Ukyo Honda, Tatsushi Oka
The study introduces X$^2$-ICL, an improved in-context learning method that uses explanations for all possible labels to enhance robustness in language models.
In-context learning (ICL) helps large language models make predictions by using examples, but it often fails when faced with new, different data. A new method, X$^2$-ICL, builds on a previous improvement called X-ICL, which uses explanations to help models understand why an answer is correct. X$^2$-ICL takes this a step further by exploring explanations for all potential answers, leading to better and more reliable predictions. Tests on various language tasks show that X$^2$-ICL is more robust when dealing with unfamiliar data compared to older methods.
