2026 · Preprint
PrismaCoder: Pushing the Boundaries of Open Multimodal Code Intelligence through Decomposed Rubric Synthesis
Trains visual-code models using separate code and rendered-output rubrics for data filtering and reinforcement learning. Relevant evaluations include rebuilding webpages, charts and vector graphics from images.
Scope and limitations
Combines visual reconstruction with text-driven generation. Judge scores do not establish behavioral equivalence. Training and evaluation share some judge families; the human check uses 100 samples. Instruction n-gram filtering cannot exclude all pretraining exposure.
Why it fits this collection
Provides a recent approach to reconstruction from rendered evidence.
Publication and review details
- Venue
- arXiv
- Status
- Preprint
- Publication / posting date
- October 14, 2026
- Date source
- First arXiv submission; no reviewed venue publication is established.
- First arXiv submission
- October 14, 2026
- arXiv
- 2610.20538
- Reviewed arXiv version
- v1
- DOI
- 10.48550/arxiv.2610.20538
Work in progress. DataCite confirms the DOI despite the arXiv pending label. The project code link is pending; the dataset API returned HTTP 401 anonymously. Reviewed model inventories include 9B and 35B-A3B-SFT with MIT card metadata. No weights were downloaded.
Sources reviewed
Source review only; experiments were not rerun for this collection. Cite the original authors using the linked paper or BibTeX.
Primary sources
- https://arxiv.org/abs/2610.20538v1
- https://arxiv.org/html/2610.20538v1
- https://api.datacite.org/dois/10.48550/arxiv.2610.20538
- https://qiushisun.github.io/PrismaCoder-Home/
- https://huggingface.co/collections/QiushiSun/prismacoder
- https://huggingface.co/QiushiSun/PrismaCoder-9B/tree/d5743a565f9f9b55da0798c4636059074f21602a
- https://huggingface.co/QiushiSun/PrismaCoder-35B-A3B-SFT/tree/e2df1e846cbfe5a2aebc860917b02c742007c62e