From Tokens to Semantics: Leveraging Complementary Signals for Hallucination Detection in Black-Box LLMs

To address the challenge of hallucination detection in black-box large language models (LLMs), recent research proposes using semantic entropy and token-based uncertainty as complementary signals. The study extends the TopK method to aggregate token-level signals in sampling responses and proposes a hybrid CoCoA method that combines target response uncertainty with semantic differences, while introducing two supervised learning strategies: Gated and Stacked. In seven tests covering five public benchmarks (including multimodal handwritten check extraction) and two self-built benchmarks (financial abstracts and long-text question answering), the researchers used four language models for validation. The results show that the Stacked method performs best in approximately half of the cases; TopK and CoCoA are competitive under unsupervised labeling, but threshold calibration requires fine-tuning. The study further analyzes the performance under different false-positive rate budgets (from 1% to 15%), the impact of generation and calibration choices, and the differences in dataset characteristics, confirming that there is no universally optimal detection method.