NVIDIA says its Agentic Variation Operators system, or AVO, reached a 100.00 RHAE score on the ARC-AGI-3 public set, completing all 183 levels across 25 environments.
The result is notable because NVIDIA presents AVO as a general-purpose agent architecture, not an ARC-specific world model. The system uses persistent memory, tool use, execution feedback, and a supervisor that can redirect the main agent when progress stalls.
NVIDIA says AVO completed the public set in 6,624 environment actions. It compares that with VISTA’s reported 7,542 actions using Claude Opus 5 across the same public-set levels, or about 12% fewer actions. NVIDIA also warns that this is not a controlled ablation because the systems differ in backend, observation representation, memory, context management, and other implementation details.
That caveat is the story. The result is less about declaring one base model better than another. It shows how much benchmark performance can come from the agent system wrapped around the model.
The harness is part of the capability
AVO was first demonstrated on software engineering and GPU-kernel optimization tasks. In that setting, an agent must inspect code, form hypotheses, make changes, run tests, interpret performance output, and revise. NVIDIA says AVO operated for seven days, explored more than 500 optimization directions, produced 40 committed kernel versions, and achieved up to 10.5% better performance than FlashAttention-4 across evaluated configurations on DGX B200 systems.
ARC-AGI-3 is different on the surface. Agents interact with unfamiliar environments without stated rules or goals. But the repeated loop is similar: observe, act, learn from feedback, preserve useful state, and recover from incorrect assumptions.
That is why AVO’s memory and supervisor design matters. A long-horizon agent can fail not because the model lacks a concept, but because it forgets useful evidence, repeats bad experiments, loses track of prior state, or spends too many actions rediscovering the same rule.
NVIDIA’s result suggests agent evaluation needs to name the harness, memory design, observation format, tool interface, and supervisor behavior. A score without those details hides what is actually being tested.
ARC scores are not product readiness
The 100.00 RHAE score is impressive, but it should not be treated as proof that an agent is ready for arbitrary real-world autonomy. ARC-AGI-3 is a benchmark with a defined interaction loop. Production agents face permissions, security boundaries, unreliable tools, messy data, user interruptions, and ambiguous success criteria.
The better takeaway is methodological. If agent systems are becoming products, then evaluation has to move from model-only leaderboards toward complete system comparisons. The base model matters, but so do memory, task decomposition, rollback, sandboxing, execution traces, and human oversight.
For buyers, that means asking vendors to show full agent traces and failure recovery, not only final scores. For researchers, it means benchmark reports should separate what belongs to the model from what belongs to the scaffold around it.
