Nvidia researchers say most AI agent failures trace back to a single bad move made early in the task, and they built a training method aimed at catching it.
The team tested three Qwen3 models, ranging from 8 billion to 235 billion parameters, on multi-turn agent tasks and found that more than half of failed runs contained what they call a pivotal mistake: an action that sends the agent off course, usually within the first few turns. Crucially, these errors were often recoverable if the model got a few turns of correct guidance right after the slip. Based on that finding, the researchers built PivotOPD, an on-policy distillation framework that trains a student model to do two things at once: avoid the pivotal mistake, and recover from it if it happens anyway. A teacher model supplies the correct action at the pivotal moment, then names recovery actions for the turns that follow.
Most agent training treats every step as equally important and tries to get each one right. PivotOPD's bet is that teaching recovery is the more efficient fix, since one early error otherwise cascades into total failure. The results suggest the bet pays off: a 1.7 billion parameter student improved by 5.5 percentage points on ALFWorld over the best of 13 baselines, and the gains carried over to a different model family, Nemotron-3.5, on the SWE-Bench Verified coding benchmark.
The catch is that PivotOPD leans on a bigger teacher model standing by to spot the mistake and point the way out, a setup that is easy to arrange in a research lab and harder to guarantee in an agent working alone in production.