U-Space: Uncovering When and Why Uncertainty Arises in Language Models [R][D]

Researchers introduced U-Lens, a method that produces a "how sure is the model about this answer?" signal for open-weight reasoning models without requiring 5x sampling, a judge model, or a labeled dataset, computing a score from how strongly a model's hidden state points toward different outcomes. The work, presented as U-Space, examines when and why uncertainty arises in language models. The approach targets builders who need uncertainty estimates from reasoning models at lower cost than existing sampling-based or judge-based methods.

If you build on open-weight reasoning models, you have probably wanted a "how sure is the model about this answer?" signal that does not cost 5x sampling, a judge model, or a labeled dataset. Hence, we built the U-Lens. In one line: score = how strongly the hidden state points toward different "dou