Brief

Researchers have developed PromptShift-CRC, a novel drift-aware conformal risk control method designed for foundation models facing evolving prompts and domain shifts. This method addresses the limitations of static calibration by embedding prompts and responses, dynamically adjusting weights for calibration examples based on relevance and recency, and updating risk levels in real-time. Evaluations on synthetic and public benchmarks demonstrate PromptShift-CRC's effectiveness in maintaining risk control where static methods fail, particularly in applications like question answering and summarization factuality. AI