NVIDIA published results for AVO (Agentic Variation Operators), a general-purpose long-horizon agent architecture, on two very different tasks: autonomous GPU-kernel optimization, and the ARC-AGI-3 interactive-reasoning benchmark. The headline number is a perfect 100.00 RHAE across all 25 public-set environments, 183 levels, running Claude Opus 5. That's a real result, and also — worth saying plainly before anything else — the third harness in two months to hit that same ceiling on this exact benchmark, by a third, incompatible design.
Three harnesses, three designs, one saturated benchmark
VISTA gets there with a lossless pixel memory the model can re-inspect at frame, region, and pixel granularity. Schema gets there by writing each game's mechanics as an executable, backtestable program and planning inside it at zero action cost. AVO gets there differently again: a persistent five-stage loop — inspect context, plan, implement, evaluate, diagnose and repair — with an external supervisor watching for stalled progress and intervening when the main loop plateaus, operating on text-only 64×64 grid observations, no images at all. NVIDIA describes its own approach as adopting VISTA's "direct-interaction" philosophy (no explicit world-model compiled up front, unlike Schema or the Tycho system NVIDIA also cites) while reimplementing the task interface independently and swapping VISTA's pixel observations for plain text grids.
Three architectures that disagree with each other about where an agent's understanding of the world should live — pixels, a compiled program, or persistent structured memory — all converge on the same perfect score. The conclusion this blog drew when VISTA and Schema first did this applies again, now with a third confirming data point: ARC-AGI-3's public set has stopped discriminating between frontier harness designs. None of these results touch a model weight, and disagreement this total between architectures that all land at 100% is itself the finding, not a footnote to it.
The comparison NVIDIA offers, and the caveat NVIDIA itself attaches
NVIDIA's own post makes a specific, checkable claim: AVO solved all 183 public-set levels in 6,624 environment actions, against VISTA's reported 7,542 for the same levels — roughly 12% fewer. Immediately after stating that number, NVIDIA adds its own caveat: "this should not be interpreted as a controlled ablation" — the two systems differ in agent backend, observation format, memory design, and context management simultaneously, so the action-count gap can't be attributed to any single architectural choice. That's a real methodological restraint worth crediting on its own — a lab citing a favorable comparison and then explicitly declining to overclaim what it proves is not the industry default this month.
The other number worth sitting with: ARC Prize's own verified figure for Claude Opus 5 at High reasoning effort, no harness, is approximately 30% on the public set. AVO took the same base model to 100%. That's NVIDIA's actual point, stated directly in its own writeup — "the model matters, but the model is not the entire agent" — and it's a real one, even if it's the same point Prime Agent, VISTA, and Schema have each already made with their own harnesses this year.
Public set, self-reported — same caveat as every harness this year
AVO's 100.00 RHAE is on the 25-environment public set, explicitly not the semi-private or private competition sets ARC Prize itself administers and verifies. That's the identical caveat that applies to VISTA's and Schema's public-set numbers, and NVIDIA states it plainly rather than letting the 100% figure imply verified competition-grade performance. Worth remembering for the same reason it's mattered every time this month: the public set is the one a harness builder can iterate against directly, which is exactly the setting where a harness — rather than the model — is doing the work.
What ARC-AGI's creator says about all of this
François Chollet — the researcher who created ARC-AGI itself — responded directly to NVIDIA's announcement, and his reaction sharpens one point this post already makes while complicating another. (Relayed from a screenshot provided directly, not an independently fetched link — worth flagging since this environment can't reach x.com to verify the post is unedited.)
On the public-set caveat, Chollet is blunter than NVIDIA or this post: "scoring 100% on the public demonstration set is not the same as 'scoring 100% on the ARC-AGI-3 benchmark.' It would be like saying you beat a videogame because you cleared the tutorial level." That's the identical point made above, but coming from the person who actually built the benchmark and administers the semi-private and private sets none of these harnesses have touched — a stronger authority for the same caveat, not a new one.
On the architecture, though, his characterization sits less comfortably against NVIDIA's own account. Chollet: "Like all high-performing approaches on ARC-AGI-3, it uses deep learning-guided on-the-fly synthesis of symbolic world models, i.e. navigating the world by generating programs to represent what you know." NVIDIA's own writeup describes AVO as adopting VISTA's "direct-interaction" philosophy specifically because it doesn't compile an explicit world model up front, unlike Schema or Tycho, and runs on plain text-grid observations rather than any symbolic program representation. Whether that's a vocabulary gap — a five-stage inspect/plan/implement/evaluate/diagnose loop could plausibly function as on-the-fly world-model synthesis even without NVIDIA describing it in those terms — or a real disagreement about what AVO is doing under the hood isn't something this post can settle from the outside. Worth flagging as an open tension rather than resolving it in either direction.
The GPU-kernel result is the more interesting news, and gets less attention
Buried beneath the ARC-AGI-3 headline is a result with a more direct, practical payoff: running continuously for seven days, exploring 500+ optimization directions, and committing 40 kernel versions, AVO evolved multihead attention kernels that outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on NVIDIA DGX B200 hardware — then adapted the result to grouped-query attention in roughly 30 additional minutes of autonomous work. Beating cuDNN and FlashAttention-4, both mature, heavily-optimized production libraries, on real hardware is a concrete engineering result with an immediate use case, in a way a saturated reasoning benchmark no longer is. NVIDIA's own framing puts the transfer between these two very different tasks — kernel optimization and interactive game reasoning — ahead of either individual score: the same core loop (hypothesize, act, observe, preserve state, revise, recover, continue) generalized across domains without redesign, which is the more durable claim than any single benchmark number.
What to expect next
- Watch for the semi-private and private ARC-AGI-3 set results. Every harness this year, AVO included, has published only public-set numbers; ARC Prize's own verified scores on the harder sets are the actual test of whether any of these designs generalizes past a benchmark their builders can practice against directly.
- Watch for independent reproduction of the kernel-optimization result. A 3.5%/10.5% improvement over cuDNN and FlashAttention-4 is specific and falsifiable in a way a saturated benchmark score no longer is — third-party benchmarking on the same hardware would settle it.
- Watch whether AVO's cross-model results (GPT-5.6 Sol alongside Claude Opus 5) get a full systematic comparison. NVIDIA's own framing calls the current Sol results "preliminary" and "limited" — a real comparison, not a subset, is the next useful data point.
References: NVIDIA — NVIDIA AVO Reaches 100% on ARC-AGI-3 · related coverage: Two Harnesses Solved the Same Benchmark by Disagreeing About Memory · Prime Agent Crossed the Human Line by 0.1 Points · ARC-AGI, Explained · Frontier Arcade: trends & predictions