Large language models are increasingly used to inform high-stakes decisions, yet their fluent and authoritative outputs can mask incorrect conclusions. Uncertainty quantification aims to estimate the reliability of individual predictions, but many existing methods require repeated generations or separately trained components, and their scalar scores do not reveal where uncertainty arises or how it evolves during reasoning.
In a new paper, researchers introduce the U-Space, a low-dimensional subspace that makes a model's evolving uncertainty measurable and interpretable. The method identifies semantic anchors for doubt and certainty, maps their unembedding directions back into the residual space, and combines their contrasts into an orthogonal basis. The U-Lens then projects each token state onto these basis vectors, producing a token-level uncertainty map that can be inspected directly or aggregated into a scalar confidence score.
The U-Space tracks four sources of uncertainty through a model's generation: ambiguity, incomplete information, conflicting evidence, and general uncertainty. The U-Lens combines this verbalizable signal with predictive entropy, a first-order distributional measure, to estimate answer reliability from a single generation. According to the authors, the approach requires no correctness labels, repeated generations, or task-specific training.
Across reasoning benchmarks, the confidence score outperforms established baselines under both standard and length-controlled evaluation, and transfers more reliably than supervised estimators. The authors report that U-Lens leads uncertainty estimation across three reasoning models — Gemma-4-31B, Qwen3.5-27B, and Magistral-Small — and four benchmarks: MMLU-Pro, Omni-MATH, SuperGPQA, and TriviaQA.
The paper also highlights the role of output length: response length is a strong predictor before controlling for it, reaching 68.6 AUROC, but falls to 52.4 under length control. Most other baselines drop in performance when length is controlled, while U-Lens drops by only 1.8 points and remains strongest. Token-level readouts reveal when and what forms of uncertainty emerge during generation, and steering along the identified directions can make models hesitate even on simple questions, suggesting the representations are behaviorally relevant.
Preliminary experiments suggest the same label-free construction may extend beyond uncertainty to emotion and reward-hacking detection. Code is available at github.com/s2labres/U-Space, with a project page at s2labres.github.io/U-Space and the paper at arXiv:2610.09087.