A new artificial intelligence (AI) model used a novel approach to AI cognition to dramatically reduce the cost of requests, suggesting that nonverbal reasoning may be the next step toward machines developing human-like intelligence.
In a new research paper published Aug. 10 on the preprint server arXiv, scientists at AI company Pathway detailed the technical foundations of its new BDH-CQ model. This follows a precursor model known as “Dragon Hatchling” that the scientists created in 2025, which was designed to accurately simulate how the neurons in the brain connected and strengthened during the learning experience.
In the new study, the scientists described how they evaluated BDH-CQ’s performance against a foundational 2019 benchmark that helped set the current standard for measuring progress toward artificial general intelligence (AGI) — the point at which AI has matched or surpassed human capabilities in all domains.
Latest Videos FromLive Science
The 2019 benchmark, known as ARC-AGI, uses nonverbal reasoning puzzles — such as rotating a series of shapes to complete a sequence — to measure the cognitive ability of AI systems. Whereas humans are highly skilled at inferring the rules of these types of puzzles through trial and error, early AI systems were historically much less skilled.
BDH-CQ scored almost 30% on the ARC-AGI-1 benchmark, successfully solving the equivalent of three out of 10 puzzles in two or fewer attempts. Although numerous models have achieved significantly better scores on this test, the underlying reasoning approach that BDH-CQ is based on makes its size and usage costs dramatically smaller than models built atop the traditional transformer-based architecture.
For example, while OpenAI’s entry-level lightweight reasoning model GPT 5.6 Luna (Low) achieved a slightly higher score, the study stated that this “modest accuracy gain” cost roughly 11 times as much as BDH-CQ in terms of relative token costs — the metering system that AI companies use to measure the cost of running AI systems. This type of AI model architecture, if adopted widely, could have a dramatic impact on the overall cost and scale of AI deployments, the scientists believe.
More than meets the eye
BDH-CQ was trained on just 150 million parameters, while parameters for the most advanced, “frontier” AI models such as Meta’s open-source Llama 3 70B or Llama 3.1 405B typically number tens of billions to hundreds of billions. In the world of AI development, fewer parameters means that models are faster to train and cheaper to run.
Get the world’s most fascinating discoveries delivered straight to your inbox.
The researchers, however, said these results also imply that the model’s cognition capabilities could scale significantly when expanded to larger parameter sizes.
The reason for this performance jump is that Pathway’s model uses what the company’s scientists describe as a “post-transformer” architecture.
Most mainstream AI models, such as those powering Claude and ChatGPT, are based on “transformer models,” so called because they transform user inputs into interconnected mathematical reference points. These systems look at every word within an input simultaneously, which allows them to infer context from position, such as deciding based on nearby words whether the word “bark” refers to dogs or trees.
Leading AI models have been criticized for being expensive to run.
(Image credit: Jaque Silva/NurPhoto via Getty Images)
A transformer model forms its responses to user queries by looking at the full prompt simultaneously and then predicting what the next word in the sequence of its reply should be. It does this word by word, using natural language to effectively verbalize a linear train of thought in the background. Transformers’ reasoning also functions sequentially, meaning they have to work through each stage of a problem in a strict linear order.
These models have significant advantages over earlier architectures, which would often forget the start of an input by the time they reached the end. However, transformer architectures can struggle with longer or more complex prompts, as the computational complexity of evaluating the prompt increases quadratically — meaning that doubling the length of an input uses four times as much processing power.
AI model usage is measured on a per-token basis, with a token representing any data fragment (equivalent to roughly four characters of text) that the AI has to ingest or output. Because more complex prompts require longer trains of thought with multiple steps, processing and responding to these queries can burn through significant amounts of tokens.
AI’s next generation?
Conventional transformer-based token generation is prone to causing memory bottlenecks, as AI re-reads every previous word in the conversation with every new word generated. Eventually, this will clog up the memory in the graphics processing units (GPUs) used for AI operations.
Because of this, scaling AI reasoning has become an expensive computational challenge. Pathway’s post-transformer approach changes how the AI’s memories of a conversation and the relationship between pieces of information are stored and processed. It replaces text logs with new tools, including an improved short-term memory and a mechanism that allows it to work through problems without consuming tokens.
Transformer-based models retain prompts and interaction histories as a long string of numerical values representing the text of requests. That string then expands as new tokens are added through processing the request. BDH-CQ uses numerical arrays to represent the underlying rules and contextual patterns of a task, using numbers to track relationships between chunks of information rather than defining them in text.
These arrays represent vectors — directional information that points to another point on a theoretical map stored inside the GPU’s memory as part of the training data, implanted during the model’s creation. The scientists said in the study that this allows the model to process complex abstract reasoning without increasing its memory footprint or computational cost.
To execute tasks, BDH-CQ implements a “latent reasoning engine” as its internal workspace. Using numbers to represent the different elements of a prompt or problem, it carries out a series of iterative recurrent loops to determine the best answer to return based on the prompt. The model takes the output of the last loop, assesses how the result could be improved based on its training data, and feeds back the previous output as the starting point for the next iteration. It repeats this for a pre-set number of loops, with each iteration theoretically closer to the desired outcome.
To tackle more complex problems requiring more thinking time, BDH-CQ can execute more loops. This increases the time taken, but the amount of memory and computational power consumed does not scale with more attempts — in theory, the model would consume a consistent proportionality of memory and power running 200 loops as it would running 20 loops. Standard transformer models, by contrast, achieve extra thinking time by generating long chains of written text tokens, which exponentially consumes GPU memory and computing power across an AI cluster.
The model’s ARC-AGI-1 benchmark results have been independently verified and reproduced by prominent researchers in the AI field, including NYU researcher Richard Zhong, and Łukasz Kaiser, a co-author of the seminal 2017 paper “Attention Is All You Need,” which introduced the concept of transformers within large language models.
“I’ve followed Pathway closely and replicated their ARC-AGI-1 results myself,” Kaiser said in a statement. “Pathway shows that model architecture, not just scale, can drive the next leap in AI reasoning.”
Pathway plans to scale the BDH architecture up to 600 billion parameters and apply its vector-based reasoning to more challenging benchmarks, such as ARC-AGI-2 and ARC-AGI-3, as well as develop a fully-fledged large language model (LLM) based on the technology, which would provide a basis for building text-based chatbots. The company hopes the technology can be applied to complex reasoning problems in sectors such as cybersecurity incident response and industrial operations.
