Read original
HuggingFace Daily PapersXuehang GuoPapers82

CoCoEvolve: Consistency-Driven Co-Evolution for Self-Supervised Cross-Representation Learning

Original title:Consistency-Driven Co-Evolution for Self-Supervised Cross-Representation Learning

The paper introduces CoCoEvolve, a self-supervised framework for improving consistency among chart images, tabular data, and visualization code. It replaces ambiguous one-to-many cross-representation mappings with explicit one-to-one correspondences and uses agreement among representations as a general optimization signal without extra annotations. CoCoEvolve@Train co-evolves the chart-table-code cycle during training, while CoCoEvolve@Test applies consistency-based co-optimization at inference time. CoCoEvolve@Eval covers all six cross-representation tasks. The abstract reports improvements across four benchmarks, but provides no numerical results, dataset names, baselines, or ablation details.

Why it's worth reading

As analytical systems increasingly connect charts, data, and generation code, this work is timely because it applies the same consistency objective during both training and inference; the full paper is needed to verify the reported gains and costs.

Tags

跨表示学习图表理解表格理解代码生成自监督学习测试时优化多模态一致性学习