Consistency-Driven Co-Evolution for Self-Supervised Cross-Representation Learning
CoCoEvolve uses representation agreement as the training and inference signal for chart, table, and code understanding.
The paper targets cross-representation learning where the same information can appear as a chart image, table, or visualization code. Its method defines explicit one-to-one correspondences instead of treating mappings as open one-to-many conversions. It applies a consistency objective during training and again at test time for co-optimization. The authors also introduce an evaluation suite covering all six pairwise tasks and report gains across four benchmarks. HF Daily Papers' note
The paper targets cross-representation learning where the same information can appear as a chart image, table, or visualization code. Its method defines explicit one-to-one correspondences instead of treating mappings as open one-to-many conversions. It applies a consistency objective during training and again at test time for co-optimization. The authors also introduce an evaluation suite covering all six pairwise tasks and report gains across four benchmarks. HF Daily Papers' note
score 4