Title:	Runtime-guided LLM-based crash identification and repair for ML notebooks
DNr:	Berzelius-2026-9
Project Type:	LiU Berzelius
Principal Investigator:	Daniel Varro <daniel.varro@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-02-10 – 2026-09-01
Classification:	10205
Keywords:

Abstract

Machine learning (ML) is widely used across domains, with Python-based Jupyter Notebooks serving as a key platform for ML prototyping. Enhancing code quality in ML notebooks is essential but challenging due to dynamic typing in Python, complex ML libraries, and notebook-specific semantics. Identifying crashes in ML notebooks prior to code execution potentially saves significant development time. Static analysis detects bugs without execution. However, it often fails in ML settings where types and API behaviors are not statically known. Large language models (LLMs) show promise for code understanding and bug detection, yet suffer from issues like hallucination and lack of execution context. We propose runtime-augmented LLM-based crash identification approach that augments LLMs with structured runtime information extracted from the notebook kernel, to achieve early crash detection and diagnosis in ML notebook development. Furthermore, we also explore how LLMs and agentic AI use runtime information and potentially crash information (after a crash occurs) to repair bugs in ML notebook context.