BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

2026-08-10

AI Summary

This work addresses the challenge of efficiently solving abstract visual reasoning tasks, such as those in ARC-AGI, without explicitly generating intermediate reasoning steps. The authors propose an implicit iterative reasoning architecture that leverages in-context learning to drive recurrent neural memory, enabling iterative refinement of representations in a high-dimensional latent space to answer queries end-to-end—without requiring language-based intermediate rationales. Evaluated on ARC-AGI-1, their 150M-parameter model achieves a pass@2 rate of 29.5%, with a per-task inference cost of merely $0.0007. This result substantially advances the state-of-the-art by establishing a new efficiency frontier, significantly outperforming prior methods on the cost–accuracy Pareto front.

Abstract

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.

Problem

• in-context learning
• recurrent latent reasoning
• abstract reasoning
• cost-efficiency
• ARC-AGI

Innovation

• in-context learning
• recurrent latent reasoning
• latent space reasoning
• cost-efficient AI
• ARC-AGI