JumpStarter: Human-AI Planning with Task-Structured Context Curation
Xuanming Zhang, Sitong Wang, Jenny Ma, Alyssa Hwang, Zhou Yu, Lydia B. Chilton
TL;DR
JumpStarter tackles the challenge of planning with large language models by introducing task-structured context curation, which organizes complex goals into a hierarchical subtask tree and localizes context management at decision points. The approach combines context elicitation, selection, and reuse to generate actionable drafts more personalized and coherent than flat long-context prompting. Empirical results show a 16% improvement in plan quality over ablations and a 79% higher quality of plans compared with GPT-4o via ChatGPT, along with reduced task load in user studies. The work demonstrates that structured, task-centered context management enhances human-AI collaboration for goal-driven planning and points toward more transparent, modular, and context-aware LLM-based assistants in real-world workflows.
Abstract
Human-AI planning for complex goals remains challenging with current large language models (LLMs), which rely on linear chat histories and simplistic memory mechanisms. Despite advances in long-context prompting, users still manually manage information, leading to a high cognitive burden. Hence, we propose JumpStarter, a system that enables LLMs to collaborate with humans on complex goals by dynamically decomposing tasks to help users manage context. We specifically introduce task-structured context curation, a novel framework that breaks down a user's goal into a hierarchy of actionable subtasks, and scopes context to localized decision points, enabling finer-grained personalization and reuse. The framework is realized through three core mechanisms: context elicitation, selection, and reuse. We demonstrate that task-structured context curation significantly improves plan quality by 16% over ablations. Our user study shows that JumpStarter helped users generate plans with 79% higher quality compared to ChatGPT.