United Minds or Isolated Agents? Exploring Coordination of LLMs under Cognitive Load Theory

The paper introduces CoThinker, a multi-agent framework designed under the principles of Cognitive Load Theory (CLT) to address performance limitations of Large Language Models (LLMs) on complex tasks. By distributing cognitive load through agent specialization and structured communication, CoThinker enhances performance on reasoning-heavy tasks while managing coordination overhead in low-intrinsic-load scenarios. This approach provides insights for practitioners on optimizing LLM interactions and system design to mitigate context rot and improve task execution efficiency.