Exploring LLM biases to manipulate AI search overview

The feed highlights multiple recent IR/LLM papers. The lead paper demonstrates a practical attack surface: a small RL-trained language model can rewrite search snippets (only editing snippet text) to bias LLM-based overview systems into preferring chosen sources; selections are driven by comparative advantages among candidates and this “context poisoning” can yield inaccurate or harmful overviews. Related items survey reasoning-intensive retrieval, argue for a denoising-first RAG paradigm, document reformulation drift and defenses (ranker feedback), introduce benchmarks for instruction-follow-