U-Space

A Clear Signal for Model Uncertainty, No Extra Sampling Needed

Building on open-weight reasoning models often leaves you wondering how confident the model actually is in its answer. The typical fix, multiple sampling rounds, a judge model, or labeled datasets, is expensive and slow. U-Lens changes that. It extracts an uncertainty signal from a single forward pass, no resampling required. The approach combines hidden-state direction analysis with token entropy, and it outperforms multi-pass baselines across several benchmarks. This is practical, label-free, and ready to use.