"Models that modify their own weights mid-inference destroy traceability, break safety reinforcement, and create trivial attack vectors for manipulating outputs at scale."

Self-Adapting Language Models (SEAL) enable large language models to modify their weights during inference by generating self-edits for finetuning. This framework uses reinforcement learning to improve model performance on tasks like knowledge incorporation and few-shot learning. Results show significant improvements in accuracy and adaptability, though challenges like catastrophic forgetting remain. Future developments aim to enhance models' reasoning about when and how to adapt their weights.