On early detection of hallucinations in factual question answering
model can provide hints that a response will contain hallucinations. Specifically, we probe LLMs at 1) the inputs via integrated gradients based token attribution, 2) the outputs via the softmax probabilities …
