2026 · Preprint
UniCoder: Unified Visual-to-Code Generation via Symbolic Rewards and Reference-Guided Code Optimization
UniCoder trains a visual-to-code model for charts, SVGs and webpages. Its reinforcement-learning objective combines rendering checks, image similarity and code-derived visual attributes; low-reward training groups receive reference-code trajectories to improve exploration.
Scope and limitations
Experiments concern static visual reconstruction with one 8B backbone, including Design2Code and ScreenBench. Reference code is used during training; successful rendering does not prove interactive behavior. Extending the attribute extractor to new domains requires new schemas/prompts. Interactive interfaces, animation and mobile-app behavior are not evaluated.
Why it fits this collection
Explores training signals for recovering code from visual website references.
Publication and review details
- Venue
- arXiv
- Status
- Preprint
- Publication / posting date
- June 30, 2026
- Date source
- arXiv v1 submission date; no separate proceedings or journal publication was verified.
- First arXiv submission
- June 30, 2026
- arXiv
- 2606.31732
- Reviewed arXiv version
- v1
- DOI
- 10.48550/arXiv.2606.31732
At 2844ebd, the repository releases ChartMimic/UniSVG evaluation wrappers, not model weights, full training code or datasets. Original repository code is Apache-2.0; adapted benchmarks retain their own terms. The arXiv title includes “UniCoder”; the HTML heading and software citation omit that prefix.
Sources reviewed
Reviewed against the paper and its primary sources. Cite the original authors using the linked paper or BibTeX.
Primary sources
- https://arxiv.org/abs/2606.31732v1
- https://arxiv.org/html/2606.31732v1
- README.md GitHub · JimmyZhengyz/unicoder · Revision 2844ebd
- MODEL_RELEASE.md GitHub · JimmyZhengyz/unicoder · Revision 2844ebd
- LICENSE GitHub · JimmyZhengyz/unicoder · Revision 2844ebd
- THIRD_PARTY_NOTICES.md GitHub · JimmyZhengyz/unicoder · Revision 2844ebd