Two-Stage Learning to Defer with Multiple Experts

Large language models (LLMs) have achieved a remarkable performance on diverse tasks across multiple domains, as reported in recent surveys [Wei et al., 2022, Bubeck et al., 2023].