Fine-grained Hallucination in Numerical Values Despite Citation and Judger Implementation

[YeLyea]

Hello,

I see that your implementation of citation and judge uses prompts to control credible output. However, there's a problem: large models are inherently susceptible to hallucinations, especially with numbers. The answer might be generally correct, but the numbers might be secretly altered (for example, changing "87%" to "85%" or "2021" to "2022"). Has the author considered and addressed this issue?

Thank you for your consideration!