Hanfei Zhu 朱晗飞

Abstract摘要

The conflict between the rapid iteration demand of prototyping and the time-consuming nature of user tests has led researchers to adopt AI methods to identify usability issues — yet these often overlook specified user characteristics and usage contexts. SimUser is an LLM-based tool with a Chain-of-Thought structure and user-modeling method that generates usability feedback by simulating interactions between users and applications, shaped by user characteristics and contextual factors. An empirical study (48 users, 21 designers) validated that, for a smartwatch interface, SimUser generates heuristic usability feedback with similarity ranging 35.7%–100% across user groups and usability categories.原型快速迭代的需求与用户测试的耗时之间的矛盾，促使研究者用 AI 方法识别可用性问题——但这些方法往往忽略了特定的用户特征与使用情境。SimUser 是一款基于大模型、结合思维链（CoT）结构与用户建模方法的工具，通过模拟受用户特征与情境因素影响的“用户—应用”交互来生成可用性反馈。一项实证研究（48 名用户、21 名设计师）验证：在智能手表界面上，SimUser 生成的启发式可用性反馈相似度在不同用户群组与可用性类别下达 35.7%–100%。

Highlights研究亮点

• Combines LLMs, Chain-of-Thought reasoning, and explicit user modeling to simulate different personas — not just a generic user.融合大模型、思维链推理与显式用户建模，模拟不同画像的用户，而非泛化的单一用户。
• Validated with 48 real users and 21 designers; feedback similarity reached 35.7%–100% across groups.经 48 名真实用户与 21 名设计师验证，各群组反馈相似度达 35.7%–100%。
• Grew into an automated interface-evaluation system (~80% coverage) and an industry collaboration with ByteDance.进一步发展为覆盖率约 80% 的界面可用性自动评测系统，并推动了与字节跳动的产业合作。

Abstract摘要

L3 automated driving lets drivers engage in non-driving-related tasks (NDRTs) but challenges the reconstruction of situation awareness (SA). Two driving-simulator studies investigated two peripheral interactions — airflow conveying vehicles' intended behavior, and surround sound conveying road-user information. Study 1 (n=21) explored airflow, surround sound, and their integration; Study 2 (n=30) tested the integrated interaction across NDRT difficulties. Airflow and surround sound each significantly improved SA, integration performed best, and the benefit was greater during hard NDRTs — while also reducing workload and improving user experience.L3 自动驾驶让驾驶员得以从事非驾驶任务（NDRT），却对情境意识（SA）的重建提出挑战。两项驾驶模拟实验研究了两种外围交互——用气流传达车辆意图、用环绕声传达道路使用者信息。研究一（n=21）探究气流、环绕声及其融合；研究二（n=30）测试融合交互在不同 NDRT 难度下的表现。结果表明气流与环绕声单独使用均能显著提升 SA，二者融合效果最佳，且在困难 NDRT 下收益更大——同时降低了负荷、改善了用户体验。

Highlights研究亮点

• Two consecutive driving-simulator studies (n=21 and n=30).两项连续的驾驶模拟实验（n=21、n=30）。
• Airflow encodes vehicle intent; surround sound encodes road-user information — each with distinct advantages.气流编码车辆意图、环绕声编码道路使用者信息，各具独特优势。
• Integrated interaction gives the best SA, cuts perceived workload, and raises trust in the automated system.融合交互带来最佳情境意识，降低主观负荷，并提升对自动系统的信任。

Abstract摘要

As LLMs increasingly generate outputs tailored to specific users, are those outputs genuinely shaped by users or merely articulating user-relevance? Using design as a testbed, we collected concepts across 20 tasks from human designers (n=200), GPT-4o, two open-weight models, and their fine-tuned variants, evaluated by experts (n=1,863). Base LLMs matched humans on usefulness and novelty but scored lower on human-centeredness and produced more similar outputs. Distilling 12,743 HCI papers to fine-tune the open-weight models improved all dimensions — most in human-centeredness — and reduced within-task similarity, yet the models still fell short of human-level diversity.当大模型越来越多地生成“为特定用户量身定制”的输出时，这些输出究竟是真正由用户塑造，还是仅仅在陈述“与用户相关”？我们以设计为试验场，在 20 项任务上收集人类设计师（n=200）、GPT-4o、两个开源模型及其微调版本的方案，并由专家（n=1,863）评分。基础大模型在有用性与新颖性上比肩人类，却在以人为本上得分更低、输出更趋同。蒸馏 12,743 篇 HCI 论文对开源模型微调后，各维度均有提升——以人为本提升最大——任务内相似度下降，但仍未达到人类水平的多样性。

Manuscript under review — full text not yet publicly available.论文正在审稿中，全文暂未公开。

Abstract摘要

LLMs open opportunities to leverage HCI studies in design practice, but designers struggle to extract usable information from papers and to synthesize fragmented cross-domain knowledge. ImpliSpark is a web-based tool that formalizes an artifact's design rationale into a unified Context–Function–Technology–Result representation (90.8% extraction accuracy), then coordinates two specialized LLM modules in a practice–academic–practice workflow to synthesize cross-paper knowledge into context-specific design implications. Two studies (n=21, n=20) found its implications more valid, original, inspiring, and actionable than GPT-4o's, helping participants produce more novel and feasible concepts.大模型为将 HCI 研究用于设计实践创造了机会，但设计师难以从论文中提取可用信息、也难以综合碎片化的跨领域知识。ImpliSpark 是一款 web 工具，将学术制品的设计逻辑形式化为统一的“情境—功能—技术—结果”结构（提取准确率 90.8%），并在“实践—学术—实践”工作流中协调两个专门的大模型模块，跨论文综合出情境化的设计启示。两项用户实验（n=21、n=20）表明，其启示在有效性、新颖性、启发性与可操作性上均优于 GPT-4o，并帮助参与者产出更新颖、更可行的方案。

Manuscript under review — full text not yet publicly available.论文正在审稿中，全文暂未公开。