Daily AI CatchupSchema Hits 99% on ARC-AGI-3 by Forcing Models to Reason Like Physicists
Schema is a harness that gets frontier models to solve ARC-AGI-3 puzzles by writing each game's hidden mechanics as an executable program, testing it against observed interactions, and planning inside the verified model — formalizing first, then acting, rather than pattern-matching. The result: 99% on the ARC-AGI-3 Public benchmark, in a setting where models receive no explanation of what they're seeing.
Why it matters
💻 Developer · The "formalize first, then act" harness pattern is worth studying and potentially adapting for your own agent design, especially for tasks involving unfamiliar or under-specified environments.
📦 Product · A harness-level technique reaching 99% on a benchmark where raw models scored near 0% is a reminder that scaffolding/architecture can matter as much as the underlying model for hard reasoning tasks.
🎨 Design · No direct design impact — this is an agent-architecture research result.
📈 Business · This is a useful counter-narrative to "just wait for a bigger model" — clever harness design can unlock capability gains that pure scale hasn't delivered yet.
🤔 Just Curious · Researchers found a clever trick to make AI much better at a notoriously hard test: instead of just guessing at patterns, they made the AI first write out the actual rules of what it's seeing as code, test those rules, and then act — like a scientist forming and checking a hypothesis before doing anything.
Sources: Schema: Harness Handbook