Large Language Models for Fuzzing: A Survey

PhD Qualifying Examination


Title: "Large Language Models for Fuzzing: A Survey"

by

Mr. Kunpeng ZHANG


Abstract:

Fuzzing can execute large numbers of tests, but its effectiveness depends on 
whether they reach meaningful program behavior. For many targets, doing so 
requires assumptions about how the software should be exercised and how 
failures should be recognized. Conventional mutation and feedback mechanisms 
often cannot recover these assumptions, while large language models (LLMs) 
offer another way to derive them from program context. Recent work has 
therefore introduced LLMs throughout the fuzzing workflow, but the value of 
these systems remains difficult to compare because the model contributes at 
different stages and is evaluated with different forms of evidence.

This survey organizes the literature around the fuzzing lifecycle, tracing 
each LLM contribution from its introduction into a campaign to its validation 
through execution. The evidence suggests that LLMs are most useful when they 
provide target-specific reasoning that conventional fuzzing components can 
execute and verify. Outputs that can be reused across many tests also reduce 
the relative cost of inference, although errors in those outputs can affect 
the rest of the campaign. These findings place system integration and 
executable validation at the center of LLM-assisted fuzzing. We conclude by 
discussing the main challenges in making such systems more reliable and their 
results easier to reproduce.


Date:                   Thursday, 20 August 2026

Time:                   9:00am - 11:00am

Zoom Meeting:
https://hkust.zoom.us/j/98825499507?pwd=gIGDGVQesrzUAVa4y3voWxGkuxmGSO.1

Committee Members:      Dr. Shuai Wang (Supervisor)
                        Prof. Charles Zhang (Chairperson)
                        Dr. Lionel Parreaux