The authors introduce two complementary program-state reasoning tasks: buggy input-output reasoning and precondition-postcondition reasoning. They incorporate these tasks into a staged post-training pipeline, resulting in Comet-9B, a 9B-parameter model based on Qwen3.5-9B Base. The model is evaluated on repository-level patch generation (SWE-bench Pro), regression-test generation (SWT-Bench Verified), and security PoC generation (CyberGym).
Adding the program-state reasoning tasks to supervised fine-tuning 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 each benchmark respectively. Despite its small size, 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. This demonstrates that explicit training on runtime program-state reasoning can significantly improve downstream software-engineering capabilities, even with a 9B parameter model.