Meta AI Unveils Coral Framework for Enhanced Collaborative Reasoning

Meta AI has introduced the Collaborative Reasoner (Coral), a new framework aimed at enhancing collaborative reasoning skills in large language models (LLMs), as detailed in a company publication. Coral is designed to address the limitations of current LLMs, which excel in single-agent tasks but struggle with multi-agent interactions that require consensus-building and negotiation.

Coral reformulates traditional reasoning tasks into multi-agent, multi-turn dialogues, where agents must reach consensus through natural conversation. This approach emulates real-world social dynamics, requiring agents to challenge incorrect conclusions and negotiate conflicting viewpoints. The framework spans five domains, including mathematics and social cognition, serving as testbeds for evaluating collaborative reasoning.

To support the generation of training data, Meta AI has developed Matrix, a high-performance serving framework that facilitates large-scale data generation. This infrastructure supports the self-collaboration approach, where a single LLM plays both roles in a conversation, generating synthetic dialogues for training purposes.

Empirical results show that models fine-tuned with Coral outperform baseline single-agent approaches, demonstrating improved generalization across various tasks. However, challenges remain in domains requiring deep symbolic reasoning, indicating that collaboration alone may not suffice for complex mathematical problems.