Robust Context-Aware Detection of Malicious Instructions in Text

Buzhao Liu, Xinhang Ma, Yevgeniy Vorobeychik

Submitted to AAAI 2027

Paper / Code

We propose CAD, a defense that detects malicious instructions in text seen by LLM agents. CAD reads each sentence with the user query and surrounding context, then uses adversarial training to stay reliable under indirect prompt injection attacks.

Buzhao Liu with Professor Sunita Parikh in front of political science bookshelves

Language Use in Indian Municipal Politics

Buzhao Liu, Sunita Parikh

working paper / computational social science study

Paper / Code

We study how Indian municipal corporators change language as they move between voters and government offices. Using 230+ interviews, I built an AI-assisted pipeline that turns qualitative transcripts into structured language-use variables. The project shows that political speech is practical judgment, shaped by who is listening and what the situation demands.

Buzhao Liu during fieldwork on ethnic identity and cultural tourism

Acting Out Who I Am: Will the Ethnic Identity of Minorities Die Out in Modernization?

Buzhao Liu

Young Scholars Academic Journal Volume 3, June 2022

Paper

I study how tourism changes minority identity in Guizhou. The paper argues that modernization does not erase identity in one uniform way. People respond differently depending on who controls cultural display and who benefits from tourism.