BDH-CQ Achieves 29.5% ARC-AGI-1 Pass@2 via Recurrent Latent Reasoning
August 9, 2026
BDH-CQ uses recurrent latent reasoning to update memory during inference without verbalizing intermediate steps. A 150M-parameter configuration achieved 29.5% pass@2 on ARC-AGI-1 at a cost of $0.0007 per task, improving the cost-accuracy Pareto frontier.
HOW THIS AFFECTS YOU
●
builderYou can achieve high-reasoning performance on ARC-like tasks with extremely low-parameter, low-cost models.
●
researcherThis method demonstrates how latent space computation can bypass the need for explicit chain-of-thought verbalization.