Nvidia’s AVO scores 100 on ARC-AGI-3, clearing all 183 levels across 25 game environments
Nvidia said its general-purpose coding agent AVO achieved a perfect score on ARC-AGI-3, finishing all 183 levels across 25 game environments in 6,624 steps with an RHAE of 100.00. The result did not come from changing the base model. Instead, Nvidia wrapped Claude Opus 5 with a system layer that adds persistent memory and a supervisor, lifting performance from 30.16% for the standalone model to a full score in the benchmark setup described in the report. According to the source material, ARC-AGI-3 places agents in unfamiliar games without giving them explicit rules or goals. Some levels allow movement and rotate the entire scene when a blue-black block is touched, while others allow only clicking to cycle cell colors into a target pattern. Nvidia also noted that AVO worked entirely in text mode, receiving each frame as an exact 64×64 text grid rather than images or image tokens. The same architecture was originally built for GPU kernel optimization. A paper uploaded to arXiv on March 25, 2026, described AVO as an agentic mutation operator for autonomous evolutionary search. In tests on Nvidia’s B200, the system ran autonomously for seven days, explored more than 500 optimization directions, and produced 40 valid kernel versions. The report said its multi-head attention kernel was up to 3.5% faster than cuDNN and up to 10.5% faster than FlashAttention-4, with separate GQA results showing gains of 7.0% over cuDNN and 9.3% over FlashAttention-4 after about 30 minutes of autonomous work.
