Verifier-Selected Self-Training Boosts AI Workflow Repair Accuracy

The study investigates how governance records, which document the lifecycle of machine-verifiable workflows (task, model attempt, verifier decision, accepted output), can be leveraged to train AI models. The goal is to consolidate occasional or expensive AI capabilities into reliable, one-shot execution, particularly for structured workflow repair. Researchers tested a method called verifier-selected self-training. They used plans generated by a Qwen3-14B model that were accepted by an independent VAL verifier to train the same model for non-thinking execution. This process did not rely on oracle targets or a more powerful teacher model. The results showed a substantial improvement: on 80 new, unseen cases, the number of VAL-accepted plans increased from 1 to 57, with no regressions. This approach significantly outperformed baseline models and even self-selection methods, demonstrating that independent semantic selection is a crucial factor. The method also proved much faster, operating at approximately 1/56th of the latency of the "thinking" model.