AttriBoT: A Bag of Tricks for Efficiently Approximating Leave-One-Out Context Attribution

How and why LLM generates such output? RAG: retrieved document might be helpful, but no insight on which part of the augmented context influenced the model's generation. Corroborative: how the statement is supported by the context. Contributive: how the context contribute to the response generation. More formally: quantify the influence of each span of text in an LLM's context on its generated output.