Imagine a world where artificial intelligence never stops learning, just like a diligent student with an endless supply of fresh notebooks.
Researchers at the Massachusetts Institute of Technology have moved a step closer to this possibility with a new approach that lets large language models find ways to teach themselves from new experiences. Simply put, the objective is to let these models adapt on the fly, improving not just from their massive initial data pool but also by ingesting fresh information as it comes.
Their novel system, Self Adapting Language Models, or SEAL, helps language models digest useful new knowledge by generating their own training materials and updating their internal mechanics. The spark for this idea came from asking whether a model’s own words could be recycled to sharpen its logic and broaden its understanding. MIT doctoral student Jyothish Pari explains, “The initial idea was to explore if tokens could cause a powerful update to a model.”
SEAL lets these models whip up synthetic passages in response to new input, then uses those passages as training ground for the model to grow. As a way to check SEAL’s effectiveness, the team had it summarize difficult moments in history, like the trials of the Apollo space missions, and then tested whether the AI could extract and apply new insights. Adam Zweiger, an undergraduate on the project, points out that while today’s AI can perform complex reasoning, those flashes of brilliance rarely become permanent features — SEAL is designed to change that.
Teaching Machines to Learn from Themselves
Pari and his colleagues likened SEAL’s approach to the tried and true method of a student reviewing class notes to lock in new material. After the language model generates and reviews these synthetic notes, it immediately gets quizzed with new questions, measuring how much has stuck and pushing the learning process a step forward.
To put SEAL through its paces, the team experimented with open-source models, including versions of both Meta’s Llama and Alibaba’s Qwen. They found the approach worked not only for regular text but also for puzzles that require abstract thinking, showing SEAL’s knack for extending learning beyond what the models originally knew.
MIT professor Pulkit Agrawal, who guided the project, sees a promising future. “LLMs are powerful but we don’t want their knowledge to stop,” he emphasizes. SEAL, he says, could help make AI tools much better at personalizing to users and situations, a quality in short supply among models trained just once.
That said, there are hurdles. As these models consume new tidbits, they sometimes forget old ones, a problem called catastrophic forgetting that remains unsolved. SEAL also demands a lot from computer resources, making it tricky to know when and how to schedule these growing periods. Zweiger suggests maybe one day AI models could set aside quiet moments, akin to sleep, to help integrate new knowledge.
With all the excitement, SEAL is still experimental, but it hints at a future where AI that learns continuously could become reality, as seen with recent progress in shrinking large language models for ongoing adaptation.
