New Artificial Intelligence Model Tests Reasoning Without Using Words
A small AI model developed by BDH researchers can solve puzzles without spelling out intermediate steps in words.
BDH-CQ, an experimental small AI model, has demonstrated the ability to solve logic puzzles without generating long chains of words, offering a more efficient approach.
A New Approach That Bypasses Words
AI systems typically process information as streams of words when solving a problem. A new model is testing whether certain reasoning processes can occur without being put into words.
Use of Fixed-Size Memory
The experimental system, named BDH-CQ, can solve puzzles without converting its intermediate thoughts into words. The model uses each instance to update a fixed-size memory, eliminating the need to look back at past data.
Internal Solution Process
Zuzanna Stamirowska, complexity scientist and CEO of Pathway, states that when a query arrives, the model does not put its thoughts into words at all. The model solves the problem internally without translating each step into language.
ARC-AGI-1 Performance
Stamirowska and her colleagues reported that, when given two attempts, the model solved roughly three out of ten puzzles in the public ARC-AGI-1 evaluation set. The model performed better on shape rotation and movement tasks.
Lower Costs and Next Steps
Researchers note that this process could help lower computing costs. Experts evaluate this as an interesting efficiency result while stating that additional testing is needed to fully assess the architecture.