Paper summary
Paper summary: BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
BDH-CQ combines in-context learning with recurrent latent reasoning in the Dragon Hatchling architecture; a 150M model reaches 29.5% pass@2 on ARC-AGI-1 at about $0.0007 per task, a new cost–accuracy frontier.
- Demonstrations update the model's recurrent memory, then it reasons in latent space without writing out steps.
- The claim is about efficiency: decent ARC-AGI accuracy at a tiny fraction of a cent per task.