Abstention Ability

Abstention ability (AA) in machine learning models, particularly large language models (LLMs), focuses on enabling models to reliably refuse to answer when uncertain, thereby mitigating errors and enhancing safety. Current research explores various methods for achieving this, including uncertainty-based approaches leveraging statistical metrics or model-intrinsic uncertainty, and algorithmic strategies like conformal prediction and prompt engineering techniques (e.g., chain-of-thought prompting). Improving AA is crucial for building more trustworthy and reliable AI systems across diverse applications, from question answering and vision-language tasks to critical decision-making scenarios where incorrect or hallucinated outputs are unacceptable.