About

I am a senior at Washington University in St. Louis, majoring in Computer Science and Global Studies. I work with Prof. Chongjie Zhang in the Machine Intelligence Group and Prof. Yevgeniy Vorobeychik on LLM agent security. Previously, I worked with Prof. Sunita Parikh on language and power in Indian politics.

I am broadly interested in trustworthy and embodied AI. My research goal is to build intelligent agents that can learn, reason, and act reliably in the real world. My current work spans robotics, recursive self-improvement, LLM agent security, and multi-agent reinforcement learning.

I also see myself as a global citizen. Fieldwork in Mexico, Italy, and China has shaped how I think about AI across cultures, needs, and values. With this global perspective, I aim to push the capabilities of AI while making its behavior more reliable.

Education

Research

* denotes equal contribution

RoboAuto programs running on a real WidowX arm (sorting balls, stacking dishes, erasing a whiteboard) and in simulation (bimanual microwave, stove and moka pot, two-arm handover)
RoboAuto: Autonomous Robot Program Search through Multi-Level Physical Understanding.

paper website poster

Coding agents write and debug robot programs but do not understand the physical world. RoboAuto gives them physical understanding at three levels: a hierarchical knowledge base of object affordances, spatial relations, and action effects; scene grounding that measures task-relevant quantities and moves the camera when needed; and staged search that runs its own experiments to diagnose failures. RoboAuto outperforms prior methods on tasks from LIBERO-PRO, LIBERO-PRO Long, LIBERO-Plus, and Robosuite and on three long-horizon real-robot tasks. Its knowledge also transfers to RoboTwin 2.0 tasks on a bimanual robot.

CAD workflow: an LLM agent reads an email containing an injected instruction; CAD embeds each sentence with the user query and surrounding context, flags the malicious sentence, and the agent completes the task safely
Robust Context-Aware Detection of Malicious Instructions in Text.

paper code

LLM agents that read emails, web pages, and tool outputs can be hijacked by instructions hidden in that text. We introduce CAD, a detector that reads every sentence together with the user's query and its surrounding context, and harden it with adversarial training against paraphrasing attacks. CAD outperforms state-of-the-art defenses on AgentDojo and AgentDyn under both static and adaptive attacks while preserving the agent's utility.

SMAX StarCraft benchmark: two teams of stalkers and zealots in combat
MATCHA: Bridging the Gap between Sample and Wall-Clock Efficiency in Model-Based Multi-Agent Reinforcement Learning.
Buzhao Liu with Professor Sunita Parikh in front of political science bookshelves
Dialects of Dominance: Navigating Hindi Supremacy and Regional Language Dominance in Indian Municipal Politics.

In India, the language a politician chooses is a statement about status. Drawing on interviews with 400+ municipal councilors across North India and Maharashtra, we combine computational sentiment analysis with structural thematic coding. Attachment to local languages is strong, but politicians shift to high-status standard forms to assert power, especially across gender and caste lines.

LLM vs human scores across 92 countries: the cultural prompt (Pearson 0.805) tracks human survey scores far better than the simple prompt (0.586)
Scalable XAI Evaluation Using LLM-Simulated User Populations.

poster

Evaluating AI explanations usually requires costly human studies. We test whether LLMs given real user profiles can stand in for human participants. Simulated users reproduce the direction of human behavior but exaggerate its size; asking them to consider cultural background raises their correlation with human survey responses across 92 countries from 0.59 to 0.81.

Atari benchmark games: Pong, Breakout, Ms. Pac-Man, Seaquest, Bank Heist, Crazy Climber, Space Invaders, and Boxing
Search-Based Model-Based RL for Single-Agent Control.

We are bringing MATCHA's GPU-native framework from multi-agent to single-agent settings.

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?

paper

Does tourism erase minority identity? Based on fieldwork in Guizhou, I argue that modernization does not erase identity in one uniform way: how communities respond depends on who controls cultural display and who benefits from tourism.

Projects

Looping robotics task visualization for multi-agent reinforcement learning evaluation
Bimanual Multi-Agent Robot Learning Benchmark.

blog code

Can multi-agent RL teach two robot arms to coordinate? I designed a new bimanual manipulation task in ManiSkill and trained PPO, MAPPO, HAPPO, and SAC from scratch, comparing single-agent and multi-agent policies on my task and two existing ManiSkill tasks.

Pages from Lists of Chinese Passenger Arrivals at San Francisco, August 1905 to April 1906 (NARA M1414, roll 10)
National Archives Immigrant Database.

website code

Chinese immigration records from 1904 to 1940 survive as scanned passenger lists that are hard to search. Working from 900+ digitized archival pages, I built a GPT-4-assisted OCR and entity-extraction pipeline, recovering 200+ Chinese migrant records with 30+ structured fields each from a noisy 56K-entry dataset, now searchable for historical research.

Life In Weeks Wallpaper project screenshot
Life In Weeks Wallpaper.

website blog

A visual calendar that turns a human life into weeks. It invites people to notice time clearly and cherish the life they are living.