Researchers at the Massachusetts Institute of Technology (MIT) have developed and open-sourced a technique called SEAL (Self-Adapting LLMs) that enables large language models (LLMs) to autonomously enhance their capabilities by generating synthetic data for fine-tuning. This method was detailed in an initial paper released in June, with an updated version published last month along with code available on GitHub under the MIT License.
SEAL allows LLMs to create and implement their own strategies for fine-tuning without relying on static external data or human-designed optimization pipelines. Instead, these models produce synthetic training data and optimization directives, enabling them to adapt and evolve according to their learning needs. The research, affiliated with MIT’s Improbable AI Lab, involved contributions from several researchers and was presented at the 39th Conference on Neural Information Processing Systems (NeurIPS 2025).
The updated SEAL framework demonstrates that self-adaptation improves with model size and effectively incorporates reinforcement learning to minimize the issue of catastrophic forgetting. This version also includes evaluations across various prompting formats and addresses practical deployment challenges.
SEAL has been tested mainly in knowledge incorporation and few-shot learning domains. In one study, performance on a comprehension task improved significantly after the model generated synthetic implications of text rather than relying solely on direct fine-tuning. Additionally, the success rate in few-shot learning increased markedly after applying reinforcement learning techniques.
The SEAL framework uses a two-loop structure: one for supervised fine-tuning based on generated self-edits, and another for refining these edits through reinforcement learning. Although SEAL shows promise, it also faces challenges such as computational overhead and the risk of catastrophic forgetting. The framework currently assumes the availability of paired tasks and reference answers, which may limit its applicability.
Experts in the AI community have shown interest in SEAL’s potential, with some describing it as a groundbreaking step towards continuous self-learning AI systems. Future research may explore its adaptability across different models and tasks, aiming to establish its efficacy in diverse and less supervised environments.
Source: https://venturebeat.com/ai/self-improving-language-models-are-becoming-reality-with-mits-updated-seal