UserHarness: Harnessing User Minds for Stronger Agent Theory-of-Mind

Researchers introduce UserHarness, a framework that improves AI agents' Theory-of-Mind capabilities by explicitly reconstructing user mental states rather than modeling behavior indirectly. The approach achieves 95.94% accuracy across five benchmarks, demonstrating significant improvements over existing methods and offering a foundation for building more adaptive AI assistants.

UserHarness addresses a fundamental challenge in AI development: enabling agents to understand user beliefs, intentions, and reasoning processes. Traditional Theory-of-Mind approaches treat user understanding as a black-box prediction problem, but this new framework reframes it as explicit mental state reconstruction. By decomposing what users observe, believe, intend, and subsequently do, the system creates a more interpretable and accurate model of human cognition.

The breakthrough emerges from a recognition that user behavior flows from a causal chain: observations shape beliefs, beliefs inform intentions, and intentions drive actions that alter the environment. This mirrors how cognitive science understands human reasoning. Rather than attempting to predict actions directly, UserHarness traces this causal sequence, enabling nested reasoning about what others believe others believe—a critical capability for complex social interactions.

The performance improvements are substantial. Achieving 95.94% accuracy represents a 15% relative improvement over competing inference methods and 20% relative improvement over prompt-based approaches. These gains suggest the framework captures genuine structural insights about how human minds process information. For AI developers, this offers a more robust foundation for building assistants that must collaborate with users, anticipate needs, and respond contextually to misunderstandings.

Looking forward, UserHarness positions user understanding as a central design principle rather than an afterthought. As AI agents move from single-turn interactions to extended collaborations requiring genuine mutual understanding, this explicit mental-state approach becomes increasingly valuable. The framework's generalizability across benchmarks hints at its potential applicability across diverse AI applications.

• →UserHarness achieves 95.94% macro accuracy by explicitly reconstructing user mental states rather than predicting behavior indirectly.
• →The framework decomposes user cognition into observations, beliefs, intentions, and actions, enabling nested social reasoning.
• →Performance improvements of 15-20% relative gains over existing methods demonstrate the effectiveness of causal-chain reasoning about user minds.
• →The approach provides a more interpretable foundation for building adaptive AI assistants that can understand and respond to user reasoning.
• →Explicit mental-state modeling enables better handling of complex social interactions requiring reasoning about what others believe or intend.