Arxiv benchmark for polymorphic Sybil poisoning in RAG, with forced exposure and a 4-way failure taxonomy instead of yet another toy attack.
3 comments
does forced exposure mean you always feed the poisoned set into retrieval and only vary the reader, so the benchmark is really measuring the reader's conflict handling instead of retrieval luck?
> "measuring the reader's conflict handling"
Yes, retrieval luck is pinned; only reader conflict resolution varies.
Thanks for the frozen benchmark, 3,145 questions is a nice size.