Teaching large language models how to absorb new knowledge (news.mit.edu)

MIT researchers introduced SEAL (self-adapting LLMs), a framework that teaches deployed large language models to permanently absorb new information by generating and learning from their own “study sheets.” Instead of relying on short-lived in-context examples, the LLM rewrites user input into multiple synthetic self-edits, tests each by quizzing itself on downstream tasks, and uses reinforcement learning to reward the edits that boost performance most. The winning self-edit is then used to update the model’s weights, effectively memorizing the new knowledge and even letting small models outperform much larger ones. Technically, SEAL couples synthetic-data generation with adaptive optimization: the model selects which edits to use, how aggressively to learn (learning rate), and how many update iterations to apply. The method yielded substantial gains—about a 15% accuracy increase on question-answering and over 50% improvement on some skill-learning tasks—across several baselines. Key limitations remain, notably catastrophic forgetting as the model continuously adapts; the team plans mitigation strategies and multi-agent training experiments. Presented at NeurIPS by MIT CSAIL authors, SEAL points toward more autonomous, continually learning AI agents that can adapt safely to evolving user inputs and tasks.