MergeSE combines coding AI models in seconds without retraining

Fine-tuned code models often become narrow specialists, degrading sharply when faced with out-of-distribution (OOD) data. In the paper's clone-detection setting, a model trained on same-language clones dropped 71% F1 on cross-language clones, while multi-task training fell to 0.151 F1 on unseen AI-generated clones. Post-hoc model merging offers a fix, recovering 93% of multi-task performance without any training data while generalizing 4x better to unseen clone types. However, until now there was no practical tool for software engineering (SE) researchers to diagnose checkpoint compatibility, merge specialists, validate results, and export models.

MergeSE, accepted at ASE 2026, fills that gap. It is an open-source command-line and web tool that works with HuggingFace encoder checkpoints. The tool provides five operations—tasks, inspect, merge, evaluate, and export—and supports five merging algorithms: TIES, DARE-TIES, Wudi, PCB, and averaging. It also detects cross-task classification-head mismatches, produces seedable deterministic outputs, and includes bundled benchmark samples for smoke testing. A full merge of two 124M-parameter checkpoints completes in under 5 seconds on CPU. End-to-end validation confirms MergeSE-produced checkpoints match reference implementations and recover cross-domain performance from specialists. The registry covers nine SE task types, including vulnerability detection, defect prediction, and code-smell detection, making it broadly useful beyond clone detection.