Schema Nears Perfect Score on ARC-AGI-3
机器之心 · wechat · 2026-07-18
Nearing a Perfect Score on ARC-AGI-3 Using a Harness
The schema mentioned by Berkeley postdoc HavenFeng is a harness wrapped around LLMs: it doesn't alter model weights, only the observation, hypothesis, experiment, correction rules, and execution flow at the outer layer.
On the ARC-AGI-3 Public set, it achieved extremely high scores when combined with different models:
- Claude Opus 4.8 + Fable 5: 98.98% RHAE
- GPT-5.6 Sol combo: 95.35% RHAE
The article emphasizes that the key to this improvement isn't "bigger models," but turning reasoning into a verifiable, replayable, and searchable programmatic flow.
How the Schema Works
ARC-AGI-3 is a grid-game benchmark with hidden rules. The schema's approach is:
- State Grounding: Organize pixels, objects, and relationships into trackable states.
- Mechanism Discovery: Write how states change after operations into runnable programs.
- Outer Loop: Observe → deduce → execute → record.
- Counterexample-Driven Correction: If reality conflicts with predictions, halt execution and return to the deduction phase to fix the model.
It also proactively conducts "hypothesis-distinguishing" experiments, finding operations that best validate/falsify candidate rules to reduce useless interactions.
Author and Community Controversies
ARCPrize President Greg Kamradt praised the idea of "writing world models as programs" but raised two critiques:
- The score utilizes fixed fallback rules, potentially injecting information about "which rounds performed poorly" into the process.
- Concepts like "player," "wall," and "counter" in the docs might stem from manual prompting, meaning the results aren't entirely model-driven.
Furthermore, the team's results are currently self-reported on the public set and have not yet undergone ARCPrize's independent review.
Core Message
The author believes the significance of this method isn't just topping leaderboards, but turning "mechanism discovery" into a general capability: letting models explicitly write their understanding of the world as programs, and continuously validating them against reality through action.
Related event: Schema Harness Sparks ARC-AGI-3 Debate(14 posts)→
More from coding & agent
- Building a Secure AI Agent Gateway: Self-Hosting OAuth for Multiple SaaS Apps — Defiant_Cod_2654 · 2026-07-22
- Rowboat launches as an open-source, local-first AI coworker with memory — ycombinator · 2026-07-22
- Scoble says AI “loops” really means long-running multi-agent workspaces — Scobleizer · 2026-07-22
- Kimi Code opens a waitlist as Moonshot rolls out its coding product — Fabulous_Bonus_8981 · 2026-07-22
- Open-source runtime lets each repo define its own AI code reviewer — ibabufrik · 2026-07-22
- Indie Dev Asks: What's Actually Broken in Your AI Agent's Memory Today? — AcceptableTime7937 · 2026-07-22