BoxComm: leaderboard

BoxComm (2026) tests video models on professional boxing commentary: given four seconds of World Boxing Championship broadcast video, the previous commentary and a requested category, the model writes the next commentary sentence. The score is the share of sentences that GPT-5.2 judges to convey the core action or tactical meaning of the reference sentence, in percent, higher is better.

Metric: Content consistency accuracy (%) over all requested categories; input video only; category-conditioned commentary generation on BoxComm-Eval (the last 40 of 445 World Boxing Championship broadcast matches, 5.6K professional commentary sentences): from the 4 seconds of video before the target moment, the 8 previous commentary sentences and a requested category (play-by-play, tactical or contextual) the model writes one commentary sentence, and GPT-5.2 judges whether it conveys the core action or tactical meaning of the reference sentence; zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 3 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B Instruct23.7#1029
2GPT-4o Mini16.4#1667
3LLaVA-OV-7B (BoxComm checkpoint unspecified)10.9

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=boxcomm · How It Works · Data refreshed daily, snapshot 2026-11-21.