An AI agent makes a mistake early in a task, then keeps going in the wrong direction. Our researchers built PivotOPD to teach agents how to avoid those mistakes and recover when they happen. During training, a teacher model shows the agent a better action and how to get... back on track over the next few steps. Read the paper and watch how it works:show more

👀

Exactly what i say since the early beginning.. We must reset the ethics from the Machine Learning a priori.. If not so.. a posteriori..teach the ethics to the LLM which did not get them! Nice understanding @NVIDIAAI Special thanks to @JensenHuang for his reactivity.

我就盼着它走歪以后 能早点自己拐回来

recovering from mistakes is the biggest thing holding agents back right now

Agents don't fail once. They fail early and keep walking, sure of themselves, straight into the wall. Teaching one to turn around is the part worth paying for.

The recovery piece is especially important. Most agents don’t fail from one bad action, but from failing to notice they’ve drifted.

would love a “steps wasted after the first mistake” metric alongside success rate. recovery can still take the scenic route.

Love that it learns to put the tomato down and go find the egg. Keeping only clean runs would throw away the bit where it actually recovers.

Demos are rehearsed; failures are not. NVIDIA says PivotOPD teaches agents to recover — production is the real test.

teaching it to notice the wrong turn early would save so many wasted steps in long tasks

the runtime half: one wrong step doesn't stay one step. everything built on top of it inherits the error, so recovery anchors to the last state that was still green, not to the moment it went wrong.

The next few recovery steps are a nice detail: correcting one action doesn't necessarily get the whole task back on track.

The tricky part is defining the reward signal so the teacher model catches compounding errors before the context window fills up.

Really nice work, so useful!

Showing a better action plus the next few steps back on track is a neat training signal. Is the evaluation on long-horizon tasks like coding or web agents, and how much did success rate improve over standard on-policy distillation?

One wrong step compounds fast, so a better move and the route back matter more than catching the drift. The hard bit is 'back on track' in tasks like a browser with no clear signal until the end. Does PivotOPD learn recovery targets from outcomes, or does it need them labelled?

What happens when the teacher model, the student agent and the overseer, the totality of human knowledge available are all wrong, and none the wiser, for example from human bias?

Agent 出错最怕的不是错一步，而是错一步之后还自信满满地一路错到底。能“掉头”的 agent 比“不犯错”的 agent 更现实——真实任务里哪有不犯错的。好奇在长链任务上，掉头成功率到底能提到多少。

Recovery is the bit most agent training ignores. An early wrong step compounds because every action after it looks reasonable in context. Teaching it to spot the drift and back out is more useful than chasing a perfect first move, since real tasks rarley allow one.