New AI Model Thinks Cheaply Without Words

Researchers at Pathway have introduced BDH-CQ, a compact artificial intelligence model that fundamentally reimagines computational reasoning by eliminating the need for verbalized chain-of-thought processes. Unlike prevailing large language models that rely on extensive parameter counts and sequential text generation to solve problems, BDH-CQ utilizes just 150 million parameters. The architecture operates by ingesting information step-by-step into a continuously updating recurrent memory. Reasoning occurs internally within a latent workspace through iterative continuous computation, allowing the system to adjust its internal state based on demonstrations before outputting a final answer. This approach decouples reasoning from language generation, significantly reducing computational overhead and latency. The model capabilities were rigorously evaluated on the ARC-AGI-1 benchmark, a standardized test designed to measure in-context learning and visual pattern recognition. During evaluation, BDH-CQ successfully completed 29.5 percent of complex visual puzzles across two attempts per task. Crucially, the model achieved this performance at a computational cost of $0.00070 per task, a figure that shatters previous cost-accuracy Pareto frontiers for the benchmark. Independent validation by researchers at Bielik AI and New York University confirmed these efficiency metrics, establishing BDH-CQ as the current state-of-the-art for resource-constrained visual reasoning. Despite its breakthrough efficiency, the architecture exhibits distinct limitations. The model demonstrates high proficiency with straightforward pattern extension, shape copying, and spatial recoloring tasks. However, performance degrades when handling operations requiring strict ordering of large object sets or multi-step visual reasoning chains. These constraints highlight the ongoing challenge of scaling non-verbal reasoning frameworks to match the cognitive flexibility of larger, text-based systems. Pathway researchers have outlined a clear development roadmap focused on scaling the BDH-CQ architecture to address these performance gaps. Future iterations will target training on increasingly complex mathematical, linguistic, and hierarchical visual datasets to expand the model latent reasoning boundaries. The underlying research, detailing the architectural mechanics and benchmark performance, has been publicly documented on the arXiv preprint server. By proving that compact models can execute sophisticated inference without generating intermediate text, this development signals a potential shift toward highly optimized, cost-effective AI systems suitable for edge deployment and large-scale industrial applications.