NVIDIA's AVO Agent Scores 100% on ARC-AGI-3: The Harness, Not the Model, Does It

新智元 · wechat · 2026-08-22

NVIDIA's general-purpose coding agent AVO achieved a perfect score on ARC-AGI-3: all 183 levels across 25 game environments in 6,624 steps, RHAE 100.00. ARC-AGI-3 gives no rules or goals. The underlying model is Claude Opus 5—which scores just 30.16% alone (still #1 on the official leaderboard). With the AVO harness, it jumps to 100, using only a 64×64 text grid, zero image tokens.

Two mechanisms matter: persistent memory—storing past versions, eval results, profiler outputs, and accumulated reasoning so the agent resumes from current state after context resets; and a supervisor—a watcher that detects stalls or loops and redirects the main agent to different strategies.

AVO was built for GPU kernel optimization: it replaces the mutation operator with an autonomous agent that reads CUDA/PTX docs, runs tests, reads profilers, and self-diagnoses. After 7 autonomous days on a B200 exploring 500+ directions, its multi-head attention kernel beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%; a GQA kernel took 30 minutes to beat both by 7.0%/9.3%. The team is largely Chinese: co-first author Bing Xu (MXNet author, GAN paper co-author), Tianqi Chen, Zihao Ye, and more.

This is the third ARC-AGI-3 perfect score in six weeks, all powered by Claude Opus 5. NVIDIA's play: open-source the model (Nemotron4), own the harness layer, and let long-horizon agents—the ideal GPU workload—settle the bill.

Original post →

NVIDIA's AVO Agent Scores 100% on ARC-AGI-3: The Harness, Not the Model, Does It

More from coding & agent