Token Level Uncertainty

Token-level uncertainty quantification in language models aims to assess the reliability of individual word predictions, addressing issues like hallucinations and inaccuracies in generated text. Current research focuses on developing methods to effectively leverage this token-level information for improved model performance, exploring techniques such as learned deferral rules, claim-conditioned probability, and adaptive temperature sampling within various architectures including autoregressive and non-autoregressive models. This work is significant because it enhances the trustworthiness and efficiency of large language models, leading to more reliable applications in diverse fields like machine translation, fact-checking, and code generation.