UserToolBench benchmark shows LLMs fail at personalized tool-use decisions

UserToolBench, proposed by Xuexiong Yin and seven colleagues in arXiv:2608.10042, aims to shift how we evaluate personalization in tool-use LLMs. Most existing benchmarks check if a model can recall user profiles or imitate a user's writing style. But when AI agents act on our behalf—booking flights, sending emails, managing calendars—they need to infer hidden preferences from past interactions, ask for clarification when info is missing, and execute multi-step tool calls that align with user intent. UserToolBench is built from privacy-sanitized real interaction traces and combines structured persona profiles, public API-style tool ecosystems, and long-horizon multi-turn trajectories. It includes 10 user profiles, 36 tool sets, 1,065 turns, and 170 unique tools, with tasks ranging from lack-of-information scenarios to single- and multi-tool coordination.

Experiments with strong tool-use LLMs reveal persistent bottlenecks. Models struggle with multi-tool coordination, fail to infer missing constraints, and lack long-horizon behavioral consistency across turns. Even when outputs sound personalized, the underlying decisions are often wrong. The authors argue that personalization evaluation should move beyond whether output sounds user-specific, and instead ask whether LLMs make correct decisions for the users they represent. This is a critical wake-up call for the agentic AI ecosystem, where delegation to LLMs is growing rapidly. UserToolBench provides a concrete testbed for measuring and improving that trust.