Schema Harness Achieves ~99% on Arc‑AGI‑3 Public

4 hours ago

Schema harness enables frontier models to achieve ~99% performance on ARC-AGI-3 Public by treating latent world representations as executable programs, allowing interpretable, verifiable, and searchable modeling.
ARC-AGI-3 challenges agents to infer game rules, objects, and goals from raw 64x64 color grids without explicit instructions, using the Relative Human Action Efficiency (RHAE) metric for evaluation.
The approach combines state grounding (identifying objects and variables from observations) and mechanism discovery (learning transition rules), solved jointly within an editable program that mirrors physicist-like reasoning.
Controlled comparisons show Schema improves Claude Code baseline by 56.15%, with efficiency gains from verifying models against complete history and planning inside reusable simulators, reducing action counts significantly.
Case studies highlight Fable models discovering world models more efficiently than Opus by better experimental decisions, representational revisions, and earlier identification of key mechanisms.
The harness reduces theory usage costs through persistence and searchability, while the underlying model affects discovery efficiency; search completeness depends on the accuracy of the world model.
Self-reported results on the Public set (98.98% with Claude, 95.35% with Sol) indicate progress but lack independent verification; extrapolation to Semi-private performance remains uncertain.

Hasty Briefsbeta