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

Problem

ARC-AGI tests learning a new transformation from a few examples. Strong scores have usually come from very large models at high cost per task.

Method

BDH-CQ builds on the Dragon Hatchling (BDH) architecture (summarised here in October 2025). Demonstrations presented at inference continuously update the model’s recurrent memory, and it then solves the query through iterative computation in a high-dimensional latent space, without verbalising its reasoning. Controlled ARC-like interventions probe what it learns from demonstrations and how consistently it applies an inferred rule.

Key results

From the abstract: a 150M-parameter model reaches 29.5% pass@2 on ARC-AGI-1 at about $0.0007 per task, which the authors say goes past the previously reported cost–accuracy Pareto frontier.

Limitations

There’s no dedicated limitations section in the extracted text. The absolute accuracy (29.5%) is well below the strongest systems. The claim is about cost efficiency.