Prism: Towards Lowering User Cognitive Load in LLMs via Complex Intent Understanding

TL;DR

Prism tackles the challenge of complex intent understanding in web-mediated LLM interactions by explicitly modeling dependencies among clarification questions through a four-module framework guided by Cognitive Load Theory. It introduces a Complex Intent Decomposition (CID) dataset to hierarchically structure intents, a Logical Clarification Generation module to sequence clarifications by dependencies, an Intent-Aware Reward with Monte Carlo Forward Sampling to generate scalable training data, and a Self-Evolved Intent Tuning loop to iteratively improve the model. Across clarification interaction, intent execution, and cognitive load, Prism achieves state-of-the-art logical consistency (e.g., reducing logical conflicts to 11.5%), higher user satisfaction, and significantly reduced task completion times. The approach generalizes across backbones and includes releasing extensive CID data and code, potentially enabling broader adoption of cognitively efficient human–LLM collaboration in complex, real-world tasks.

Abstract

Large Language Models are rapidly emerging as web-native interfaces to social platforms. On the social web, users frequently have ambiguous and dynamic goals, making complex intent understanding-rather than single-turn execution-the cornerstone of effective human-LLM collaboration. Existing approaches attempt to clarify user intents through sequential or parallel questioning, yet they fall short of addressing the core challenge: modeling the logical dependencies among clarification questions. Inspired by the Cognitive Load Theory, we propose Prism, a novel framework for complex intent understanding that enables logically coherent and efficient intent clarification. Prism comprises four tailored modules: a complex intent decomposition module, which decomposes user intents into smaller, well-structured elements and identifies logical dependencies among them; a logical clarification generation module, which organizes clarification questions based on these dependencies to ensure coherent, low-friction interactions; an intent-aware reward module, which evaluates the quality of clarification trajectories via an intent-aware reward function and leverages Monte Carlo Sample to simulate user-LLM interactions for large-scale,high-quality training data generation; and a self-evolved intent tuning module, which iteratively refines the LLM's logical clarification capability through data-driven feedback and optimization. Prism consistently outperforms existing approaches across clarification interactions, intent execution, and cognitive load benchmarks. It achieves stateof-the-art logical consistency, reduces logical conflicts to 11.5%, increases user satisfaction by 14.4%, and decreases task completion time by 34.8%. All data and code are released.

Figures (5)

• Figure 1: Comparison of user intent clarification: (a) Q&A Pair Interactions, (b) Table-based Interactions, and (c) Logical Interactions Design.
• Figure 2: The Prism Framework: Given a user instruction, the policy model first hierarchically decomposes complex intents with retrieving CID dataset. It then organizes clarification questions and enables logical interaction through interactive table. Intent-aware rewards (IRs) are estimated via Monte Carlo sampling. Finally, self-evolved intent tuning iteratively enhances the training data quality and improves LLM's capacity for complex intent understanding.
• Figure 3: Our cognitive-load study includes 20 participants interacting with four anonymized models (Mistral-Interact, ITIU, CollabLLM, Prism). Each participant is randomly assigned tasks from TIN, IN3, or ABP. We measure (a) user spent time, (b) conversation token count. Participants rate (c) overall interaction experience, with (d) additional assessments every two turns.
• Figure 4: Trend of absolute power spectral density (PSD) across the all band in the cognitive-load experiment. Each point on the blue solid line represents the mean PSD value of valid frequency points within the total frequency band at a given moment, and the shaded area indicates the standard deviation. Higher mean and variance of PSD indicate more active brain activity and higher cognitive load.
• Figure 5: Performance comparison of Mistral-Interact and ITIU across simple and complex intent scenarios. From the simple intent scenario to the complex intent scenario, the increments of Mistral-Interact and ITIU decreased from 18.8% and 24.8% to 6.2% and 12.4% (ref. Section \ref{['sec:Evaluation on Intent Execution']}).