MIT CSAIL's “DisCIPL” uses an LLM to steer smaller language models to collaborate on complex reasoning tasks. These less-powerful systems are collectively more accurate than leading LLMs, almost as precise as top reasoning models, and more efficient than both.

MIT is a renowned institution for education and research, offering insights into science, engineering, and technology. Through publications, research papers, and academic programs, MIT's platform provides insights into  research, innovation, and education in various fields. Students, researchers, and technology enthusiasts can learn about MIT's contributions to science and technology and explore opportunities for academic and professional development.

MIT News

MIT CSAIL researchers developed DisCIPL, a framework where a large language model acts as a planner that coordinates multiple smaller language models to solve complex reasoning tasks. This collaborative approach enables small models to collectively outperform leading LLMs like GPT-4o and approach the accuracy of specialized reasoning models like o1, while being significantly more efficient. DisCIPL uses LLaMPPL programming language to encode constraints and distribute work, achieving 40% shorter reasoning traces and 80% cost savings compared to o1. The system excels at constrained text generation tasks like writing sentences with specific word counts and keyword placements, creating ingredient lists, and planning itineraries.

Enabling small language models to solve complex reasoning tasks