According to MIT researchers, LLMs will soon no longer need static datasets or human intervention. Through the SEAL technique, they improve themselves.
Researchers at the Massachusetts Institute of Technology (MIT) have developed the Self-Adapting LLMs (SEAL) technique. This allows LLMs like ChatGPT to improve themselves by generating their own training data.
Self-improving AI
Unlike classical models that rely on static datasets and human training, SEAL can tune itself with synthetic data it creates. The model formulates ‘self-edits’: natural language descriptions indicating how it should adjust its knowledge.
The method uses two loops, reports VentureBeat. In the inner loop, the model tunes itself based on the self-edits, while the outer loop uses reinforcement learning to learn which adjustments effectively improve performance. This allows the model to continuously evolve without human intervention.
‘Performance beyond Expectations’
In a paper, MIT shares its test results. SEAL improved accuracy in question-and-answer sessions from 33.5 percent to 47 percent, better than synthetic data from GPT-4.1. Also, in few-shot learning tasks, where the model learns from a few examples, the success rate increased from 20 percent to 72.5 percent after applying reinforcement learning.
Although SEAL is still experimental and requires enormous computing power, researchers consider it a breakthrough towards self-adapting AI systems. The availability of data on the web is declining, making these autonomous techniques crucial for the further evolution of LLMs.
read also