Researchers at KAIST and Naver AI Lab have found that different kinds of reasoning correspond to distinguishable patterns inside language models. The team identified eight operations, including extracting facts, decomposing a problem, recalling a formula, deducing and calculating, then examined models as they solved mathematics questions.
The experiments covered Qwen2.5-7B, Qwen3-8B and Gemma4-31B, with GPT-5 used to label segments of each written solution. Classifiers working from internal activations separated the operations most clearly in the models’ middle layers and outperformed classifiers that saw only the output tokens. The same common word could acquire a different representation depending on the surrounding reasoning operation.
Blocking attention to the preceding 30 tokens weakened the signal, suggesting each step builds on earlier context rather than forming independently. An incorrect calculation still looked internally like calculation, so recognizing an operation does not establish that it was performed correctly. The result also replicated on Llama-3-8B, and a Qwen classifier transferred to two other math tests. However, the study covers a small set of models and mathematics tasks. Whether these signatures can detect mistakes, monitor hidden reasoning or steer generation remains unproven.