Science News · Science

AI Model Explores Reasoning Without Explicit Language Steps

An experimental AI system, BDH-CQ, is being tested to see if artificial intelligence can reason through problems without generating word-by-word explanations for each step, potentially reducing computational costs.

Current AI often uses 'chain-of-thought' reasoning, producing lengthy textual steps. This method is computationally intensive, making AI responses slower and more expensive.

The BDH-CQ model processes reasoning puzzles internally, updating a fixed-size memory with examples. This allows it to solve problems without converting intermediate steps into language.

Researchers report BDH-CQ solved nearly 30% of puzzles on the ARC-AGI-1 evaluation set, demonstrating that smaller models can tackle new reasoning problems without explicit textual intermediate steps.

The model showed varied performance, excelling at shape manipulation but struggling with color changes and complex rule combinations.

BDH-CQ's approach may reduce computing costs, with an estimated cost per query of $0.00070, about one-eleventh of GPT-5.6 Luna on the same benchmark.

Experts note BDH-CQ's specialized architecture makes direct comparison with general AI difficult. Further testing is needed to isolate design effects from training effects.

The study raises questions about whether AI requires language for all reasoning processes, suggesting language might not be the most efficient internal representation for every computational step.

AI-samenvatting op basis van de bron.