ThoughtTrace: Understanding User Thoughts in Real-World LLM Interactions
Abstract
AI models are now widely deployed to interact with humans. While many datasets, analyses, and methods focus on what people say during these interactions, understanding and modeling what people actually think remains a largely unexplored dimension of the human-AI interaction experience.
User thoughts are the unspoken context behind each message: the underlying intent and desired outcome, motivations and stakes, constraints and preferences shaping the request, assumptions about the situation and the assistant's capabilities, as well as interpretations of and reactions to the assistant's reply. Capturing these thoughts provides a new lens on assistance by exposing the mental context that shapes user messages. This paper describes the construction of WildThoughts, a framework and dataset for capturing and analyzing user thoughts during real-world human-AI interactions. WildThoughts is a rich corpus of first-person cognitive traces that illuminate the lived experience of interacting with AI systems. We analyze these high-quality interactions which encompass diverse users performing real-world tasks. WildThoughts unlocks deeper cognitive understanding, including improved Theory of Mind for understanding user states, and provides the foundation for user-centric model assessment beyond surface-level response quality.
User thoughts are the unspoken context behind each message: the underlying intent and desired outcome, motivations and stakes, constraints and preferences shaping the request, assumptions about the situation and the assistant's capabilities, as well as interpretations of and reactions to the assistant's reply. Capturing these thoughts provides a new lens on assistance by exposing the mental context that shapes user messages. This paper describes the construction of WildThoughts, a framework and dataset for capturing and analyzing user thoughts during real-world human-AI interactions. WildThoughts is a rich corpus of first-person cognitive traces that illuminate the lived experience of interacting with AI systems. We analyze these high-quality interactions which encompass diverse users performing real-world tasks. WildThoughts unlocks deeper cognitive understanding, including improved Theory of Mind for understanding user states, and provides the foundation for user-centric model assessment beyond surface-level response quality.
