collaboration (ICLR 2025 · similarity 0.86)