NVIDIA researchers introduced PivotOPD, an on-policy distillation method that trains multi-turn LLM agents to avoid early pivotal mistakes and recover from them, posting the best average against 13 ...