Small Transformer achieves 44% on ARC-AGI-1 via test-time training
September 1, 2026
A small transformer trained from scratch in 1.5 hours on a single 5090 reaches 44% on ARC-AGI-1 and 7% on ARC-2. The approach uses test-time training to match performance of TRM/HRM models while focusing on extreme sample efficiency within the 1,000-puzzle ARC dataset.
HOW THIS AFFECTS YOU
●
builderYou can achieve competitive ARC benchmarks using minimal compute and rapid iteration cycles.
●
researcherThis provides a baseline for testing the limits of sample efficiency within standard transformer architectures.