Takyoung Kim, Jinseok Nam, Chandrayee Basu, Xing Fan, Chengyuan Ma, Heng Ji, Gokhan Tur, Dilek Hakkani-Tür
ReIn is a method for conversational agents to recover from errors by integrating an external reasoning module without altering the model's parameters or prompts.
Conversational agents often struggle with unexpected user errors, which can disrupt the dialogue flow. Instead of trying to prevent these errors, this research focuses on how agents can recover from them effectively. The proposed method, Reasoning Inception (ReIn), uses an external module to detect errors and suggest recovery plans, integrating these into the agent's decision-making process without changing its core settings. This approach allows agents to handle errors better and improve task success rates, even with new types of errors, without the need for costly model adjustments.
