Head-to-Head Comparison of B1 Token Entropy and Logit-Based Uncertainty Methods

Develop and run a formal head-to-head comparison between B1 token-entropy–based uncertainty detection and logit-based alternatives such as Semantic Energy, LogTokU, and PRO on matched datasets, with special attention to multi-cluster regimes, to quantify relative discrimination performance.

Background

The paper argues that B1 token entropy and logit-based approaches (e.g., Semantic Energy) share mechanisms that retain signal in single-cluster regimes where sampling-based diversity collapses, and suggests B1 as a zero-cost approximation.

However, a direct experimental comparison across matched data—particularly in cases without single-cluster collapse—has not been conducted, leaving the relative strengths of these methods unresolved.