VisualProg Distiller: Learning to Fine-tune Non-differentiable Visual Programming Frameworks

As an interpretable and universal neuro-symbolic paradigm based on Large Language Models, visual programming (VisualProg) can execute compositional visual tasks without training, but its performance is markedly inferior compared to task-specific supervised learning models. To increase its practicality, the performance of VisualProg on specific tasks needs to be improved. However, the non-differentiability of VisualProg limits the possibility of employing the fine-tuning strategy on specific tasks to achieve further improvements. In our analysis, we discovered that significant performance issues in VisualProg's execution originated from errors made by the sub-modules at corresponding visual sub-task steps. To address this, we propose ``VisualProg Distiller", a method of supplementing and distilling process knowledge to optimize the performance of each VisualProg sub-module on decoupled visual sub-tasks, thus enhancing the overall task performance. Specifically, we choose an end-to-end model that is well-performed on the given task as the teacher and further distill the knowledge of the teacher into the invoked visual sub-modules step-by-step based on the execution flow of the VisualProg-generated programs. In this way, our method is capable of facilitating the fine-tuning of the non-differentiable VisualProg frameworks effectively. Extensive and comprehensive experimental evaluations demonstrate that our method can achieve a substantial performance improvement of VisualProg, and outperforms all the compared state-of-the-art methods by large margins. Furthermore, to provide valuable process supervision for the GQA task, we construct a large-scale dataset by utilizing the distillation process of our method.