Researchers at AI company Pathway have developed a new reasoning model that could challenge the transformer architecture behind most large language models. Called BDH-CQ, the experimental system uses a different approach to processing information and can operate with up to 11 times fewer computing resources than a leading OpenAI model, tells Live Science.
Most modern AI models break information into tokens, which are processed through transformer networks. BDH-CQ instead represents information as numerical vectors. Its architecture is based on the Baby Dragon Hatchling, or BDH, framework, which is designed to perform iterative reasoning while keeping computational and memory requirements relatively low.
The researchers tested BDH-CQ on ARC-AGI-1, a benchmark designed to evaluate an AI system’s ability to recognize patterns and solve unfamiliar reasoning problems. The model scored about 30%. Although this result trails some larger reasoning systems, BDH-CQ contains only 150 million parameters, making it much smaller than today’s leading large language models.
Its efficiency is particularly significant. According to the researchers, the model can tackle reasoning tasks without continually expanding the amount of memory required as problems become more complicated. This could reduce the infrastructure and energy costs associated with running advanced AI systems.
The architecture has also attracted attention from AI researchers, including Łukasz Kaiser, one of the authors of the influential 2017 paper that introduced transformers. Independent examination of the approach supports the researchers’ claim that it represents a genuine architectural departure rather than another variation of existing transformer models.
Pathway now plans to scale the BDH architecture dramatically, potentially reaching 600 billion parameters. The company envisions applications ranging from conversational AI to cybersecurity. If the architecture retains its efficiency at much larger scales, BDH-CQ could suggest an alternative path for advancing AI through architectural changes rather than simply building ever-larger models.
