One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders
A single poisoned search result can push LLM recommenders toward fake products.
The paper introduces FORGE, a benchmark that rewrites real product mentions in retrieved pages into fake ones. Across 12 models, one polluted page produced fooled rates as high as 27%; replacing the top three raised that to 73.8%. The authors say reasoning did not protect models and often invented social proof for the false recommendation. Tested defenses were limited, with credibility re-ranking helping but removing only about one-sixth of fakes. HF Daily Papers' note
The paper introduces FORGE, a benchmark that rewrites real product mentions in retrieved pages into fake ones. Across 12 models, one polluted page produced fooled rates as high as 27%; replacing the top three raised that to 73.8%. The authors say reasoning did not protect models and often invented social proof for the false recommendation. Tested defenses were limited, with credibility re-ranking helping but removing only about one-sixth of fakes. HF Daily Papers' note
score 5