Integrated Gradients for Llama2 May Produce Unstable Explanations for Long Contexts

[WYT8506]

❓ Questions and Help

Hi Authors,

When I apply LayerIntegratedGradients on Llama for long context QA, the feature attribution seems not meaningful. For example, when I copy a news from wiki page and ask a question about that news:
"Kingdom of the Planet of the Apes is a 2024 American science fiction action film directed by Wes Ball and written by Josh Friedman. A standalone sequel to War for the Planet of the Apes (2017), it is the fourth installment in the Planet of the Apes reboot franchise and the tenth film overall. It stars Owen Teague in the lead role alongside Freya Allan, Kevin Durand, Peter Macon, and William H. Macy. The film takes place 300 years after the events of War and follows a young chimpanzee named Noa, who embarks on a journey alongside a human woman named Mae to determine the future for apes and humans alike.
Development on a new Planet of the Apes film began in April 2019, following Disney's acquisition of 20th Century Fox, with Ball attached as writer and director that December. Much of the script was written during the COVID-19 pandemic in 2020, with casting commencing in June 2022, following the script's completion..... When is the development on a new Planet of the Apes film began? Give a short answer"
And I set the target to the correct answer, which is "April 2019". LayerIntegratedGradients would not mark "April 2019" in the context as the most important. Could you provide more information for when LayerIntegratedGradients would work well/or not well for long inputs?