Self-Adapting AI: SEAL Enables Language Models to Update Themselves

Researchers at MIT have introduced a new framework, SEAL, that allows large language models to update their own behavior by generating and applying their own training instructions. This approach, which combines reinforcement learning with supervised finetuning, marks a significant step toward models that can autonomously adapt to new information or tasks—without needing external retraining.

SEAL, short for Self-Adapting Language Models, enables a model to modify its internal weights using what the researchers call “self-edits”—custom-generated data and optimization directives produced by the model itself. These edits are then applied through supervised finetuning, allowing the model to adapt in a way that persists over time.

Unlike previous methods that rely on external modules or auxiliary networks to drive adaptation, SEAL uses the model’s own text generation capabilities to define how it should be updated. A reinforcement learning loop helps the model learn which kinds of self-edits lead to better performance, using success on downstream tasks as the reward signal.

How SEAL Works

At the heart of SEAL is a lightweight reinforcement learning algorithm called ReST-EM. Each training iteration involves:

• The model receiving a new task or input context.

• Generating a self-edit, which may include restructured information, hyperparameter settings, or data augmentation strategies.

• Applying that self-edit using supervised finetuning.

• Evaluating the updated model’s performance on a downstream task.

• Reinforcing successful edits that improve performance.

This loop teaches the model not only to generate training data, but to learn which data and training configurations are effective for different tasks.

The researchers tested SEAL in two domains:

• Knowledge Incorporation – where the model learns new facts from text passages and uses them to answer related questions.

• Few-Shot Learning – where the model generates its own training strategies to solve abstract reasoning tasks with minimal examples.

Results: Better Adaptation in Two Key Domains

• Knowledge Incorporation: In tasks requiring the assimilation of factual information, SEAL significantly improved performance over static baselines. After two rounds of training, the model raised question-answering accuracy from 32.7% to 47.0% in single-passage tests—outperforming models finetuned on either raw passages or synthetic data from GPT-4.1. In a larger-scale setup with 200 passages, SEAL again led the field with a 43.8% accuracy, showing that its self-editing policy scales well beyond its initial training conditions.
  

• Few-Shot Learning: On a simplified version of the ARC benchmark for abstract reasoning, SEAL achieved a 72.5% success rate. That’s compared to 0% for standard in-context learning and just 20% for models using untrained self-edits. These results highlight SEAL’s ability to configure effective training setups from scratch, enabling strong generalization from very limited data.

For detailed graphs, please click here.

Limitations and Open Challenges

While SEAL enables persistent self-modification, it also raises a key issue: repeated updates can cause the model to forget earlier knowledge. This "catastrophic forgetting" limits the model’s ability to retain prior information over time.

The researchers acknowledge this as an open problem in continual learning and suggest that future solutions might involve replay mechanisms, update constraints, or more robust internal representations to preserve older knowledge.

Looking Ahead: Toward Continually Improving Models

The MIT team envisions future models that not only generate their own updates, but can decide when adaptation is needed—potentially mid-inference. These models could turn ephemeral reasoning steps into permanent learning, gradually improving through ongoing interaction with new data and tasks.

This vision moves language models closer to becoming autonomous learners—systems capable of adjusting to changing environments not just by consuming new information, but by actively deciding how to evolve.

What This Means: A Step Toward Self-Improving AI

SEAL introduces a new paradigm for language models: one in which the model doesn’t just use data, but actively decides how to learn from it. This is a shift from static systems that rely on human-curated training to adaptable systems capable of self-directed improvement.

By giving models the tools to update their own weights through self-generated training instructions, SEAL could reduce the need for costly retraining pipelines and enable faster, more flexible adaptation in real-world settings. For example, a model could update itself after encountering new scientific findings, internal company documents, or evolving user behavior—without engineers needing to intervene.

This approach also opens the door to more agent-like systems—models that learn continually and improve based on their own reasoning traces, not just on external feedback. While challenges like memory retention remain, SEAL offers a practical starting point for developing models that evolve over time, much like human learners do.

This shift—from models trained by us to models training themselves—could reshape how intelligence is built, guided, and scaled.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.