Postdoctoral Researcher · National University of Singapore
I design and build evaluation infrastructure and coding-agent systems that make AI-generated software more reliable and effective. My aim is to understand how these systems behave in realistic software workflows and identify where they can be improved.
I am currently a postdoctoral researcher at SPARTAN, National University of Singapore, advised by Professor Abhik Roychoudhury. I completed my PhD at Singapore Management University as part of the SOAR Research group, under the guidance of Professors Lingxiao Jiang and David Lo.
ExplainBench evaluates whether coding-agent explanations accurately describe the intended and actual behavior of their patches.
We found that explanation quality does not always align with patch efficacy, and propose ExplanationAuditAgent, which improved explanation scores by 10.9% on average across five coding agents.
c-CRAB evaluates whether feedback from code-review agents leads to correct fixes using executable tests derived from human reviews.
Current review agents collectively solve only around 40% of its tasks, leaving substantial room for human-agent collaboration.
BigCodeBench evaluates practical code generation across 1,140 tasks involving complex instructions and diverse function calls.
Top models reached about 60%, compared with 97% human performance.
Trusted by teams at Zhipu AI, Alibaba Qwen, DeepSeek, Amazon AWS AI, Snowflake AI Research, ServiceNow Research, Meta AI, Cohere AI, Sakana AI, and AI2.
RustMap
RustMap combines LLM-based translation with program analysis to support project-scale migration from C to Rust.
LLM4CBI
LLM4CBI combines language models, program analysis, and adaptive prompt selection to generate effective witness programs for compiler bug isolation.
ArduinoProg
This project explored how domain knowledge and learning techniques can connect natural-language intent and hardware setups with reusable examples and generated Arduino programs.
RecipeGen++
This project developed domain-adapted learning and retrieval techniques to generate trigger-action programs from natural-language requirements.
BiasFinder
BiasFinder uses metamorphic testing to reveal demographic bias in sentiment-analysis systems.
See all publications on Google Scholar .