What Matters in Data Curation for Multimodal Reasoning? Insights from the DCVLR Challenge
This work provides insights into data-efficient multimodal reasoning for AI researchers, though it is incremental as it builds on existing challenge frameworks.
The authors tackled data curation for multimodal reasoning by analyzing the DCVLR challenge, where they won first place using a compact dataset. Their results showed that difficulty-based example selection on aligned data drives performance gains, while increasing dataset size mainly reduces variance without improving mean accuracy.
We study data curation for multimodal reasoning through the NeurIPS 2025 Data Curation for Vision-Language Reasoning (DCVLR) challenge, which isolates dataset selection by fixing the model and training protocol. Using a compact curated dataset derived primarily from Walton Multimodal Cold Start, our submission placed first in the challenge. Through post-competition ablations, we show that difficulty-based example selection on an aligned base dataset is the dominant driver of performance gains. Increasing dataset size does not reliably improve mean accuracy under the fixed training recipe, but mainly reduces run-to-run variance, while commonly used diversity and synthetic augmentation heuristics provide no additional benefit and often degrade performance. These results characterize DCVLR as a saturation-regime evaluation and highlight the central role of alignment and difficulty in data-efficient multimodal reasoning.