Beyond Turns: Fragment-Response Alignment for Evolving Intent and Traceable Review in LLM Chat
Xichun Xu (Xi'an Jiaotong-Liverpool University), Ruofei Du (Xi’an Jiaotong-Liverpool University), Jiaqi Zheng (Xi'an Jiaotong-Liverpool University), Qingping Wang (Xi'an Jiaotong-Liverpool University), Lijie Yao (Xi'an Jiaotong-Liverpool University), Liu YANG (Xi'an Jiaotong-Liverpool University)
Links
Sign in to access the Paper PDF.
Abstract
LLM chat interfaces are increasingly used for complex tasks, yet remain tethered to rigid, chronological transcripts. This structure creates a "traceability gap," hindering users from mapping evolving intent to fragmented model outputs. In this work, we introduce Fragment-Response Alignment, a conceptual framework that maps user intent fragments to specific AI response units. We report findings from a formative study of 16 participants, which reveal three core phenomena: intent is continuously evolving, interaction functions as a traceability reconstruction task, and users naturally employ fragment-based mental models.