BDH-CQ is a reasoning model combining in-context learning with recurrent latent reasoning. A 150M-parameter configuration achieves 29.5% pass@2 on the ARC-AGI-1 evaluation set at an inference cost of $0.0007 per task, establishing a new state of the art in benchmark cost efficiency.
Researchers introduced BDH-CQ, a model that performs reasoning through iterative computation in a high-dimensional latent space without verbalizing intermediate steps. Inputs at inference time update a recurrent memory. Evaluated on the public ARC-AGI-1 set, the 150M-parameter model reaches 29.5% pass@2 at $0.0007 per task, breaking the previously reported cost-accuracy Pareto frontier.
The model uses recurrent latent reasoning, where inference-time inputs continuously update a recurrent memory, and queries are solved via iterative latent-space computation. This avoids explicit chain-of-thought, potentially enabling more efficient reasoning. Controlled ARC-like interventions were used to analyze learning from demonstrations and consistency of inferred transformations.
Achieving a new cost-accuracy Pareto frontier on ARC-AGI-1 with a small 150M-parameter model suggests that efficient reasoning architectures can challenge larger models on abstract reasoning benchmarks. This could shift focus toward inference-time compute efficiency in model development.
The model's low inference cost ($0.0007 per task) and state-of-the-art efficiency make it attractive for applications requiring cost-effective reasoning, such as automated problem-solving, educational tools, or embedded AI systems.
Further scaling of the recurrent latent reasoning approach may improve performance on ARC-AGI-1 and similar benchmarks. The low inference cost could enable deployment in resource-constrained settings. Future work may explore integration with larger models or application to other reasoning tasks.