NVIDIA paper details VERA framework for co-evolving agent harnesses and models

VERA converts benchmark trajectories into more than 9,000 restartable sandboxes with rubric scoring, retaining only environments that execute and can be evaluated from observable evidence. During training, the framework co-evolves the agent harness and model weights simultaneously. A harness modification is kept only if it passes self-tests and yields at least a 5-point gain on the development set, while model checkpoints are rejected if performance drops by more than 20%.

According to the [research paper](https://arxiv.org/abs/2610.05923), a 9B co-evolved agent outperformed the strongest single-axis baseline by 10.3 points on AutoCoWorkBench and 13.0 points on AutoMedBench. At 27B, the agent reached a score of 71.6 on AutoCoWorkBench, surpassing Claude Opus 4.8. NVIDIA has open-sourced the environment corpus.