Tell Me More! Towards Implicit User Intention Understanding of Language Model Driven Agents (2402.09205v2)

Abstract: Current LLM-driven agents often lack mechanisms for effective user participation, which is crucial given the vagueness commonly found in user instructions. Although adept at devising strategies and performing tasks, these agents struggle with seeking clarification and grasping precise user intentions. To bridge this gap, we introduce Intention-in-Interaction (IN3), a novel benchmark designed to inspect users' implicit intentions through explicit queries. Next, we propose the incorporation of model experts as the upstream in agent designs to enhance user-agent interaction. Employing IN3, we empirically train Mistral-Interact, a powerful model that proactively assesses task vagueness, inquires user intentions, and refines them into actionable goals before starting downstream agent task execution. Integrating it into the XAgent framework, we comprehensively evaluate the enhanced agent system regarding user instruction understanding and execution, revealing that our approach notably excels at identifying vague user tasks, recovering and summarizing critical missing information, setting precise and necessary agent execution goals, and minimizing redundant tool usage, thus boosting overall efficiency. All the data and codes are released.

• The paper introduces the Intention-in-Interaction (IN3) benchmark to quantitatively assess how well agents discern vague user instructions.
• The paper presents Mistral-Interact, an expert model that actively queries users to extract hidden intentions before task execution.
• Empirical results demonstrate that integrating Mistral-Interact reduces execution errors and improves overall task efficiency.

Enhancing LLM-Driven Agents with Implicit User Intention Understanding

Introduction

LLM-driven agents have significantly advanced in executing tasks directly from user instructions. However, a notable challenge persists in these agents' inability to effectively solicit user participation, especially when instructions are vague. This limitation frequently leads to "fake success" instances where the outcome superficially meets the instruction but misses the user's true intention. Addressing this, the paper proposes a new concept, Intention-in-Interaction (IN3), designed to aid in understanding users’ implicit intentions through structured interaction, thus paving the way for more effective task execution.

User-Agent Interaction Gap

Despite the remarkable capabilities displayed by state-of-the-art LLMs in text, code generation, and logical reasoning, their application in agent systems often fails to account for the nuanced and varied intentions different users might have. This oversight limits the robustness and efficiency of the agent, as it cannot discern the actual intention behind vague instructions or engage the user effectively to clarify. Current benchmarks for assessing agent designs do not consider the importance of clarifying user intentions, leading to a critical gap in the evaluation methodology.

Intention-in-Interaction Benchmark

To address this gap, the paper introduces the Intention-in-Interaction (IN3) benchmark. IN3 is designed with a user-centric perspective, providing a structured framework that focuses on quantitively measuring how well an agent can discern task vagueness and interact with the user to uncover hidden intentions. The benchmark includes a wide range of tasks across various categories, each annotated with details regarding its vagueness and missing critical details, simulating real-world scenarios where users might not provide complete instructions.

Enriching Agent-User Interaction

Building on the IN3 benchmark, the research innovates further by introducing an expert model designed specifically for interacting with users to extract those implicit intentions before executing any tasks. Named Mistral-Interact, this model is trained on simulated dialogues and employs strategies like explicit initial thought, querying with options, and accommodating diverse user tones to effectively communicate with users. This model differentiates itself by actively questioning to fill in the gaps of user instructions, significantly enhancing the clarity of tasks before they are executed by the agent.

Empirical Validation and Implications

Extensive experiments validate the effectiveness of incorporating Mistral-Interact into the XAgent framework. Compared to baseline agents, the enhanced agent system demonstrates superior performance in identifying and clarifying vague tasks, reducing unnecessary or overly general task components, and streamlining tool usage during task execution. These improvements underline the practical benefits of integrating specialized interaction expertise into agent systems, suggesting a promising direction for future developments in agent design.

Future Directions

The paper offers a novel approach to enhancing user-agent interaction through explicit intention understanding, represented by the development and implementation of Mistral-Interact. Despite this progress, further research is necessary to explore the integration of user interactions during agent task execution, expand and refine metrics for assessing interaction quality, and consider the usage of LLMs for simulating realistic user-agent dialogues. These areas hold potential for significantly advancing the field of LLM-driven agent systems, contributing to more personalized, efficient, and user-aligned task execution.

Conclusion

The paper presents a comprehensive approach towards closing the user-agent interaction gap in task execution systems, highlighted by the introduction of the IN3 benchmark and the integration of Mistral-Interact into the agent design. The empirical success of these contributions signifies a critical step towards realizing fully functional, user-centric agent systems capable of navigating the complexities of real-world tasks and user intentions. This work lays a foundation for future explorations into enhancing the interaction between users and AI agents, ultimately contributing to the development of more intuitive, effective, and efficient AI-driven solutions.