Experimental AI Model Solves Reasoning Puzzles Without Writing Out Its Steps

A small experimental AI system called BDH-CQ can work through some reasoning puzzles without spelling out its intermediate thinking, researchers report in a paper submitted August 10 to arXiv.org. The work questions whether AI needs language at every step of reasoning, and whether doing more internally could make some reasoning cheaper.

Many of today's AI systems generate lengthy chains of words before answering, which takes computing power and makes answers slower and more expensive. BDH-CQ takes a different route: it uses each training example to update a fixed-size memory, then works through a problem internally. "Nothing in between ever converts into language," says complexity scientist Zuzanna Stamirowska, CEO of AI company Pathway.

The model solved nearly three in 10 puzzles on the public ARC-AGI-1 evaluation set when given two attempts. It handled tasks involving turning or moving shapes but struggled more with color changes and some rule combinations. Researchers estimate each puzzle query costs about $0.00070 to run, roughly one-eleventh as much as GPT-5.6 Luna on the same test, though the costs were calculated differently. The study has not been peer-reviewed.

Computer scientist Yuntian Deng of the University of Waterloo calls the finding "an interesting efficiency result" but says it does not show the architecture is better than other approaches. Machine learning researcher Jonas Geiping notes the model was built specifically for ARC-style problems, making direct comparisons with general-purpose systems difficult, though he called the approach "neat." Deng adds that without written reasoning the model is harder to inspect.

Source: Science News