CoCoEvolve trains chart-table-code models without labels via view agreement

Something that stood out from a new Hugging Face paper listing this week: a team from William & Mary, IBM, the Allen Institute for AI, Oak Ridge National Laboratory, and Northwestern is arguing that you can train chart, table, and visualization-code models without the labelled correspondences those benchmarks normally require. Their framework, CoCoEvolve, treats the three formats as views of the same underlying semantics and uses mutual agreement between the views as the reward signal.

The setup runs two models in a chart to table to code cycle and rewards them for producing outputs that agree once the code executes back into a rendered chart. Consistency is measured through a mix of execution success, embedding similarity, and image comparisons using CLIP, SSIM, DINO, and OCR. On the code-generation side, the authors report Qwen3-VL-2B moving from 58.56% to 87.04% sandbox execution, Qwen3-VL-4B from 78.47% to 93.52%, and Qwen3-8B hitting 100.00% when the train-time and test-time variants are combined. Reported gains reach 37.91% on non-overlapping test sets and 46.88% in out-of-domain settings.

Why the annotation-free framing matters: chart-to-table and code-to-chart datasets are expensive to label and inherently one-to-many, and the paper's own analysis says naive SFT on those correspondences can actually degrade performance, with drops of up to 25.19%. If mutual consistency across formats is a strong enough proxy, teams already running post-training loops with GRPO, DAPO, or GSPO can bolt on a reward without paying for more human labels, and the authors say the method integrates with all three.

The honest caveats are in the paper itself. The models are small (up to 8B), the charts are static, and applicability to animated or dynamic visualizations is called out as unexplored. Cross-view agreement can also plateau on outputs that are wrong in correlated ways, which the paper flags as a failure pattern where errors in the table propagate into both the code and the rendered image. What the write-up does not give you is a head-to-head against large closed multimodal systems, or the compute cost of the test-time optimization pass.

Still, the direction is worth watching. If agreement-as-supervision holds up on larger models and messier data, this is the kind of trick that quietly makes the next generation of BI copilots noticeably more reliable.