Teaching LLMs to Understand Code Execution
Training Models with Dynamic Program Behavior, Not Just Static Code
This research introduces Execution Tuning (E.T.), a novel approach that trains language models to understand how code actually runs, rather than just analyzing static code.
- Trains LLMs on real-world program execution traces without manual annotations
- Enhances code generation and understanding capabilities
- Improves model performance on execution-aware tasks
- Demonstrates better reasoning about dynamic program behavior
Why it matters: This breakthrough enables AI systems to truly understand software execution, leading to more reliable code generation, better debugging assistance, and enhanced software development tools for engineers.
What I cannot execute, I do not understand: Training and Evaluating LLMs on Program Execution Traces