Beyond Score: A Dataset for Joint Action and Score Predictions in 2v2 Sports

Current sports analysis methods often rely on high-level statistics and therefore provide limited insight into the action-level interactions underlying tactical outcomes. To enable fine-grained modeling of tactical processes, we present JAS-2v2, the first public dataset for action-based tactical analysis across 2v2 sports. By unifying beach volleyball, badminton doubles, tennis doubles, and table tennis doubles, JAS-2v2 enables comparative study of diverse tactical structures spanning both open and closed tactical systems under a shared sequence modeling framework. This dataset reframes score prediction as a sequence understanding problem, emphasizing how fine-grained action flows contribute to final tactical outcomes rather than only what the final outcome is. We further propose a unified action-to-score modeling framework based on graph reasoning and temporal modeling. Experiments show that our method outperforms strong baselines on both action prediction and score outcome prediction, and further demonstrate that fine-grained action semantics and sequential action dependencies are critical for tactical reasoning. In particular, using action prediction as an intermediate objective provides more effective representations for score outcome prediction.