FULL STORY

NVIDIA's AVO Agent Takes On ARC-AGI-3

NVIDIA released its AVO coding agent, topping the ARC-AGI-3 public set. Follow-up analysis revealed its architecture of persistent memory, supervision and tool use, enabling Claude to score full marks on the benchmark.

2026-08-21 ~ 2026-08-24 · 2 episodes · 24 posts

Episode 1 · NVIDIA's AVO Agent Hits 100% on ARC-AGI-3 Public Set (2026-08-21, 22 posts)

On August 21, NVIDIA released AVO (Agentic Variation Operators), a general-purpose coding agent that scored 100% on the public set of the ARC-AGI-3 interactive reasoning benchmark, completing all 183 levels across 25 environments (RHAE 100.00). The benchmark provides no instructions, explicit rules, or goal hints; AVO had to infer objectives on its own. With no open-source release or disclosed cost, and with the same underlying model performing vastly differently under different harnesses, the community debated the result's significance.

Confirmed

  • Multiple posts including @MagicZhang, @theologi, @NVIDIAAI, and @wavefnx confirm the 100% score: 183 levels across 25 public environments; @MagicZhang shared the official scoreboard screenshot
  • Per @danielmac8, AVO's core idea replaces fixed mutation/crossover operators in evolutionary search with autonomous coding agents that consult lineage, knowledge bases, and execution feedback; the architecture integrates persistent memory, supervisor-based feedback (@HaktanSuren quipped that this AI itself needs a "middle manager"), and tool use, with Claude models integrated into the system design
  • Per @eigenhector, AVO was not built specifically for the benchmark and solved tasks with minimal inputs and tools by reviewing past grids represented as text
  • Per @huggingface, AVO originated as NVIDIA's in-house coding tool for optimizing CUDA GPU kernels; Clement Delangue commented that with agents, we move from models that barely run to systems that can actually perform complex tasks
  • @danielmac8's evaluation notes the AVO harness with Claude Opus 5 lifted Opus 5's roughly 30% score under a standard harness to 100%; @imjustnewatai relayed the same comparison (183/183 as 100% RHAE)
  • @mhmazur noted community skepticism about the claim NVIDIA "solved" ARC-AGI-3: the score is mainly attributable to the AVO agent framework (cross-context persistent memory and inspection) rather than the Opus 5 model itself
  • Per @jamestagg, benchmark author François Chollet stated that, like all strong methods on ARC-AGI-3, the 100% covers only the demonstration (public) set and does not mean the full benchmark is solved
  • @danielmac8 added that AVO has no open-source repository and no cost disclosure, but its concepts resemble Prime Intellect's open-source Prime Agent (17.6k stars), a self-improving coding/research agent usable as a reproducible alternative

Unconfirmed

  • Posts differ on the expansion of AVO's acronym (@danielmac8 wrote "Agentic Voxel Objects" in one post); NVIDIA's official "Agentic Variation Operators" is taken as authoritative

Why it matters

  • @NVIDIAAI argues the result shows system design (agent architecture, memory, tool integration) may matter more than raw model capability, pointing a new direction for general-agent research
  • The jump from 30% to 100% for the same underlying model shows how much the evaluation harness affects measured LLM capability; yet, as Chollet cautioned, the perfect score covers only the public demonstration set and should not be read as solving ARC-AGI-3
  • Achieving a full score without benchmark-specific tuning suggests notable autonomous goal inference and generalization, but the lack of open source and cost data limits independent reproduction; the open-source Prime Agent offers a verification path

2 more related posts →

Episode 2 · NVIDIA's AVO Architecture Achieves Full Score on ARC-AGI-3 (2026-08-24, 2 posts)

NVIDIA's AVO (Agentic Variation Operators) architecture wraps frontier models with persistent memory, oversight, and tool use, enabling long-horizon autonomous tasks. It helped Claude Opus 5 achieve a full score on ARC-AGI-3, showing system design can outweigh raw model capability.