Ziwei Gu

Hi, I'm Ziwei.

I am a fourth-year Ph.D. student in Computer Science at Harvard University, advised by Dr. Elena Glassman. Currently, my research focuses on augmenting human cognition and efficiency by leveraging large language models (LLMs) and interactive techniques, aiming to ensure that potential AI errors can be easily noticed, judged, and recovered from, a concept I formalize as AI-resiliency.

Before coming to Harvard, I graduated with a bachelors in Mathematics and Computer Science (December 2020) and masters in Computer Science (May 2021) from Cornell University.

Reading documents is too cognitively demanding to do at scale. But what if the documents share a common structure? We design a system that enables comparative close reading of similarly structured documents at scale.

Through an eye-tracking study, we uncover unique gaze patterns when people read with GP-TSM, an LLM-powered reading support tool, and explain how it shapes the reading experience.

Reading is challenging. AI summaries can be helpful but also unreliable. We design an LLM-powered technique that enhances reading efficiency while making readers resilient to errors in summaries.

Large language models (LLMs) are capable of generating multiple responses to a single prompt. In this paper, we explore how to present many LLM responses at once to support sensemaking at scale.

Interaction logs are useful but can be extremely complex. We break down event logs into goal-directed segments to make it easier to understand user workflow.

How do core design elements of debiasing systems shape how people reason about biases? We find distinctive sensemaking patterns for different systems through a think-aloud study.

We present Silva, an interactive tool that utilizes a causal graph linked with quantitative metrics to help people find and reason about sources of biases in datasets and machine learning models.

We trained a Transformer model for open information extraction, uniquely modeled as a sequence-to-sequence transduction task. Our model was competitive with state-of-the-art systems, yet did not rely on other NLP tools.