A team from MIT built a model that scores 61.9% on ARC-AGI-PUB using an 8B LLM plus Test-Time-Training (TTT). Previous record was 42%. (via Localllama)

→ View original post on X — @kimmonismus