Researchers Developed AI Model That Reasons Without Text
The experimental system, BDH-CQ, avoids generating intermediate text to reduce computational overhead.
Updated on Sept. 22, 2026 in Quantum Computing
Live Poll
Should AI systems be required to display their reasoning process to be considered trustworthy?
On August 10, 2026, researchers submitted a paper to arXiv detailing BDH-CQ, an experimental AI system that performs reasoning without creating intermediate language outputs. This research-stage model currently solves nearly 30 percent of ARC-AGI-1 puzzles within two attempts.
Why it matters
The model aims to reduce the massive computational power, time, and costs typically required to generate reasoning outputs. By bypassing text-based chains of thought, this design seeks to make complex AI tasks more efficient.
The system processes queries by updating a fixed-size memory with examples rather than maintaining previous inputs within a standard context window. It performed at 1/11th the cost of GPT-5.6 Luna, solving 3 in 10 puzzles on the ARC-AGI-1 evaluation set.
The players
BDH-CQ
An experimental AI model designed to reason through fixed-size memory without generating intermediate text.
ARC-AGI-1
A benchmark evaluation set used to measure general intelligence and reasoning capabilities in AI systems.
GPT-5.6 Luna
A proprietary large language model serving as the cost-comparison baseline for current AI reasoning queries.
The details
BDH-CQ functions by conducting problem-solving operations internally, entirely skipping the step of converting thoughts into language or intermediate tokens. Instead of keeping all historical inputs in a context window—a segment of memory used by models to track previous conversation turns—the system utilizes a fixed-size memory structure. This design forces the model to store and recall specific examples, rather than re-processing entire sequences of text for every reasoning step.
Timeline
August 10, 2026: Researchers submitted the BDH-CQ paper to arXiv.org.
September 22, 2026: Article publication date.
The Tech Race
This research follows a growing effort to move beyond the current industry standard of chain-of-thought prompting by testing if models can reach similar logic conclusions without language generation. It sits in direct competition with transformer-based architectures that rely heavily on the ARC-AGI-1 evaluation set to prove reasoning capabilities.
This development is currently in the research stage and not available for commercial use. If validated through further study, it could eventually lower the compute costs associated with running high-reasoning tasks on cloud AI platforms.
The takeaway
The study suggests that future reasoning models might shed the heavy computational tax of text-based chains, provided the design proves robust in peer review. Researchers must next clarify the distinction between the architecture's efficiency and the training methodology used.
Further reading
For more on evolving model architectures, see Quantum Computing.
Live Poll
Should AI systems be required to display their reasoning process to be considered trustworthy?
