代理評測
NVIDIA AVO Completes All 183 Levels in the ARC-AGI-3 Public Set With About 12% Fewer Actions Than VISTA
NVIDIA adapted the AVO agent architecture, originally used for evolutionary search over GPU kernels, to interactive reasoning and achieved a 100.00 RHAE score with Claude Opus 5. The results show that agent memory and verification loops can substantially change the performance of the same model, but they cover only public environments, and the comparison with VISTA is not a controlled ablation.
NVIDIA says its Agentic Variation Operators (AVO) system completed all 183 levels across ARC-AGI-3’s 25 public environments, achieving a 100.00 RHAE score with 6,624 environment actions. ARC-AGI-3 provides no game rules, operating instructions, or explicit goals. Agents must probe the environment, observe state changes, and carry their discoveries forward to subsequent levels. RHAE accounts for both completion and action efficiency relative to a human baseline, so “100%” does not refer solely to the final completion rate.
AVO was originally designed to automate the evolution of GPU kernels. Its research implementation does not restrict the language model to generating one-off candidate programs. Instead, the agent can consult the lineage of existing versions, domain knowledge, and execution feedback as it repeatedly proposes, repairs, critiques, and validates modifications. NVIDIA has now applied the same long-horizon workflow to unfamiliar interactive environments, using Claude Opus 5 as the backend. The company used VISTA, which also runs on Opus 5, as a reference: VISTA completed the public set in 7,542 actions, while AVO used about 12% fewer. This gap supports an important engineering conclusion: leaderboard results measure a composite system comprising the model, observation representation, memory, context management, and control loop—not the standalone capability of the foundation model.
However, this is not a controlled experiment sufficient to declare ARC-AGI-3 solved. NVIDIA explicitly notes that AVO and VISTA differ in their agent backends, memory, and representation methods, making it impossible to attribute the action savings to any single module. The evaluation also covers only the 25 publicly available environments, so prolonged tuning of the agent architecture may absorb characteristics of the public set. Neither the original AVO paper nor this blog post provides a public execution package that can fully reproduce the 6,624-action result. Researchers should next focus on performance in hidden environments, ablations with fixed models and budgets, total token usage and wall-clock cost, and whether memory genuinely transfers across environments rather than merely preventing repeated exploration within the same game.