Teaching NVIDIA's Nemotron 3.5 Super-VL ML Engineering

The short version

The problem

A machine-learning agent may spend hours trying experiments, but it is hard to teach it which choices were useful and how to recover when tests fail. Training from final answers alone misses the evidence and decisions that led to them.

The idea

The paper turns research attempts into lessons: it trains on selected decisions from a stronger model, and on Nemotron’s own successful attempts after removing extra hints the model would not normally have. It also teaches the model to review what went wrong, while replaying concrete examples of actions—for instance, learning to fix a slow test before making more model changes.

What it got

On 22 extended machine-learning tasks, the best trained model scored 0.358, compared with 0.286 for the untrained base model. That is a 24.8% relative improvement under the same agent setup.