BDH-CQ was submitted to arXiv on August 10, 2026 (paper 2608.09888) by Bjorn Engdahl, Adrian Kosowski, Jan Chorowski, Zuzanna Stamirowska, Przemyslaw Uznanski, Junlin Jiang, Rohan Phadke, Remigiusz Kinas and Richard Zhong. It describes a reasoning model that departs from the dominant recipe in two ways at once: it is very small, and it does not think in words.
The mechanism is recurrent latent reasoning. As the paper puts it, “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.” Where a chain-of-thought model spends test-time compute emitting tokens that are then fed back to itself, BDH-CQ iterates inside a latent state. Context is absorbed into recurrent memory rather than re-attended over a growing prompt, which is what makes the compute budget per task so small.
The reported result is a 150 million parameter version scoring 29.5 percent pass@2 on ARC-AGI-1 at a computed inference cost of 0.0007 dollars per task, which the authors claim breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier. The accuracy is nowhere near frontier systems, which have long since passed the high 80s on that benchmark. The claim is specifically about the frontier curve: what accuracy is reachable per dollar, not what accuracy is reachable at any price. Two things are unverified as of this writing - whether the cost accounting is comparable to how large-model ARC runs are priced, and whether the approach holds up on ARC-AGI-2, which was built precisely to defeat solutions that worked on the first version.
For a technical leader, the interest is not in this particular model but in what it probes. Nearly all current reasoning capability is bought with test-time token generation, which is why reasoning-mode API bills scale so badly. A result showing that a sub-billion-parameter recurrent model can occupy a corner of the cost-accuracy plane that no large model reaches is evidence that verbalized chain of thought is one implementation of test-time compute rather than the only one. If latent recurrence generalizes past puzzle benchmarks, the economics of deployed reasoning change; if it does not, this remains a well-executed data point on a curve.