Imperfect Visual Verification for Code Edition : A Case Study on TikZ

The paper presents an empirical study on the effectiveness of iterative refinement in visual code customization using TikZ, focusing on the challenges posed by unreliable verifiers. It evaluates various LLM-based and tool-augmented visual verifiers, finding that even imperfect verifiers can achieve F1-scores up to 0.815 and improve customization outcomes, particularly for weaker models. This research highlights the importance of precise feedback in enhancing the performance of LLMs in tasks requiring localized, semantics-preserving edits, which is critical for practitioners working with visual programming and code generation.