# Self-Adapting Language Models

**Authors:** Adam Zweiger, Jyothish Pari, Han Guo, Ekin Akyürek, Yoon Kim, Pulkit Agrawal  
**arXiv:** [2506.10943](https://arxiv.org/abs/2506.10943)  
**Date:** 2025-06-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.10943v1/

## Key points

- Large language models (LLMs) lack the ability to adapt their weights in response to new tasks, knowledge, or examples.
- Self-Adapting LLMs (SEAL) is a framework that enables LLMs to self-adapt by generating their own finetuning data and update directives.
- SEAL allows the model to produce self-edits when given a new input, involving restructuring information, specifying optimization hyperparameters, and utilizing tools for data augmentation and gradient-based updates.
- Through supervised finetuning (SFT), these self-edits result in persistent weight updates, facilitating lasting adaptation.
- SEAL uses a reinforcement learning loop with downstream performance as the reward signal to train the model to generate effective self-edits.
- Experimental results show promise in enabling language models capable of self-directed adaptation through SEAL.
- Acknowledgments are made to various individuals and funding sources for support in conducting the research.
- SEAL represents a significant advancement in enabling large language models to adapt autonomously through self-generated training data and update directives.

## Summary

Large language models (LLMs) are powerful tools for language understanding and generation. However, they lack the ability to adapt their weights in response to new tasks, knowledge, or examples. To address this limitation, Self-Adapting LLMs (SEAL) has been introduced as a framework that enables LLMs to self-adapt by generating their own finetuning data and update directives. This allows the model to produce self-edits when given a new input, which can involve restructuring information, specifying optimization hyperparameters, or utilizing tools for data augmentation and gradient-based updates. Through supervised finetuning (SFT), these self-edits result in persistent weight updates, facilitating lasting adaptation. To train the model to generate effective self-edits, a reinforcement learning loop is employed with the downstream performance of the updated model serving as the reward signal. Unlike previous approaches that rely on separate adaptation modules or auxiliary networks, SEAL directly utilizes the model's own generation to control its adaptation process. Experimental results on knowledge incorporation and few-shot generalization demonstrate that SEAL shows promise in enabling language models capable of self-directed adaptation. Acknowledgments are made to various individuals for their valuable discussions and feedback, as well as support from ARO MURI grant number W911NF-23-1-0277 and other funding sources. The research was conducted under Cooperative Agreement Number FA8750-19-2-1000 with contributions from the Stevens Fund for MIT UROP research and the MIT-IBM Watson AI Lab. In conclusion, SEAL represents a significant advancement in enabling large language models to adapt autonomously through self-generated training data and update directives. The framework shows promise in enhancing language models' capabilities for adapting to new tasks and incorporating new knowledge effectively. More information about SEAL can be found on their website at https://jyopari.github.io/posts/seal.