Machine Learning

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

Björn Engdahl, Adrian Kosowski, Jan Chorowski, Zuzanna Stamirowska, Przemysław Uznański, Junlin Jiang, Rohan Phadke, Remigiusz Kinas, Richard Zhong

Featured August 12, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new AI system learns new visual puzzles just by seeing a few examples, remembering them in a special internal "thought space," and then figuring out the answer by thinking deeply without saying its steps out loud, making it super cheap and fast.

In depth

The paper introduces BDH-CQ, a reasoning model that combines in-context learning with recurrent latent reasoning. It continuously updates a recurrent memory with inference-time demonstrations and then iteratively solves queries in a high-dimensional latent space without verbalizing intermediate steps. This approach achieves a new state of the art in cost efficiency on ARC-AGI-1 by avoiding the token consumption and latency associated with chain-of-thought prompting.

Key Takeaways

1
BDH-CQ integrates in-context learning with recurrent latent reasoning, allowing models to acquire new skills from demonstrations at inference time without explicit intermediate verbalization.
2
The model establishes a new cost-accuracy Pareto frontier on ARC-AGI-1, reaching 29.5% pass@2 at a significantly lower cost ($0.00070 per task) compared to previous state-of-the-art systems.
3
The system's latent reasoning enables parallel exploration of solution paths in a continuous hidden state, offering an efficient alternative to autoregressive token generation for complex reasoning tasks.

Conceptual Flow

HIGH LEVEL

Methodology: How BDH-CQ Works

The system learns from examples by updating its internal memory, then thinks through the problem in a hidden space to find the answer.

Example Puzzles

New Puzzle

Learn & Think

Puzzle Answer

Results: Impact on Puzzle Solving

This new way of thinking solves puzzles much cheaper and almost as accurately as bigger, more expensive systems.

Old Way (Expensive)

New Way (Cheap)

Compare Performance

Better Value

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.