LLM biases exploited to manipulate AI search overviews

A recent study published on arXiv (2605.00012) examines the biases present in large language models (LLMs) that generate overviews of search results, referred to as LLM Overview systems. This research delves into how these biases can be leveraged to influence the selection of sources. The authors employ reinforcement learning to train a smaller language model capable of rewriting search snippets, thereby enhancing their chances of being favored by an LLM Overview. The experimental design purposefully limits the policy to showcase the potential for such manipulation. The findings underscore the vulnerabilities of AI-driven search systems that depend on LLMs for both source selection and answer formulation.