
NVIDIA PivotOPD trains agents to recover from early mistakes
arXiv 2609.40285, posted 30 September 2026, introduces PivotOPD from NVIDIA, Princeton, and Maryland authors. They report best averages versus 13 baselines on ALFWorld, WebShop, and search QA, plus a +3.2% SWE-Bench Verified lift for a Nemotron-3.5 student. Code is listed as coming soon.




































