When LLMs Modify Themselves: Convergence, Values, and Stopping in Iterative Self-Improvement
An empirical study of 54 self-modification experiments across six language models found that whether and when a model stops rewriting its own system prompt is tied to model identity rather than the optimization mode assi…