Hugging Face researchers have introduced BDH-CQ, a novel reasoning model that merges in-context learning with recurrent latent reasoning. Unlike traditional models that verbalize intermediate steps, BDH-CQ performs iterative computation in a high-dimensional latent space, continuously updating its recurrent memory with inputs presented at inference time.
The model was evaluated on the public ARC-AGI-1 benchmark, a challenging test of abstract reasoning. A 150M-parameter configuration achieved 29.5% pass@2 at a computed inference cost of $0.0007 per task. This operating point breaks through the previously reported cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.
The researchers also used controlled ARC-like interventions to analyze what the model learns from demonstrations, how consistently it applies inferred transformations, and which concepts remain difficult. This provides insights into the model's reasoning capabilities and limitations.
BDH-CQ's success suggests that recurrent latent reasoning can be a cost-effective alternative to larger, more verbose models, potentially enabling advanced reasoning on edge devices with limited computational resources.