U-Space: Uncovering When and Why Uncertainty Arises in Language Models

U-Space. Uncovering When and Why Uncertainty Arises in Language Models. By Tobias Braun and six others, from the Technical University of Darmstadt, University College London, and three other institutions. Published on the sixth of October.

Large language models now inform decisions with ever-higher stakes, and as the cost of their errors grows, so does a plain question: how far can one individual answer be trusted? Knowing when to defer is difficult, because a model can deliver a wrong conclusion with a fluent explanation and an authoritative tone. Uncertainty quantification tries to close that gap by estimating how reliable each prediction is.

The existing methods carry three limitations. Many need repeated generations or separately trained components. Their output is a single scalar, which says nothing about where uncertainty arises or how it changes during reasoning. And recent work has shown that generation length can be strongly tied to both uncertainty estimates and correctness, which raises a pointed question: how much of an estimator's predictive power is about uncertainty at all, and how much is just output length?

Mechanistic interpretability offers a way in, since it connects concepts people can understand to the model's intermediate states. Building on that, the authors introduce the U-Space, a low-dimensional subspace in which a model's uncertainty, as it evolves, becomes measurable and interpretable. The aim answers each of the complaints above. It should be cheap enough to need no extra generations, local enough to show where doubt sits and how it moves, and grounded enough in identifiable concepts that what it measures is uncertainty rather than something incidental, such as length.

Construction starts from semantic anchors, one set for doubt and one for certainty. Their unembedding directions are mapped back into the residual space, and the contrasts between them are combined into an orthogonal basis. A tool called the U-Lens then projects each token state onto those basis vectors. The result is a token-level map of uncertainty that can be read directly, or aggregated into one scalar score when a single number is wanted. In other words, the measuring stick is built from the model's own representation of what doubt and certainty mean, and every token in a chain of reasoning gets a position along it. None of this requires correctness labels, repeated generations, or any training. The model runs once, as it normally would, and the reading is taken from states it already produces.

On reasoning benchmarks, the confidence score derived this way outperforms established baselines. It does so under standard evaluation and also under evaluation controlled for length, which speaks directly to the worry that length alone could be doing the work. It also transfers more reliably than supervised estimators, and it gets there without the correctness labels that supervision depends on. The code is public, under the s2labres account on GitHub.

The shift is from a verdict to a trace. Instead of one number attached to a finished answer, the U-Space offers a reading at every token of where doubt and certainty sit inside the model, and that reading still collapses into a competitive score when one is needed, without sampling the model again or training anything on top of it.