In their study titled LM2, researchers argue that Large Language Models often stumble over complex reasoning tasks due to lack of coordination between problem decomposition and solution.

• Decomposition Module: Creates step-by-step questions based on reasoning requirements.
• Solution Module: Addresses the broken-down problems with guidance.
• Verification Module: Checks the validity of solutions, refining reasoning process.
• Policy Learning-Based Coordination: Enables effective communication between modules.

The innovative use of a society of language models offers a window into more complex applications for LLMs in problem-solving, enhancing their reasoning by drawing out the best of collaborative efforts.