MIT Unveils SEAL: The AI Framework That Self-Updates Without Waiting for Datasets

MIT researchers have introduced a framework called **Self‑Adapting Language Models (SEAL)**. This isn’t just another tool for training models; SEAL empowers AI to… learn independently from itself. A language model equipped with SEAL acquires new information and, instead of merely absorbing it and moving on, initiates a self-editing process.

The model autonomously:

— Reconstructs the acquired information in various ways.
— Produces its **own** data for fine-tuning based on this new information.
— Drafts its own optimization guidelines.
— Implements data augmentation and updates based on gradients.

Then, with these self-generated learning resources, the model employs reinforcement learning (reinforcement learning) through the **ReST^{EM}** algorithm developed by the team at Google DeepMind. The essence of ReST^{EM} lies in executing multiple training cycles and selecting the best outcomes. The model learns **efficiently**, filtering the top updates for itself.

The MIT team tested SEAL on Alibaba’s Qwen2.5-7B model. The results were remarkable. Initially, the «bare» **Qwen2.5-7B** (represented by the red line in the graph) lagged behind the seasoned **GPT-4.1** (shown by the green line). However, once SEAL was applied and the model underwent just two cycles of self-training, it not only caught up but surpassed GPT-4.1 in effectively assimilating new knowledge.

Although it may seem feasible to continue self-training indefinitely, the phenomenon of **catastrophic forgetting** began to emerge. In simpler terms, after numerous update cycles, the model started to lose track of previously learned information. Researchers acknowledge the need for a mechanism to ensure that the model retains old skills while acquiring new ones.

*Links:*

Official post and details: [Self‑Adapting Language Models (SEAL)](https://jyopari.github.io/posts/seal);

Article on arXiv: [2506.10943 Self‑Adapting Language Models](https://arxiv.org/abs/2506.10943);

Source code on GitHub: [Continual‑Intelligence/SEAL](https://github.com/Continual-Intelligence/SEAL).

Subscribe to our Telegram channel [BotHub AI News](https://t.me/bothub) to stay updated on the latest trends in the IT world.