New AI Model Uses Nonverbal Reasoning to Cut Inference Costs

A new artificial intelligence (AI) model uses a brain-inspired approach to reasoning that could significantly reduce the cost of AI requests. The system also suggests that nonverbal reasoning may help machines develop more human-like intelligence.

In a research paper published on the arXiv preprint server on Aug. 10, scientists from AI company Pathway described the technical foundation of the new BDH-CQ model.

BDH-CQ builds on a precursor system called “Dragon Hatching,” which Pathway researchers created in 2025. The model was designed to simulate how neurons in the brain form connections and strengthen those connections during learning.

In the new study, researchers evaluated BDH-CQ against established benchmarks used to measure progress toward artificial general intelligence (AGI). AGI is the theoretical point at which an AI system matches or exceeds human performance across a broad range of tasks.

One of the evaluations was ARC-AGI-1, a benchmark based on visual, nonverbal reasoning puzzles. These tests may ask an AI system to rotate shapes, identify patterns or complete a sequence. Humans are generally effective at inferring the rules behind these puzzles through trial and error, while many earlier AI systems have struggled with them.

BDH-CQ scored almost 30% on ARC-AGI-1, solving the equivalent of three out of 10 puzzles in fewer than two attempts. Although some AI models achieve higher scores, Pathway says BDH-CQ’s underlying inference method makes it substantially smaller and less expensive to operate than many traditional transformer-based models.

For comparison, OpenAI’s lightweight GPT 5.6 Luna (Low) reportedly achieved a slightly higher ARC-AGI-1 score. However, the researchers said that this moderate improvement required approximately 11 times more tokens than BDH-CQ. Tokens are data units that AI systems process when reading prompts and generating responses.

If this architecture can be scaled successfully, it could lower the cost of AI inference and make advanced reasoning systems more practical for large-scale deployment.

A Smaller AI Model With Room to Grow

BDH-CQ was trained with approximately 150 million parameters. By comparison, leading frontier models such as Meta’s open-source Llama 3 70B and Llama 3.1 405B contain tens or hundreds of billions of parameters.

In general, models with fewer parameters can train more quickly and cost less to run. However, smaller models often have less capacity for complex tasks. Pathway researchers said their results suggest that BDH-CQ’s reasoning ability could improve significantly if the architecture is expanded to a larger number of parameters.

The performance gains appear to come from Pathway’s “post-transformer” architecture. Most mainstream AI models, including Claude and ChatGPT, rely on transformer technology.

Transformers convert text into interconnected numerical representations, allowing them to analyze relationships between words. Because they examine the context surrounding a word, they can determine whether “bark,” for example, refers to a dog or a tree.

A transformer generates a response by processing the prompt and predicting the next token in a sequence. The model repeats this process one token at a time, creating a largely linear reasoning path.

This approach improved on earlier AI architectures that could lose track of information as inputs became longer. However, transformer models can become increasingly expensive when processing lengthy or complicated prompts because their computational demands grow rapidly as input length increases.

Longer reasoning chains also require more tokens. As an AI system generates additional text to work through a problem, it consumes more memory and processing capacity, increasing the cost of each request.

Could Nonverbal Reasoning Shape the Next Generation of AI?

Traditional transformer models can create memory bottlenecks because they repeatedly refer back to earlier parts of a conversation whenever they generate a new token. Over time, the memory of the graphics processing units (GPUs) running the model can become overloaded.

Pathway’s post-transformer design changes how the system stores conversational memory and relationships between pieces of information. Instead of relying exclusively on text-based logs, BDH-CQ uses internal memory mechanisms that allow it to reason without generating a lengthy written chain of thought.

Transformer models typically represent prompts and conversation histories as long numerical sequences corresponding to text tokens. BDH-CQ instead uses numerical arrays to represent the underlying rules and contextual patterns of a task. These arrays track relationships between information rather than storing every relationship as text.

The arrays function as vectors — directional representations within a mathematical space. Pathway’s researchers say this allows BDH-CQ to work with abstract relationships while limiting increases in memory use and computational demand.

To solve a problem, BDH-CQ uses what the researchers call a “latent inference engine” as an internal workspace. The system runs iterative loops that manipulate numerical representations of the task, evaluate the results and use each output as the starting point for the next iteration.

BDH-CQ can run additional loops when a problem requires more reasoning. This may increase the time needed to produce an answer, but the researchers say it does not require the model to generate an equally long sequence of text tokens. In contrast, conventional reasoning models often use extended written chains of thought, which can increase GPU memory use and computational costs.

The ARC-AGI-1 results have reportedly been independently reproduced by researchers in the AI field. Łukasz Kaiser, a co-author of the seminal 2017 paper “Attention Is All You Need”, said he had replicated Pathway’s ARC-AGI-1 results.

“Pathway shows that model architecture, not just scale, can drive the next leap in AI inference,” Kaiser said in a statement.

Pathway plans to scale its BDH architecture to as many as 600 billion parameters and test it on more difficult benchmarks, including ARC-AGI-2 and ARC-AGI-3. The company also plans to develop a large language model based on the technology for text-based chatbots.

If the approach performs reliably at larger scales, it could support lower-cost AI applications in areas such as cybersecurity response, industrial operations and complex automated decision-making. For now, BDH-CQ remains an early research model, but its combination of nonverbal reasoning, compact architecture and token-efficient inference offers a potential alternative to conventional transformer systems.