why the geometry of reasoning might matter

A model is working through a problem. At step four it carries a quantity across incorrectly — nothing dramatic, no confusion, just a small wrong turn. Steps five through nine build on it, fluently and confidently. The final answer is wrong.

Now read the transcript. Step four looks like every other step: same measured tone, same plausible shape. The error is not hiding. It is simply invisible in the medium we happen to be watching.

Why reading the text is not enough

A model producing a reasoning chain is satisfying two constraints at once. One is fluency: does this look like a well-formed continuation? The other is coherence: does this actually follow from what came before? Most of the time the two travel together, which is why reading a chain of reasoning works as well as it does.

But they are different constraints, and they can come apart. A step can be entirely fluent and not follow at all.

Text is where fluency lives. It is what was optimized, and it is what you are reading. So when fluency and coherence separate, a detector that reads the text is not merely unlucky — it is measuring the wrong quantity.

Which raises the obvious question: is there somewhere else to look?

Reasoning as a path

Inside a language model, every step is represented by a large collection of numbers called a hidden state. As the model moves from one step to the next, those states move too. A chain of reasoning traces a path through a very high-dimensional space.

That path is what we studied in GeoReason.

The premise is a claim about motion, not location. Correct reasoning ranges widely — it has to, since it covers different territory as it goes. What stays regular is the step-to-step transition: the kind of move that follows from a given prefix belongs to a comparatively narrow family. A first error is a move outside that family. It shows up as a localized departure — an excursion — and once the trajectory has left, later steps tend to travel further out.

The everyday version of this: you do not need to read someone’s mind to notice they have walked off the trail.

The part that surprised us

It would be tidy to stop there. The more useful result is the one that complicates it.

We built two detectors. A teacher uses step labels to construct a contrastive geometric lens and score how unstable each transition is. It is accurate, but it is not deployable — at deployment you do not have labels. So we distilled it into a student that scores raw hidden states in a single pass, with no labels at all.

In-domain, both beat entropy-based, probing-based, and attention-based baselines. Then we changed the model and changed the dataset. The teacher held up. The student collapsed.

That gap is the finding worth carrying out of the paper. The teacher’s transfer says the geometric signal is not an artifact of one model or one benchmark — the structure is shared. What fails to transfer is the decision boundary. The margin separating coherent transitions from excursions shifts when the distribution shifts, and a student that has learned where the boundary sat no longer finds it.

So the obstacle to deploying this is not detecting the signal. It is preserving the margin under shift. That is a narrower problem than “generalization is hard,” and a more honest description of where the work stands.

The question this changes

These are not the same question, and the second one is answerable from a quantity the first cannot see.

There is a broader habit underneath this. We tend to evaluate reasoning the way we grade an exam: check the final answer, maybe skim the working. But reasoning is a process, and processes have dynamics. If a failure leaves a signature in how a model moves, then what matters is not only that it arrived somewhere wrong.