U-Space: Uncovering When and Why Uncertainty Arises in Language Models
U-Space maps language-model residual states into a subspace that tracks token-level doubt and certainty without extra training.
U-Space is a training-free method that locates a low-dimensional residual subspace associated with doubt and certainty in language models. Semantic anchors are mapped from unembedding directions into residual space and combined into an orthogonal basis, and a U-Lens projects each token state onto that basis. The result is a token-level uncertainty map that can also be aggregated into a scalar score, without correctness labels or repeated generations. On reasoning benchmarks the confidence score outperformed established baselines under standard and length-controlled evaluation and transferred more reliably than supervised estimators.
- Builds an orthogonal residual basis from doubt and certainty anchors
- Projects each token state into an interpretable uncertainty map
- Needs no correctness labels, extra training, or repeated generations
- Beats baselines under standard and length-controlled evaluation
- Transfers more reliably than supervised uncertainty estimators
From the original article
1 min read at huggingface.co
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.
By Tobias Braun, Nils Loose, Alexander Herzog · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.