MIT Announces Self-Updating AI Foundation 'SEAL'

Researchers at MIT have announced 'SEAL,' a framework that enables large language models (LLMs) to rewrite their own parameters—the collection of numerical values that record internal knowledge and decision criteria. Conventionally, AI models have had their internal knowledge fixed once training is completed. SEAL transcends this constraint and provides a mechanism for models to continuously update themselves.

At the core of SEAL is a technique called reinforcement learning. Reinforcement learning is a mechanism in which an AI receives feedback in the form of rewards or penalties for the results of its actions and learns progressively better behaviors; it has been widely used in game AI and other applications. In SEAL, this mechanism is applied to the action of a model editing its own parameters. In other words, the model can autonomously learn through trial and error how to rewrite itself to improve its outputs.

In conventional AI model operations, updating knowledge required human engineers to prepare new training data and perform retraining—a process that is time-consuming and costly. As a result, models struggle to keep their knowledge current with real-world changes. If SEAL becomes practical, the possibility exists that models could autonomously update themselves, thereby significantly reducing such costs and delays.

On the other hand, the nature of AI rewriting its own parameters requires careful consideration from the perspective of safety and controllability. If unintended self-updates occur, there is a risk that model behavior could become unpredictable. In moving this research toward practical implementation, a critical challenge is determining how to limit the scope of self-updates and ensure reliability.

As an institution at the forefront of AI research, MIT has produced numerous achievements in the fields of model architecture and learning methods. This SEAL research challenges the previous assumption that 'trained models are static' and is drawing researcher interest as it offers new directions for future AI system design.