WhiteFox: White-box Compiler Fuzzing Empowered by Large Language Models
summary
WhiteFox is a white‑box compiler fuzzer that reads optimization source code, uses an analysis agent (GPT‑4) to summarize the precise triggering requirements, and uses a generation agent (StarCoder‑15B) with a feedback loop to synthesize tests that exercise those conditions. It targets PyTorch Inductor, TensorFlow‑XLA and TensorFlow Lite, exercises up to eight times more optimizations than prior fuzzers, and surfaces 101 bugs — 92 previously unknown and about 70 already fixed — demonstrating practical, scalable bug finding.
AI-written summary · about
chapters (15)
- 0:00WhiteFox project intro
- 0:20Deep learning compilers background
- 1:24Fuzzing method limitations
- 2:51WhiteFox approach insights
- 3:32WhiteFox system overview
- 4:37Requirement analysis agent
- 5:21Generation agent inputs
- 7:10Implementation details
- 8:13Evaluation vs baseline
- 9:16Ablation study results
- 10:20Bug detection results
- 11:06Fuse attention bug case
- 11:49Linear ReLU bug case
- 13:12WhiteFox summary
- 13:53Q&A
related talks
53mLocknote: How Badly Do We Want Correct Compilers?
15mInterleaving Large Language Models for Compiler Testing
14mFinding Compiler Bugs through Cross-Language Code Generator and Differential Testing
22mSatisfiability Modulo Fuzzing: A Synergistic Combination of SMT Solving and Fuzzing
16mQuteFuzz: Fuzzing quantum compilers using randomly generated circuits with control(…)
14mFuzzing C++ Compilers via Type-Driven Mutation
50mD2T2 - Exposing Hidden Exploitable Behaviors Using EDF
22mBoosting Compiler Testing by Injecting Real-World Code