Researchers find reasoning circuits in AI models 1,000x faster than before

A new technique called CircuitProbe can identify which parts of a language model are responsible for reasoning in under 5 minutes on a regular computer, instead of 25 hours on expensive GPUs. This matters because it makes it practical to understand how these models actually work, and potentially to make smaller models run better by duplicating their reasoning circuits.

Right now, understanding what's happening inside AI models is mostly trial-and-error — expensive, slow, and inaccessible to anyone without industrial-scale compute. This technique opens a path to reverse-engineering model behavior with ordinary hardware, which means smaller organizations and researchers can actually see what their models are doing rather than treating them as black boxes. For practical deployment, it suggests that small language models (under 3 billion parameters) can be made more capable by duplicating specific internal structures, which is a concrete way to get better reasoning without training a bigger model.