TennisVAR

A Stroke-Evidence-Grounded Multimodal Large Language Model for Tactical Reasoning in Tennis Videos

Yifan Mei1, Qinglin Shi1, Changli Wu1,2, Jiayuan Rao3, Jiayi Ji1, Liujuan Cao1

1 Xiamen University   ·   2 Shanghai Innovation Institute   ·   3 Shanghai Jiao Tong University

Abstract

Sports-video understanding is moving beyond event recognition toward explaining how actions collectively shape match progression. TennisVAR bridges this perception-to-understanding gap by modeling ordered stroke events, their tactical relations, and the evidence supporting each conclusion. The model produces grounded answers, hierarchical tactic labels, supporting strokes, and decisive key actions, making rally-level reasoning explicit and verifiable.

TRACE Benchmark

TRACE organizes rally-level perception, tactical understanding, and decision reasoning around ordered evidence chains.

TRACE benchmark overview

TennisVAR

TennisVAR follows an event - relation - evidence - tactic reasoning paradigm for traceable tactical answers.

TennisVAR model overview

Qualitative Results

TennisVAR links temporally distributed strokes to its tactical answer and rationale.

TennisVAR qualitative results

BibTeX

@article{mei2026tennisvar,
  title  = {TennisVAR: A Stroke-Evidence-Grounded Multimodal Large Language Model for Tactical Reasoning in Tennis Videos},
  author = {Mei, Yifan and Shi, Qinglin and Wu, Changli and Rao, Jiayuan and Ji, Jiayi and Cao, Liujuan},
  year   = {2026}
}