Language models reveal when they are uncertain during reasoning
U-Space: Uncovering When and Why Uncertainty Arises in Language Models
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Language models sometimes make mistakes while sounding confident, making it hard to know when to trust them. The authors developed a way to find and measure uncertainty inside the model’s thought process without needing extra training or repeated tests. They created a special space called U-Space that shows where and when the model feels doubt or certainty as it generates answers. This method gives better and more understandable confidence scores compared to older techniques.
What this means in practice
- •For ai system developers: Detect unreliable or uncertain outputs from language models to improve decision-support tools without extra model training or repeated runs.
- •For customer service platforms: Identify when language models might be unsure in chatbots to trigger escalation or human review before providing critical information.
Authors
Tobias Braun, Nils Loose, Alexander Herzog, Virginia Ceccatelli, Marcus Rohrbach, Thomas Eisenbarth, Lorenzo Cavallaro
Abstract
Large language models are informing decisions with ever-higher stakes. As the consequences of their errors grow, a central question becomes harder to ignore: how much can we trust an individual answer? Yet recognizing when to defer remains difficult because language models can present incorrect conclusions with fluent explanations and an authoritative tone. Uncertainty quantification seeks to address this disconnect by estimating the reliability of individual predictions. However, many existing methods require repeated generations or separately trained components, and their scalar estimates do not reveal where uncertainty arises or how it evolves during reasoning. Recent work has also shown that generation length can be strongly associated with uncertainty estimates and correctness, raising the question of how much of an estimator's predictive power comes from uncertainty-specific information rather than output length alone. Mechanistic interpretability offers a way to address these limitations by connecting human-interpretable concepts to intermediate model states. Building on this capability, we introduce the U-Space, a low-dimensional subspace that makes a model's evolving uncertainty measurable and interpretable. We identify semantic anchors for doubt and certainty, map their unembedding directions back into the residual space, and combine their contrasts into an orthogonal basis. The U-Lens projects each token state onto these basis vectors, yielding an interpretable token-level uncertainty map that can be inspected directly or aggregated into a scalar uncertainty score. Our approach requires no correctness labels, repeated generations, or training. Across reasoning benchmarks, its confidence score outperforms established baselines under both standard and length-controlled evaluation and transfers more reliably than supervised estimators. Code: https://github.com/s2labres/U-Space.