BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
A 150M-parameter BDH-CQ model reports 29.5% pass@2 on ARC-AGI-1 at an estimated $0.0007 per task.
The paper introduces a reasoning model that updates recurrent memory from examples shown at inference time. It then works through the query in latent space, without spelling out intermediate reasoning. The authors test it on ARC-AGI-1 and with controlled ARC-like interventions to probe what the model learns from demonstrations and where it remains weak. They say this result sets a new cost-efficiency state of the art on the benchmark. ArXiv · AI/CL/LG's note
The paper introduces a reasoning model that updates recurrent memory from examples shown at inference time. It then works through the query in latent space, without spelling out intermediate reasoning. The authors test it on ARC-AGI-1 and with controlled ARC-like interventions to probe what the model learns from demonstrations and where it remains weak. They say this result sets a new cost-efficiency state of the art on the benchmark. ArXiv · AI/CL/LG's note
score 6