Researchers at the Massachusetts Institute of Technology (MIT) have developed a technique that allows large language models to improve on their own. Carl Franzen reports for VentureBeat that the method, called SEAL, enables AI to autonomously generate its own training data.
Instead of relying on fixed external datasets, models using the SEAL framework create instructions for themselves called “self-edits”. These edits specify how the model should update its internal knowledge. The model then fine-tunes itself based on these instructions. A process known as reinforcement learning guides this self-improvement, rewarding changes that lead to better performance on specific tasks.
The researchers report significant performance gains with this approach. In one test involving reading comprehension, the model’s accuracy increased from 33.5% to 47.0%. For another task requiring reasoning from a few examples, the success rate rose to 72.5%. The system even surpassed results from synthetic data generated by the advanced GPT-4.1 model in one setting.
While the technique shows promise, challenges remain. The process is computationally intensive, and there is a risk that the model might forget previously learned information, a problem known as catastrophic forgetting. The researchers state that reinforcement learning helps to reduce this issue.
SEAL is considered a key step toward more adaptive AI systems. Such models could continuously evolve and integrate new knowledge without constant human supervision, making them more useful in rapidly changing environments. The project’s code has been released with an open-source license.
