Large language models improve software tasks by reasoning about program states
Improving Large Language Models for Code through Runtime Program-State Reasoning
Artificial Intelligence
Summary
Large language models usually don’t get much training on understanding how computer programs behave when they run. The authors studied whether teaching these models to think about how programs execute helps them do software jobs better. They created two new tasks that make the model find bugs by testing inputs and explain what causes these bugs. By training a 9-billion-parameter model with these tasks, the model got better at fixing code, creating tests, and demonstrating security problems compared to previous versions.
What this means in practice
- •For software developers: Generate more accurate bug fixes and regression tests by using code models trained to reason about program states and execution behavior.
- •For security researchers: Create reproducible security proof-of-concepts automatically by leveraging models that identify bug-triggering conditions and explain their effects.
Authors
Hongwei Li, Spandan Garg, Yufan Huang
Abstract
Large language models receive limited explicit training in reasoning about runtime program states. We study whether training models to reason about runtime program states improves downstream software-engineering capabilities. We introduce two complementary program-state reasoning tasks. Buggy input-output reasoning requires a model to generate a concrete input that exposes a behavioral difference between a buggy program and a hidden correct implementation and to predict the resulting execution behavior. Precondition-postcondition reasoning requires an agent to symbolically characterize a bug-triggering precondition, predict the expected postcondition, explain their causal connection, and instantiate this reasoning as an executable regression test. By incorporating these two tasks into a staged post-training pipeline, we develop Comet-9B, a 9B language model based on Qwen3.5-9B Base. We evaluate the resulting checkpoints on repository-level patch generation, regression-test generation, and security PoC generation. Adding both program-state reasoning tasks to supervised fine-tuning (SFT) on issue resolution improves success rates by 7.25 percentage points on SWE-bench Pro and 9.70 points on SWT-Bench Verified. Sequential reinforcement learning on the two tasks yields further gains of 7.25, 26.79, and 4.67 percentage points on SWE-bench Pro, SWT-Bench Verified, and CyberGym, respectively. Despite having only 9B parameters, Comet-9B achieves a score comparable to the reported GPT-5.2 result on SWE-bench Pro and matches the reported success rate of a GPT-4o-based agent on SWT-Bench Verified.