VCode: a multimodal coding benchmark with SVG as symbolic visual representation

Conference item

Code has emerged as a precise, executable medium for linguistic-centric tasks, leaving visual-centric coding underexplored. Conventional image representations rely on RGB pixels that capture visual appearance but offer limited symbolic abstraction. In this work, we advocate SVG code as a compact, interpretable, and executable visual representation. We introduce VCode, a benchmark that reframes multimodal understanding as code generation: given an image, a model must produce SVG that preserves symbolic meaning for downstream reasoning. VCode covers general commonsense, professional disciplines, and visual-centric perception. To assess symbolic fidelity, we propose CodeVQA, a novel evaluation protocol where a policy model answers questions over rendered SVGs; correct answers indicate faithful symbolic preservation. We also introduce VCoder, an agentic framework that augments VLMs via test-time revision and visual tool use, yielding substantial improvements over strong baselines. The models are available at https://csu-jpg.github.io/VCode.

Authors: Lin, KQ; Zheng, Y; Ran, H; Zhu, D; Mao, D

• Publication status: Accepted
• Peer review status: Peer reviewed
• Publisher: IEEE
• Acceptance date: 2026-04-24
• Event title: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2026)
• Event location: Denver, Colorado, USA
• Event website: https://cvpr.thecvf.com/Conferences/2026
• Event start date: 2026-06-03
• Event end date: 2026-06-07
• Language: English