When a model writes out that it is thinking step by step, do those steps correspond to anything inside it? A study published by researchers at KAIST and Naver AI Lab in South Korea gives a measurable answer.
The finding: reasoning operations such as calculation, formula recall and deduction are clearly separable in a model's internal states.
The method
Models were made to solve maths problems step by step. The resulting solution paths were segmented and each segment labelled with a reasoning operation, using a separate model for the labelling.
Then each segment was examined for how it appears in the model's internal states. The question was simple: can these operations be told apart from the inside?
| Detail | Value |
|---|---|
| Institutions | KAIST and Naver AI Lab |
| Models tested | Qwen2.5-7B, Qwen3-8B, Gemma4-31B |
| Replication | Repeated with Llama-3-8B |
| Operations separated | Eight reasoning types |
The signal sits in the middle layers
The clearest part of the finding is the layer distribution. The separation is strongest in the middle layers and weakens towards the early and late ones.
That held across all three models. So the result does not look like a quirk of one architecture but like a shared pattern.
That the separation concentrates in the middle layers sits about where you would expect. Early layers deal more with surface language features and late layers with choosing the next word; information about the type of operation falls between the two.
The most interesting detail
The study's standout finding sits at the word level. The same word gets a different internal representation depending on which reasoning step it belongs to.
In other words, the model does not encode a number the same way when recalling a formula as when performing a calculation. Context carries not only meaning but the type of operation.
Why it matters for safety
This work ties directly to one concern: what a model does internally may differ from the chain of thought it writes out.
Earlier research had already shown that gap. Claude Opus 4.6 was found to process more than it displays, and in Anthropic's measurement models disclosed hints they had been given in only 25 to 39 percent of cases.
Read together, those two findings give this: the written chain of thought is not all of what is inside the model but a summary of it. The summary is usually correct yet incomplete, and which part is missing cannot be seen from outside.
What it means
If reasoning types separate in internal states, it may become possible to measure what a model is actually doing without reading the text it writes. Not having to trust the written output for oversight is something interpretability work has been after for a long time.
The limit is clear too: the study used maths problems, and the classification of steps rests on another model. Whether the same separation holds on open-ended tasks has not yet been shown.