U-Space Framework Introduces U-Lens for Label-Free Uncertainty Estimation in Reasoning Models

The U-Space project introduces U-Lens, a method for estimating uncertainty in open-weight reasoning models without requiring correctness labels, fine-tuning, or multiple sampling passes. The approach calculates a score by combining the strength of hidden state projections onto four specific "doubt" directions with the mean token entropy of reasoning tokens.

How U-Lens Works

U-Lens operates as a single-pass, label-free estimator. It identifies the type of uncertainty present in a model's output, categorizing it into four distinct types:

• Ambiguity
• Missing information
• Conflicting evidence
• General uncertainty

The method relies on a Jacobian lens (J-Lens), a model-specific ingredient derived from Anthropic's global workspace research. Once a J-Lens is obtained, the U-Space basis can be constructed on CPU using only the unembedding rows of a few hundred words, avoiding the need to load the full model for this step. At inference, the system extracts one hidden state at the end of the thinking process, projects it onto four vectors, and multiplies the result by the mean entropy derived from logits.

Performance and Availability

In evaluations across 3 seeds, 3 models (Gemma 4, Qwen3.5, and Magistral), and 4 benchmarks, U-Lens served as the best estimator, outperforming 5-pass and 10-pass sampling methods as well as supervised baselines.

• Pre-fitted Lenses: Over 40 pre-fitted J-Lenses are available on Hugging Face (neuronpedia/jacobian-lens), covering models such as Gemma 2/3/4, Qwen3/3.5/3.6, Llama 3.1/3.3, gpt-oss-20b, and OLMo 3.
• Custom Fitting: For unsupported models, users can fit a lens using Anthropic's reference code (anthropics/jacobian-lens) with approximately 100 unlabeled prompts.
• Implementation: The code is available on GitHub (s2labres/U-Space). It is currently released as research code rather than a pip package.