Researchers at South Korea's KAIST and Naver AI Lab have shown that the reasoning steps a language model writes out in text correspond to separable patterns inside its numerical representations. The team tested three models — Qwen2.5-7B, Qwen3-8B and Gemma4-31B — and found the separation is strongest in the middle layers. For companies that rely on a model's chain of thought as an oversight tool, this matters because it suggests the written explanation is not the only place where the reasoning can be observed.
What the study measured
The team defined eight recurring reasoning operations, among them extraction, decomposition, formula recall, deduction and computation. Three models were set to solve math problems, their solution paths were split into segments, and GPT-5 was then used to label each segment with one of those eight operations. The result held across all three models: the operations can be reliably told apart in internal representations, and the separation peaks in the middle layers rather than at the input or output end. The work therefore rests on a fixed taxonomy of eight labels applied to segmented solution paths, not on free-form descriptions of what a model appears to be doing.
To rule out simpler explanations, the researchers ran a classifier that looked only at the tokens used. It performed worse than the classifier analyzing internal representations. Position within the solution path did not explain the effect either. That means the internal states carry information about the type of reasoning step that goes beyond surface-level wording — a distinction that matters when the same phrase can appear in a deduction and in a calculation.
Common function words such as "a," "is" or "the" appear across very different reasoning steps. In the early layers their representations are still jumbled together, but by the middle and later layers they separate according to the surrounding operation. The same word gets a different internal representation depending on which reasoning step it belongs to. A separate intervention blocked attention to the preceding 30 tokens; the signal for the operation weakened, indicating that reasoning steps do not emerge in isolation but build on the preceding context.
What this means for business
Even on incorrectly solved problems, the type of step the model was performing stayed identifiable — whether it was computing, retrieving a formula or deducing. A flawed computation step still looked like a computation step internally, even when the result was wrong. For teams that audit model output, this points to a possible distinction between a wrong answer produced by a correct type of operation and a failure that comes from the model performing the wrong operation altogether. The two cases call for different fixes, and today both are usually visible only as a wrong final number.
The separability replicated with Llama-3-8B, and for Qwen3-8B the trained classifiers transferred successfully to GPQA-Diamond and MATH-500. Those are benchmarks outside the original training setup, which suggests the signal is not tied to one dataset. The limits are equally clear: the experiments cover math tasks and a handful of models. Whether the findings can be used to catch errors or steer a model mid-generation remains an open question for future work, so no vendor can yet promise runtime control built on this method.
The relationship between text output and internal computation is already a live issue for AI safety. Reading the chain of thought is one of the few oversight tools available, according to OpenAI, but Anthropic showed that models disclose the hints they used in only 25 to 39 percent of cases. A method that translates a model's internal vectors into readable text revealed that Claude Opus 4.6 processes more than what shows up in its output reasoning. With OpenAI's Astra model, the Recurrent Depth technique shifts part of the reasoning into internal numerical representations — the same space the KAIST study investigates.
The marker to watch is whether the classifier approach moves from math benchmarks to production tasks and from a handful of open models to the systems companies actually deploy. If classifiers trained on internal representations transfer to domains such as contract review or financial analysis, buyers of AI tools will have a second, independent channel for checking what a model did, not just what it said. Until that transfer is demonstrated, the practical reading is narrower: the written chain of thought is a useful but incomplete record, and the internal one is measurable in the lab.
