Researchers have developed a new benchmark to study how well-retrieved passages help large language models (LLMs) generate accurate answers. They argue that the usefulness of these passages depends on the specific LLM being used, due to differences in knowledge and reasoning abilities. The team created a dataset of 'utilitarian passages' for four different LLMs and found that each model performs best with its own evidence, while evidence optimized for other models is often suboptimal. This suggests that current methods for selecting useful passages may not be effective for all models. The study's findings highlight the need for more tailored approaches to retrieval-augmented generation. --- Why it matters: This research matters because it sheds light on the limitations of current methods for selecting useful evidence for large language models, which are critical components of many applications, including chatbots and virtual assistants. Source: https://arxiv.org/abs/2510.11358