Adaptive Conformal Semantic Entropy for LLM Uncertainty
A novel approach called Adaptive Conformal Semantic Entropy (ACSE) has been introduced by researchers to gauge uncertainty in the outputs of large language models. This technique involves clustering semantic entropy from various responses and dynamically modifying scores according to semantic characteristics. Additionally, ACSE employs conformal calibration to ensure statistical reliability, effectively tackling issues of overconfidence and hallucination in applications where safety is paramount.
Key facts
- LLMs exhibit overconfidence, especially when hallucinating.
- Existing uncertainty quantification methods prioritize lexical or probabilistic measures.
- ACSE estimates prompt-level uncertainty by measuring semantic dispersion.
- The method clusters semantic entropy from multiple diverse responses.
- ACSE adaptively adjusts uncertainty scores based on cluster semantic features.
- Conformal calibration ensures statistical reliability for accept/abstain decisions.
- The approach targets safety-critical settings.
- The paper is available on arXiv with ID 2605.04295.
Entities
Institutions
- arXiv