U-Space maps LLM uncertainty token by token, no training
TL;DR
- U-Space is a low-dimensional subspace that renders a language model's token-level uncertainty measurable and interpretable without any additional training.
- A probe called the U-Lens projects token states onto basis vectors to produce both per-token uncertainty maps and an aggregated scalar confidence score.
- The authors claim the score outperforms established baselines and transfers more reliably than supervised estimators across reasoning benchmarks.
"Language models can present incorrect conclusions with fluent explanations and an authoritative tone." That one line, from a new arxiv preprint titled "U-Space: Uncovering When and Why Uncertainty Arises in Language Models," is the whole motivation for the method it introduces.
The authors, Tobias Braun, Nils Loose, Alexander Herzog, Virginia Ceccatelli, Marcus Rohrbach, Thomas Eisenbarth and Lorenzo Cavallaro, define U-Space as "a low-dimensional subspace that makes a model's evolving uncertainty measurable and interpretable," and read it with a probe they call the U-Lens, which projects token states onto basis vectors to produce both per-token uncertainty maps and an aggregated scalar confidence score. The score, they report, "outperforms established baselines under both standard and length-controlled evaluation and transfers more reliably than supervised estimators" on reasoning benchmarks. The method needs no correctness labels, no repeated generations, no additional training.
The preprint names no specific models in its public abstract and publishes no per-benchmark figures. Code is posted at github.com/s2labres/U-Space.
Originally reported by paper
Read the original article →Original headline: U-Space Maps LLM Uncertainty to Individual Tokens, No Training Required