Why teaching AI agents to use tools keeps blowing up in training

A new paper pins the sudden collapse of multi-step tool-use training on runaway probabilities in a few control tokens, and shows that mixing in supervised examples stabilizes it.

Raccontata dadev.to