Towards Evaluation of Implicit Software World Models in Coding LLMs

2026-06-25

AI Summary

This work addresses a critical gap in current large code models, which predominantly focus on syntactic correctness and control flow while lacking deep understanding of software execution behaviors—such as memory usage and runtime resource consumption. For the first time, it extends the evaluation of software world models from control flow to the dimension of execution resources. Leveraging the SWE-bench Verified dataset, the study constructs a multidimensional execution behavior prediction task by integrating test outcomes, exception types, and performance profiling tools to assess whether large language models possess implicit understanding of software execution dynamics. Experimental results reveal that both mainstream and state-of-the-art models exhibit limited and unstable performance on this task, underscoring their significant deficiency in comprehending real-world software execution mechanisms.

Abstract

Software engineering, whether performed by humans or by AI agents, requires reasoning about how software behaves. We call the internal model that supports such reasoning the software world model, and view current code-execution benchmarks as covering one well-studied slice of it -- control flow. In this paper, we take a step toward a broader evaluation by shifting the observable axis to execution resources: alongside test outcome and exception class, we predict peak memory, wall-clock time, and ranked profiler outputs at method and line granularity. We use SWE-bench Verified as the source of data to hold the test close to real-world software engineering tasks. All tested models, frontier ones included, show modest performance and brittle behaviour, suggesting a notable lack of understanding of how software is executed, as opposed to how its source code is written.

Problem

• software world model
• code LLMs
• execution resources
• model evaluation
• program behavior

Innovation

• software world model
• execution resources
• code LLMs
• performance prediction
• SWE-bench Verified