Self-Adapting Language Models Is the Way to AGI? (jyopari.github.io)

MIT researchers introduced SEAL (Self-Adapting LLMs), a framework that lets a language model generate its own finetuning data and optimization instructions—called self-edits—and then apply those edits as supervised finetuning (SFT) to produce persistent weight updates. SEAL is trained with a lightweight reinforcement learning loop (ReST), where the model proposes self-edits, the model is updated with those edits, and downstream performance of the updated model is used as the reward. ReST uses rejection sampling to select high-reward edits for reinforcement, so the model learns to parameterize its own adaptation process without external adapters or auxiliary modules. SEAL shows promising gains on two tasks: knowledge incorporation (converting passages into finetuning examples) and few-shot learning (autonomously selecting augmentations and hyperparameters for ARC reasoning tasks). In experiments, two rounds of ReST-EM lifted QA accuracy from 32.7% to 47.0% in a single-passage setting and reached 43.8% in a 200-passage continued-pretraining test, outperforming baselines including GPT-4 generated data; on a simplified ARC subset SEAL hit 72.5% vs. 0% for in-context learning and 20% for untrained self-edits. A key limitation is catastrophic forgetting from repeated self-edits, highlighting the need for retention mechanisms (replay, constrained updates, etc.). SEAL points toward agentic models that decide when and how to adapt, potentially distilling chain-of-thought into lasting capabilities.