PrismaCoder: Pushing the Boundaries of Open Multimodal Code Intelligence through Decomposed Rubric Synthesis

One-line takeaway

HKU NLP's PrismaCoder targets multimodal code (charts, webpages, animations, documents): the PrismaForge pipeline synthesizes checkable rubrics used both to filter SFT data and as RL rewards on visual-coding tasks, yielding the PrismaCoder-2M dataset and released 9B/35B models that the authors say lead open SOTA by a wide margin and rival top proprietary systems.

港大 NLP 组的 PrismaCoder 面向“代码→可视结果”（图表、网页、动画、文档）：PrismaForge 数据管线合成可逐项检查的评分细则，既用来筛选 SFT 数据，也作为视觉编程任务的 RL 奖励；据此构建 PrismaCoder-2M 数据集，并开源 9B 和 35B 模型，作者称其大幅领先开源 SOTA、接近领先闭源系统。