U-Lens

U-Space offers label-free uncertainty detection for LLMs

Researchers released U-Space and its U-Lens, which combine hidden-state doubt directions with mean token entropy to estimate LLM confidence in a single pass, with no labels, resampling, or fine-tuning, and distinguish ambiguity, missing information, and conflicting evidence. It beat multi-pass and supervised baselines on three models and four benchmarks; the only model-specific ingredient is a pre-fitted Anthropic Jacobian lens, with 40+ available.