Wei-Hao Chen

Computer Science PhD student @ Purdue University Interested in Human-Computer Interaction, Human-AI Interaction, and Data Science.

[Wei-Hao Chen @_weihao_chen_]

Thrilled to share that our paper "MUSE: An Interactive Meta-Agent for Understanding and Steering LLM-powered Data Science Systems" has been accepted to UIST 2026! 🎉 (1/8)
Preprint: https://t.co/CpPCkjWj11

[Wei-Hao Chen @_weihao_chen_]

In a between-subjects study (N=15), MUSE raised task success from 50% to 90%, cut completion time by 35%, and significantly boosted users' confidence and trust in the agent's outputs. (8/8)

[Wei-Hao Chen @_weihao_chen_]

Thrilled to share that our paper "MUSE: An Interactive Meta-Agent for Understanding and Steering LLM-powered Data Science Systems" has been accepted to UIST 2026! 🎉 (1/8)
Preprint: https://t.co/CpPCkjWj11

[Wei-Hao Chen @_weihao_chen_]

MUSE also proactively monitors execution and surfaces localized warnings on suspicious steps, letting users inspect what triggered the warning, understand why it matters, and fix it through the agent, all without digging through the detailed log. (7/8)

[Wei-Hao Chen @_weihao_chen_]

Users can adjust the semantic granularity of the log summary based on their need, and verify any issue through contextualized, agentic verification instead of manual verification. (6/8)

[Wei-Hao Chen @_weihao_chen_]

Inspired by the information-seeking mantra: "overview first, zoom and filter, details on demand." MUSE restructures agent raw logs into multi-level explanations from high-level overviews to low-level traces, spanning workflow summaries, step lists, and execution logs. (5/8)

[Wei-Hao Chen @_weihao_chen_]

This was extremely difficult for non-experts unfamiliar with agent workflows. Many simply started over, or gave up on using agents. (4/8)

[Wei-Hao Chen @_weihao_chen_]

Our formative study revealed the real pain: it wasn't just reading over the logs, but figuring out which step went wrong, verifying that it was indeed wrong, and translating their insights and findings into actionable feedback. (3/8)

[Wei-Hao Chen @_weihao_chen_]

Agentic data science systems can now complete complex workflows from just a few natural language prompts. But when something goes wrong, users are left sifting through walls of raw agent logs, including code, thoughts, plans, and outputs, to figure out what happened. (2/8)

[Wei-Hao Chen @_weihao_chen_]

We conducted a within-subjects user study (N=38) and found that Dango helped users complete tasks faster (45%) and with fewer hallucinations (72%). We observed that users actively engaged with multiple-choice clarification questions, which helped them clarify their intent. (7/7)

[Wei-Hao Chen @_weihao_chen_]

I will present our paper "Dango: A Mixed-Initiative Data Wrangling System using Large Language Model" at 5:08 pm today in the Annex Hall F206. If you are around, please come and see my talk! (1/7)

Paper: https://t.co/LmqEozsZz0
Code: https://t.co/3xEHwT8qLv

#CHI2025

[Wei-Hao Chen @_weihao_chen_]

We propose Dango, an LLM-based data wrangling system that supports mixed-initiative interaction. Users can proactively point out incorrect data wrangling steps and provide feedback, and Dango can also proactively ask clarification questions when confused about user intent. (6/7)

[Wei-Hao Chen @_weihao_chen_]

To solve these usability issues, we draw inspiration from the mutual grounding theory in communication. In human communication, people build understanding by formulating their utterances, asking clarification questions, and confirming understanding through acknowledgments. (5/7)

[Wei-Hao Chen @_weihao_chen_]

For instance, PBD systems often support only single-table operations, are difficult for users to fix errors, and struggle to handle complex tasks. LLM-based systems can misinterpret user intent, hallucinate, and are challenging for non-experts to prompt effectively. (4/7)

[Wei-Hao Chen @_weihao_chen_]

Over the last decades, many AI systems have been proposed to help users clean data, such as Programming by Demonstration (PBD) and more recently LLM-based systems. However, we found several limitations. (3/7)

[Wei-Hao Chen @_weihao_chen_]

Data wrangling is time-consuming and complex. A recent study shows that data cleaning can take up to 80% of the time spent on data work, especially for non-experts. They need to deal with various data issues, such as missing values, duplicates, and inconsistent formats. (2/7)

[Wei-Hao Chen @_weihao_chen_]

@ryanyen22 Interesting work!

[Wei-Hao Chen @_weihao_chen_]

@jiayouwyhit Very interesting work!

[Wei-Hao Chen @_weihao_chen_]

@bryanhaoenwang https://t.co/t6M85p3WiH
I found this upcoming work to be relevant. It utilized free-form sketching to shape the code. The paper referred to this as "code-shaping," but I believe the concept is quite similar to the idea of vibe coding.

[Wei-Hao Chen @_weihao_chen_]

@yuwen_lu_ lol

[Renwen (Alice) Zhang @renwenzhang]

Our study led by @zichengstephy shows that prosocial behavior towards AI chatbots can improve well-being.