Investigating Evolutionary Characteristics of Deep Learning Framework Bugs in the Large Language Model Era

Since ChatGPT’s release in November 2022, deep learning frameworks like PyTorch and TensorFlow have undergone extensive updates to support the transition from traditional pre-trained models to large language models (LLMs). To investigate how framework bugs have evolved in this new era, we conducted a large-scale empirical study analyzing 552 bugs from these two frameworks spanning December 2022 to September 2025. We categorized bugs using a comprehensive taxonomy covering 13 root cause types and 6 symptom categories. Our findings reveal significant distributional shifts in bug characteristics driven by the adaptation to LLMs, providing actionable insights for framework testing and maintenance in the evolving landscape of deep learning infrastructure.