Search papers, labs, and topics across Lattice.
The paper introduces BDH-CQ, a novel reasoning model that integrates in-context learning with recurrent latent reasoning, allowing it to continuously update its memory during inference. This model iteratively computes solutions in a high-dimensional latent space without verbalizing intermediate steps, significantly enhancing efficiency. Evaluated on the ARC-AGI-1 set, a 150M-parameter version achieved a pass rate of 29.5% at a remarkably low inference cost, setting a new benchmark for cost-effective reasoning models.
Breaking the cost-accuracy Pareto frontier, BDH-CQ achieves state-of-the-art efficiency in reasoning tasks with minimal computational expense.
We introduce BDH-CQ, a reasoning model that combines in-context learning with recurrent latent reasoning. Inputs presented at inference time continuously update the model's recurrent memory; the model then solves a query through iterative computation in a high-dimensional latent space, without verbalizing its intermediate reasoning. We evaluate the model on the public ARC-AGI-1 evaluation set and use controlled ARC-like interventions to study what it learns from demonstrations, how consistently it applies an inferred transformation, and which concepts remain difficult. A 150M-parameter configuration reaches 29.5% pass@2 at a computed inference cost of \$0.0007 per task. This operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.
