There may soon be a new kind of artificial intelligence in town, one that uses a novel approach to thinking that could save massive amounts of computing power and money.
The AI that most people use every day, like ChatGPT, Claude or Google Gemini, relies on large language models that can work through difficult problems step by step. While this may ultimately give us the answers we are searching for, the process can increase response times and use a lot of computing power.
So scientists at AI company Pathway decided to test an alternative approach. The research and technical foundations of the team’s work are in a paper posted on the arXiv preprint server.
A new approach to AI thinking
The team developed a new model called BDH-CQ. It has just 150 million parameters, which are the internal values that an AI adjusts as it learns patterns. Some of today’s largest AI models use hundreds of billions of these settings.
Instead of generating a long chain of written words to work through a problem, BDH-CQ reads information step by step into a constantly updating internal memory. It then runs its reasoning inside an internal workspace before revealing its answer. As the authors explain in their paper, “Demonstrations modify recurrent memory at inference time, and the resulting task is solved through iterative continuous computation rather than a verbalized chain of thought.”
Putting BDH-CQ to the test
To test the model, the researchers used a benchmark called ARC-AGI-1. This test shows the AI a few visual examples on colored grids and asks it to figure out the hidden pattern to complete a new one.
The BDH-CQ model solved 29.5% of the tasks it was tested on when given two attempts per puzzle. According to the Pathway team, it costs $0.00070 per task, or less than one-tenth of a cent, setting a new accuracy-cost efficiency record for the ARC-AGI-1 benchmark. An independent team from Bielik AI and New York University separately tested the model and confirmed the results.
“BDH-CQ demonstrates that in-context learning and recurrent latent reasoning can coexist in a compact, practical system.”
As good as this may be, BDH-CQ did have its shortcomings. While it easily handled basic visual tasks like copying shapes or extending lines, it struggled with more demanding tasks, particularly those involving ordering larger sets of objects or several steps of visual reasoning.
The next steps for the researchers are to train larger versions of their model on more complex mathematical, linguistic and visual problems. https://techxplore.com/news/2026-09-kind-ai-cheaply-words.html
Recent Comments