Vector-Based AI Could Cut Reasoning Costs by 11 Times
A new vector-based AI model could make advanced reasoning far cheaper to run. Researchers at Pathway developed the BDH-CQ system using a different approach to machine cognition. The team published its findings on arXiv on August 10. They tested the model using ARC-AGI-1, a benchmark for nonverbal reasoning. BDH-CQ scored nearly 30% on the test. However, its smaller design makes the result especially notable. The model uses only 150 million parameters. By comparison, many leading AI systems use billions of parameters. The researchers also compared its costs with a leading OpenAI reasoning model. BDH-CQ needed roughly 11 times fewer relative tokens for a similar benchmark result.
Moving Beyond Transformers
Most modern AI systems rely on transformer architecture. These models process text and generate responses through sequences of tokens.
That method works well, but complex tasks can require significant computing power. Longer reasoning can also increase memory use across large GPU systems. Pathway takes a different route with BDH-CQ. Its system uses numerical vectors to represent relationships and patterns.
The model also includes a latent reasoning engine. It repeatedly reviews its internal results before producing a final answer.
As a result, the system can spend more time reasoning without creating long written token chains. This approach could help reduce memory pressure and operating costs.
Researchers Independently Reproduced The Model’s Benchmark Results.
Richard Zhong and transformer researcher Łukasz Kaiser also reviewed the findings. Pathway plans to scale its architecture significantly. The company also wants to test harder reasoning benchmarks and build a full language model. If those efforts succeed, this architecture could support cheaper AI systems. It may also find uses in cybersecurity, industrial operations, and other complex fields. The findings do not prove that transformers will disappear. Instead, they suggest that AI progress may depend on smarter architectures, not only larger models.
