Self-Execution Simulation Improves Coding Models

A team of researchers, including Gallil Maimon, Ori Yoran, and contributors from Meta AI, has introduced a novel training method called Self-Execution Simulation (SES) to address a core weakness in large language models for code (Code LLMs). The key insight is that these models are poor at estimating how the code they generate will actually execute. The new approach directly trains models to simulate program execution in a detailed, step-by-step manner, grounding their understanding in actual runtime behavior.

The technique combines two main components: supervised fine-tuning on 'natural language execution traces'—textual explanations of what happens when code runs—and reinforcement learning using verifiable rewards from test outcomes. This equips the model with two complementary skills: predicting a program's output given specific inputs, and solving competitive programming tasks by using either correct execution feedback or its own predictions to guide the solution.

This execution-aware training enables powerful new capabilities for Code LLMs. Most notably, a model can now generate multiple candidate solutions for a problem and then 'self-verify' them by simulating their execution against test cases. If a solution fails, the model can engage in 'iterative self-fixing,' using the simulated execution feedback to understand the bug and revise its code. The paper reports that this method yields consistent improvements across multiple competitive programming benchmarks compared to standard reasoning techniques like chain-of-thought, marking a significant step toward more reliable and autonomous AI coding assistants.