LLM-Aided Study of TorchDynamo Uncovers 23 Previously Unknown PyTorch 2 Frontend Bugs, 15 Confirmed

arXiv 2607.25651·medium signalXBlueskyLinkedInCopy link

A paper posted 2026-07-28 (arXiv 2607.25651) delivers the first systematic empirical study of frontend bugs in TorchDynamo, the default deep-learning-compiler frontend for PyTorch 2, a stage prior work skipped in favor of low-level operators or treating compilers as monolithic. Using a domain-knowledge-enhanced LLM-aided methodology the authors analyze 123 frontend bugs and build a taxonomy of 7 root-cause categories and 15 subcategories. They then use the LLM to generate root-cause-aware test cases and uncover 23 previously unknown frontend bugs in recent releases across eight (sub)categories, with 15 already confirmed — a concrete demonstration that taxonomy-guided LLM test generation finds real defects in production compiler infrastructure.