Position: Limitations of LLMs Can Be Overcome by Carefully Designed Multi-Agent Collaboration

Edward Y Chang, Arihant Choudhary, Parth Behani, Yash Vardhan Pansari

TL;DR: LLMs and agent systems lack reasoning capabilities for planning tasks. MACI's meta-planning architecture achieves System-2 planning through specialized agents.

Abstract:

Position Statement:
Current Large Language Models (LLMs) face three fundamental limitations: (1) reliance on pattern matching rather than deliberate reasoning, (2) inability to self-validate their output, similar to Gödel’s incompleteness constraints, and (3) inconsistent constraint management in planning tasks. These deficiencies prevent LLMs from achieving system-2 level reasoning and planning. We introduce Multi-Agent Collaborative Intelligence (MACI), a structured framework designed to overcome these challenges through meta-planning and distributed validation. MACI comprises three key components: (1) a metaplanner (MP) that formulates and refines all task roles and constraints while generating a dependency graph augmented with common-sense reasoning; (2) a collection of specialized agents to facilitate domain-specific planning and task execution; and (3) a run-time monitor that dynamically adjusts plans as needed. By structuring problem solving into specialized roles and coordinating agent collaboration, MACI enables robust constraint awareness, self-verification, and adaptability, capacities absent in monolithic LLM architectures. The experimental results validate the effectiveness of MACI in improving planning consistency and satisfaction with constraints.

Primary Area: Research Priorities, Methodology, and Evaluation

Keywords: Temporal Planning, system-2 reasoning, AGI, multi-agent collaboration

Submission Number: 129