MIT researchers enable LLMs to dynamically adjust computation for harder problems, boosting efficiency.
MIT researchers have developed a novel method for large language models (LLMs) to dynamically allocate computational resources as they reason through problems. Their approach allows models to spend more compute on difficult questions and promising solution paths, while using fewer resources on easier tasks. This is achieved by integrating a process reward model (PRM) that scores partial solutions and guides the LLM to focus on the most viable reasoning paths in real time.
The technique, called inference-time scaling, lets LLMs generate multiple solution attempts and select the best ones, adapting on the fly rather than at the outset. This not only reduces computational costs by up to 50% but also enables smaller models to match or outperform larger ones on complex tasks. The method is already influencing frontier models like GPT-5.1, which has adopted similar adaptive reasoning strategies. Future applications could include code generation, AI agents, and reinforcement learning, making this a significant leap in both efficiency and scalability for LLMs.
Source: MIT News