S2 Lab Proposes U-lens: Single-Pass Uncertainty Estimation Combining J-lens and Entropy Outperforms 12 Baselines in Self-Tests

Researchers Virginia Ceccatelli, Xiaoyu Liang, and Tobias Braun from S2 Lab published a new uncertainty estimation method called U-lens on October 17, 2026. The method aims to monitor the confidence of Large Language Models (LLMs) during reasoning in real-time, requiring no training labels and only a single forward pass.

Building on Anthropic's global workspace theory and the J-lens technique, the team constructed four orthogonal directions in the model's hidden state space: ambiguity, incompleteness, conflicting evidence, and general uncertainty. These directions form the "U-space." By measuring how closely the model's hidden states align with these directions and combining this with the average entropy of the next-token probability distributions, U-lens generates a scalar score indicating the likelihood of an answer being wrong.

According to the authors' self-tests, U-lens separated right from wrong answers better than 12 other methods across three reasoning models and four benchmarks. The method held up even when controlling for reasoning trace length to remove potential confounds. Early tests also suggested similar readouts could help flag suspicious behaviors like reward hacking. The code has been open-sourced.

This result is currently a first-party benchmark (E2) and has not yet been independently reproduced. The authors note that while their experiments are reproducible, whether the approach can become a broader tool for monitoring models remains an open question.