Logit Leakage through Floating-Point Bias Addition

Shaomin Chen, Xiaorui Gong

Language model APIs can expose token-specific biases while withholding logits and token probabilities. Existing bias-based extraction recovers relative logits, leaving a common add…