BDH-CQ Model Uses Silent Latent Reasoning to Slash ARC-AGI Inference Costs

Researchers have developed BDH-CQ, a model that performs visual reasoning without generating intermediate language tokens, instead absorbing patterns silently into its recurrent hidden state. The approach addresses a longstanding tradeoff in AI benchmarks where cheaper inference typically meant lower accuracy and vice versa. Tested on the ARC-AGI-1 benchmark — a visual abstract reasoning challenge — a 150-parameter variant achieved 29.5% pass@2 accuracy at a cost of just $0.0007 per task. Unlike standard language models that process examples as text tokens, BDH-CQ updates its internal memory state directly from demonstrations, bypassing linguistic bottlenecks. The result establishes a new cost-efficiency frontier, suggesting that making AI reasoning invisible to the user does not necessarily compromise its effectiveness.