← all projects
ARC-AGI-3 Research System
- Timeline
- June 2026 – present
- Role
- Solo researcher & builder
- Stack
- Python, world models, search, open-weight LLMs
- Links
- GitHub
Training-free system for ARC-AGI-3, where frontier models score under 0.5%. LLM-proposed world-model rules verified by exact replay, evidence-seeking exploration, forward-search planning. Zero LLM calls at inference. Along the way: reverse-engineered undocumented evaluation semantics and refuted two claims from a published arXiv paper.