New study shows coding LLMs lack software execution understanding

A new paper accepted at the DL4Code workshop (ICML 2026) challenges the assumption that advanced coding LLMs truly understand how software runs. Researchers Egor Bogomolov and Yaroslav Zharov propose evaluating 'implicit software world models'—the internal reasoning about program behavior. Instead of only checking test outcomes or exception classes (control flow), their benchmark predicts peak memory usage, wall-clock execution time, and ranked profiler outputs at both method and line granularity. Using the real-world SWE-bench Verified dataset as a foundation, they tested multiple frontier models.

Results were sobering: all models showed modest accuracy and brittle performance, especially when predicting resource consumption. The study suggests that current LLMs excel at writing syntactically correct code but lack a robust model of how that code executes inside a machine—a critical gap for autonomous software engineering agents. This work points toward a more holistic evaluation paradigm that could drive future improvements in AI-assisted development.