cv

Reviewer & Grant Support

Academic Services

Reviewed papers for EMNLP-2024, ICLR-2025, NAACL-2025, and ACL-2025; supported the Amazon Trustworthy AI Challenge Fellowship proposal.

Bachelor

University of California, Davis

Computer Science and Engineering

• Statistical Machine Learning
• Natural Language Processing
• Trustworthy AI

International Amazon Trustworthy AI Fellow

Amazon

Awarded for contributions in Trustworthy AI with a 250k grant.

2025 CRA Outstanding Undergraduate Researcher Award Honorable Mention

CRA

Honorable mention for outstanding undergraduate research.

UC Davis Provost Undergraduate Fellowship

University of California, Davis

Received a fellowship award of 1.5k.

UC Davis Hong Kong Undergraduate Scholarship

University of California, Davis

Scholarship awarded for academic excellence, amounting to 4k.

UC Davis Vincent V.C. Woo Undergraduate Scholarship

University of California, Davis

Scholarship awarded for academic performance, amounting to 3k.

UC Davis James Leung Family Endowed UG Scholarship

University of California, Davis

Scholarship awarded for academic excellence, amounting to 2k.

UC Davis Dean Witter / Student Development Fund

University of California, Davis

Award of $500 from the Dean Witter / Student Development Fund.

UC Davis 2024 Fall Travel Award

University of California, Davis

Award of $500 for fall travel.

Dean's Honor Roll

University of California, Davis

Maintained position on the Dean's Honor Roll three times.

Chairman's Award for Excellence, 45/45 IB, KGV High School

KGV High School

Received the Chairman's Award for excellence.

BadJudge: Backdoor Vulnerabilities of LLM-as-a-Judge

ICLR

Identified vulnerability in the NLP evaluation pipeline with automatic judges; proposed a principled defense reducing attack success rate to ~0% and achieving state-of-the-art evaluation performance.

Mitigating Backdoor Threats to Large Language Models: Advancement and Challenges

Allerton

Analyzed backdoor attacks in large language models, covering both training-time and test-time threats, and proposed mitigation strategies.

Securing Multi-turn Conversational Language Models from Distributed Backdoor Attacks

EMNLP-Findings

Exposed a threat to multi-turn chatbots with a 97%+ attack success rate and proposed a novel linear-time defense reducing ASR to as low as 0.36%.

ISVP: Influence Selection via Proxy for Efficient Preference Adaptation

Developing influence functions for preference adaptation with a novel derivation and methodology to approximate test-time loss using unobserved preference data.

• Preference-data influence functions
• Test-time loss approximation methodology

Influence Functions for Indirect and Noisy In-Context Learning

Exploring the application of influence functions in in-context learning with indirect and noisy signals, including an EKFAC influence function approach.

• Literature review
• EKFAC influence function