Abstract
Large Language Models (LLMs)-based agents have made impressive progress in reasoning and tool use, enabling them to solve complex tasks. However, their ability to proactively collaborate with users, especially when goals are vague, evolving, or indirectly expressed, remains underexplored. To address this gap, we introduce UserBench, a user-centric benchmark designed to evaluate agents in multi-turn, preference-driven interactions. UserBench features simulated users who start with underspecified goals and reveal preferences incrementally, requiring agents to proactively clarify intent and make grounded decisions with tools. Our evaluation of leading open- and closed-source LLMs reveals a significant disconnect between task completion and user alignment. For instance, models provide answers that fully align with all user intents only 20% of the time on average, and even the most advanced models uncover fewer than 30% of all user preferences through active interaction. These results highlight the challenges of building agents that are not just capable task executors, but true collaborative partners. UserBench offers an interactive environment to measure and advance this critical capability.
Introduction
While LLM agents have become proficient at using tools and executing code, they remain surprisingly poor at understanding the messy reality of human communication. Most current agent benchmarks focus on "task completion"—whether the agent can call the right API or write the correct snippet of Python. However, in the real world, users rarely provide perfect instructions. Human communication is inherently a "joint activity" where meaning is co-constructed.
UserBench addresses three core traits of user instructions that current agents struggle with:
- Underspecification: Users often start with vague or incomplete goals.
- Incrementality: Intent emerges and changes over the course of a conversation.
- Indirectness: Users may soften or obscure their true desires for social or strategic reasons.
Existing benchmarks often overlook the user's role, assuming fully specified prompts. UserBench fills this gap by creating a sandbox where an agent must proactively clarify intent and align with a simulated user’s hidden, nuanced preferences to succeed.
Methods and Framework
UserBench is built on the standard OpenAI Gymnasium framework, providing a modular and reproducible environment for agent evaluation. The framework simulates a complex travel-planning scenario where agents must navigate user preferences across five aspects: flights, hotels, apartments, car rentals, and restaurants.
The core components include:
- Simulated Oracle User: An LLM-driven user that holds a set of "ground-truth" preferences. It only reveals these preferences if the agent asks the right questions or after a specific number of turns.
- Interactive Loop: Agents can perform three types of actions:
search(querying a mock database),action(communicating with the user), andanswer(submitting a final recommendation). - Preference Curation: Over 100 distinct travel preferences were curated and paired with "implicit expressions" to test if agents can read between the lines.
- Scoring Mechanism: A normalized score based on how well the agent's final selection matches the user's hidden constraints.
Evaluation and Key Results
We evaluated leading models including GPT-4o, Claude 3.5 Sonnet, and DeepSeek-V3 across 417 high-quality data points categorized into Easy, Medium, and Hard tiers based on preference complexity.
Our evaluation revealed several sobering truths about the current state of user-centric agents:
- The "Alignment-Execution" Disconnect: There is a massive gap between an agent's ability to use a tool and its ability to satisfy a user. Models often execute perfect searches for the wrong things because they failed to clarify the user's intent. On average, models provide answers that fully align with all user intents only 20% of the time.
- The Guessing Trap: Smaller or less capable models tend to "guess" an answer early to appear efficient, rather than investing turns in clarification. This leads to fast but incorrect results.
- Poor Elicitation: Even the most advanced models uncover fewer than 30% of hidden user preferences through active questioning.
- Preference Overload: Agents reason more effectively when preferences are spread across different aspects rather than concentrated in one. Concentrated preferences (e.g., four specific requirements for a single hotel) seem to "overload" the model's local reasoning capabilities.
Conclusion
UserBench highlights that even the most advanced models struggle to act as true "collaborative partners." While GPT-4o and Claude 3.5 Sonnet lead the field, there is significant room for improvement in proactive intent clarification and dialogue planning. Future work will involve using the UserBench training set (comprising 2,651 scenarios) to develop "user-aware" agents via Reinforcement Learning, teaching them specifically when to ask questions versus when to act. UserBench is publicly available to encourage the community to move toward more human-centric AI design.
Summary updated