ALMC: Adaptive LLM-based Multi-Agent Collaboration Across Diverse Task Domains

Feng He, Tianqing Zhu, Bo Liu, Dayong Ye, Zhen Fang, Huajie Chen, Wanlei ZhouOpenReview 2026 · Submission to ICLR 2026 · Rejected · OpenReview

Abstract

Large language model-based multi-agent systems (LLM-MAS) are effective at solving complex tasks by coordinating specialized agents. However, existing frameworks rely on a small set of predefined scenarios with static role configurations and rigid collaboration structures, limiting their adaptability across diverse task domains. We propose the Adaptive LLM-MAS Collaboration (ALMC) framework, which dynamically recruits agents and configures collaboration patterns according to task demands through three collaborative components: a Manager Agent that synthesizes task-specific role compositions and an executable workflow, a Judge Agent that evaluates execution quality, and a Solution Optimizer Agent that persists and reuses high-quality configurations via retrieval-augmented generation. The framework supports human-in-the-loop review and creates a learning loop where previous superior configurations improve future executions on similar tasks. By using ALMC, collaborations become adaptive, auditable, and reusable across domains. Code is available at: https://anonymous.4open.science/r/ALMC-2E0F.
BibTeX
@misc{he2026almc,
  author = {Feng He and Tianqing Zhu and Bo Liu and Dayong Ye and Zhen Fang and Huajie Chen and Wanlei Zhou},
  title = {{ALMC: Adaptive LLM-based Multi-Agent Collaboration Across Diverse Task Domains}},
  year = {2026},
  url = {https://openreview.net/forum?id=jXZGgxTjiK},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2026},
}

Loading related content...

Figure