tokenization

Hugging Face outlines a safer way to train tool-using agents with RL

A token-handling mistake in multi-turn RL can silently corrupt gradients when models call tools mid-rollout.