black-box-llms

This paper proposes methods for detecting hallucinations in black-box LLMs by combining semantic entropy and token-level uncertainty signals, evaluating techniques like TopK, CoCoA, Gated, and Stacked across multiple benchmarks to find that no single method is universally strongest but Stacked often performs best.

Introduces Stepwise Confidence Attribution (SCA), a framework for assigning step-level confidence to reasoning traces from black-box LLMs without internal access, using the Information Bottleneck principle to distinguish legitimate variability from errors. Experiments show SCA reliably identifies low-confidence steps and improves self-correction success rates by up to 13.5% over answer-level feedback.