Worked on enhancing Dynamic Programming Optimization training within the PaddlePaddle/PaddleFormers repository by implementing a filtered label loss approach. Refactored the loss function to replace the sparse head and loss function, and updated response indexing logic to support the new method, which improved both training efficiency and model accuracy for DPO tasks. Collaborated closely with team members to ensure seamless integration with existing training loops and maintained compatibility with regression tests. Utilized Python and deep learning frameworks to streamline the training pipeline, enabling more scalable experimentation and supporting higher-quality decision making in dynamic programming-based production workflows. No bugs were reported or fixed.