More about HKUST
Towards Meta-evaluation for LLM-as-a-judge in Web Development Quality
The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
MPhil Thesis Defence
Title: "Towards Meta-evaluation for LLM-as-a-judge in Web
Development Quality"
By
Mr. Chunyang LI
Abstract:
Large language models (LLMs) are increasingly used as automated evaluators
(LLM-as-a-judge) to reduce the cost of human annotation in model development
and assessment. However, their reliability remains unclear in complex,
open-ended domains where quality is subjective and cannot be fully observed
from static text. This thesis studies LLM-as-a-judge in web development
quality evaluation, a setting that requires reasoning over user intent,
source code, visual rendering, and interactive behavior. To support this
study, we construct WebDevJudge, a meta-evaluation benchmark for assessing
whether automated judges can approximate expert human preference judgments
between paired web implementations. Starting from a crowdsourced web
development preference dataset, we filter unsafe, unclear, non-interactive,
and deployment-invalid instances, then annotate a curated subset with expert
judgments guided by hierarchical query-grounded rubric trees. The final
benchmark contains 654 preference instances with balanced labels and supports
multiple observation forms, including code, screenshots, and live interactive
evaluation. Experiments with different evaluators show that current automated
judges still fall short of expert human judgment. Pairwise comparison is
consistently more reliable than single-answer grading because it provides
relative evidence between two implementations, and rubric helps in single
answer grading due to more concrete decision boundary. Comparison between
different observation forms reveals that source code is the most important
observation form, while screenshots provide complementary visual evidence in
pairwise settings. An agentic workflow can collect interaction-based
evidence, but does not outperform strong vanilla LLM judges because errors in
planning, execution, and summarization accumulate. Further analysis
identifies several failure patterns. Positional bias remains present but is
not the dominant source of error. More importantly, current evaluators often
struggle to recognize functional equivalence and to verify whether web
functions are actually feasible. These findings suggest that reliable web
development evaluation requires stronger integration of code understanding,
visual perception, contextual requirement interpretation, and robust
interactive verification.
Date: Friday, 7 August 2026
Time: 10:00am - 12:00noon
Venue: Room 3494
Lifts 25/26
Chairman: Dr. May FUNG
Committee Members: Dr. Yangqiu SONG (Supervisor)
Dr. Binhang YUAN