MIT Researchers Introduce SEAL: A Step Toward Self-Improving AI

Researchers from the Massachusetts Institute of Technology (MIT) have introduced the SEAL (Self-Adapting LLMs) framework, which allows large language models (LLMs) to update their own weights. SEAL uses reinforcement learning (RL) to generate "self-edits" — synthetic data on which the model fine-tunes itself. An outer RL loop optimizes the edit generation process, while an inner loop applies edits via gradient descent. For training stability, the authors chose DeepMind's ReST^EM method over PPO and GRPO. The framework was tested in two scenarios: few-shot learning (on Llama-3.2-1B-Instruct) and knowledge integration (on Qwen2.5-7B). In few-shot learning, adaptation accuracy reached 72.5%, compared to 20% for the model without RL. For knowledge integration, SEAL outperformed the baseline even with synthetic data, and after two RL iterations it surpassed approaches using GPT-4.1. The work sparked widespread discussion, including on Hacker News, amid growing interest in AI self-evolution — recent projects from Sakana AI, CMU, SJTU, and others have emerged. OpenAI's Sam Altman also wrote about the future of self-improving AI and robots in his blog "The Gentle Singularity." The article notes SEAL's limitations, including catastrophic forgetting and computational costs.