NVIDIA PivotOPD Teaches Multi-Turn AI Agents to Recover From Pivotal Mistakes

What happened

NVIDIA researchers, with Princeton University and the University of Maryland, have introduced PivotOPD , an on-policy distillation method for multi-turn LLM (the kind of AI system trained on text to produce text) agents. Tested on Qwen3-1.7B and Qwen3-8B students, plus a Nemotron-3.5-SFT student on SWE-Bench Verified. NVIDIA is a semiconductor company based in Santa Clara.

Adds 0 inference (running a trained model to get an answer, rather than training it) cost, so the trained agent runs wherever its base model runs. PivotOPD on-policy distillation trains an agent to avoid its most damaging early mistake, and to recover when it happens anyway.

The takeaway: recovery is learnable, and standard OPD rarely teaches it. Runs on: Trained on NVIDIA H100 nodes.

Key facts

• NVIDIA researchers, with Princeton University and the University of Maryland, have — introduced: PivotOPD , an on-policy distillation method for multi-turn LLM agents