Artin Tajdini

LLM agents trained on function calling forget nearly 25% of their capability when action names are perturbed, but a simple self-supervised regularization recovers over half of that loss.