Language model APIs frequently expose token-specific biases while keeping logits and token probabilities hidden. Attackers can exploit these biases to extract information about the model's internal scoring, but existing techniques have a limitation: they recover only relative logits, leaving a common additive component undetermined under ideal arithmetic.

The new arXiv paper, "Logit Leakage through Floating-Point Bias Addition," addresses this gap. The abstract describes how bias-based extraction works and notes the unresolved additive component. The title points to a method that uses floating-point bias addition to resolve this ambiguity, potentially allowing an attacker to recover the full logit distribution.

If successful, this approach would strengthen existing leakage attacks by removing the last unknown constant, giving a complete picture of the model's token probabilities. The abstract is brief, but it highlights a concrete weakness in how APIs handle bias parameters while protecting logits.