14
BDH-CQ: In-Context Learning with Recurrent Latent Reasoning
arxiv.orgWe introduce BDH-CQ, a reasoning model that combines in-context learning with recurrent latent reasoning. 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. We evaluate the model on the public ARC-AGI-1 evaluation set and use controlled ARC-like interventions to study what it learns from demonstrations, how consistently it applies an inferred transformation, and which concepts remain difficult. A 150M-parameter configuration reaches 29.5% pass@2 at a computed inference cost of \$0.0007 per task. This operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.
29% is probably somewhere between meh and ok-ish.
150M is extremely small. The B stands for billion parameters, so the M stands for million parameters. So we’re talking 3% the size of recent mid-tier models.
What this means is someone made a tiny model, and it scored way better than we thought. The conclusion is (I assume) we can make models way smaller at the same quality.
That’s huge, because that would mean in a matter of years anyone can deploy their own model on the most dog shit hardware. This would kill AI companies, it would neuter Nvidia and it would mean PC components will decrease to suitable prices. And it also means even in poor countries people have access to AI which I personally think is one of the most noble things to finally make people more equal by giving them lots of knowledge.
yes, models can be much smaller than people think. esp when they’re fit to purpose. I have a couple models I deploy in apps that perform better than the 27 gig models models and are only 20-100MB. its about training, data processing, and understanding how information theory plays into the domain space you’re solving for.