Scientific Graphics Program Synthesis via Dual Self-Consistency Reinforcement Learning

TL;DR

This paper introduces a new dataset, benchmark, and a dual self-consistency reinforcement learning method for scientific graphics program synthesis, significantly improving the accuracy and fidelity of generated TikZ code.

Contribution

The paper presents a high-quality dataset, a comprehensive benchmark, and a novel reinforcement learning paradigm for improved scientific graphics synthesis.

Findings

SciTikZer-8B outperforms proprietary and large models in graphics synthesis.

The dataset covers 11 scientific disciplines with high-quality, executable data.

The reinforcement learning approach enhances structural and visual fidelity.

Abstract

Graphics Program Synthesis is pivotal for interpreting and editing visual data, effectively facilitating the reverse-engineering of static visuals into editable TikZ code. While TikZ is the de facto standard for scientific schematics due to its programmatic flexibility, its requirement for rigorous spatial precision presents a significant challenge for Multimodal Large Language Models. Progress is currently stifled by two primary gaps: (1) Data Quality Gap: existing image-TikZ corpora often lack strict executability and reliable visual alignment; (2) Evaluation Gap: a lack of benchmarks for both structural and visual fidelity. To address these, we present a closed-loop framework featuring: SciTikZ-230K, a large-scale, high-quality dataset from our Execution-Centric Data Engine covering 11 diverse scientific disciplines; SciTikZ-Bench, a multifaceted benchmark spanning from basic geometric constructs to intricate hierarchical schematics to evaluate both visual fidelity and structural logic. To further broaden the scope of visual-code optimization methodology, we introduce a novel Dual Self-Consistency Reinforcement Learning optimization paradigm, which utilizes Round-Trip Verification to penalize degenerate code and boost overall self-consistency. Empowered by these, our trained model SciTikZer-8B achieves state-of-the-art performance, consistently outperforming proprietary giants like Gemini-2.5-Pro and massive models like Qwen3-VL-235B-A22B-Instruct.