U-Lens
An uncertainty estimator built on the U-Space framework. It uses the Jacobian Lens and the U-Space basis to produce an interpretable uncertainty classification from a single forward inference pass.
Timeline (last 90 days)
The core scoring formula of U-Lens is: score = strength of the hidden state pointing toward different directions of doubt × average entropy of the reasoning tokens.
U-Lens requires no labeled labels, no training or fine-tuning, can be completed in a single forward pass, and requires no resampling.
U-Lens can distinguish four types of uncertainty: ambiguity, missing information, conflicting evidence, and general uncertainty.
In evaluations across 3 seeds, 3 models, and 4 benchmarks (involving Gemma 4, Qwen3.5, and Magistral), U-Lens is the best estimator, ahead of 5-round sampling, 10-round sampling, and supervised baselines.
The only model-specific component U-Lens relies on is the Jacobian Lens (J-Lens), from Anthropic's global workspace paper.
All Facts (5)
The core scoring formula of U-Lens is: score = strength of the hidden state pointing toward different directions of doubt × average entropy of the reasoning tokens.
50%UnverifiedU-Lens requires no labeled labels, no training or fine-tuning, can be completed in a single forward pass, and requires no resampling.
50%UnverifiedU-Lens can distinguish four types of uncertainty: ambiguity, missing information, conflicting evidence, and general uncertainty.
50%UnverifiedIn evaluations across 3 seeds, 3 models, and 4 benchmarks (involving Gemma 4, Qwen3.5, and Magistral), U-Lens is the best estimator, ahead of 5-round sampling, 10-round sampling, and supervised baselines.
50%UnverifiedThe only model-specific component U-Lens relies on is the Jacobian Lens (J-Lens), from Anthropic's global workspace paper.
50%