BESPOKE: Benchmark for Search-Augmented Large Language Model Personalization via Diagnostic Feedback

Search-augmented large language models (LLMs) remain insufficient for fully addressing diverse user needs, which requires recognizing how the same query can reflect different intents across users and delivering information in preferred forms. While recent systems such as ChatGPT and Gemini attempt personalization by leveraging user histories, systematic evaluation of such personalization is under-explored. To address this gap, we propose BESPOKE, the realistic benchmark for evaluating personalization in search-augmented LLMs. BESPOKE is designed to be both realistic, by collecting authentic chat and search histories directly from humans, and diagnostic, by pairing responses with fine-grained preference scores and feedback. The benchmark is constructed through long-term, deeply engaged human annotation, where human annotators contributed their own histories, authored queries with detailed information needs, and evaluated responses with scores and diagnostic feedback. Leveraging BESPOKE, we conduct systematic analyses that reveal key requirements for effective personalization in information-seeking tasks, providing a foundation for fine-grained evaluation of personalized search-augmented LLMs. Our code and data are available at https://github. com/augustinLib/BESPOKE.

When people use AI to search for information, the same question can mean very different things depending on who is asking. In our paper, two users ask the identical question, "What are the main differences between EV and hybrid vehicles?" One user, based on their past activity, really wants to know about environmental impact, while the other cares about driving performance. They also prefer different formats, such as a flowing explanation versus a short bullet summary. Today's AI search assistants try to handle this by looking at a user's past conversations and searches, but there has been no reliable way to measure how well they actually succeed. To address this, we built BESPOKE, a benchmark for testing how well AI search assistants tailor answers to individual users. It is based on real people: volunteers shared their own genuine chat and search histories, wrote questions reflecting their true needs, and judged the AI's answers with detailed scores and feedback. Using BESPOKE, we found that current systems can tailor answers well when a user's needs are spelled out clearly, but struggle to infer those needs from messy real-world history.