Research map / Vision-language and multimodal / Multimodal LLMs

SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SportR, a multi-sport image and video QA benchmark with human-authored chain-of-thought rationales and bounding-box grounding, and shows zero-shot MLLMs struggle while SFT+GRPO on Qwen2.5-VL-7B helps.

Problem

MLLMs lack fine-grained visual perception and rule-based reasoning needed for sports adjudication; existing sports benchmarks cover single sports or lack CoT rationales and visual grounding annotations.

Approach

A multi-sport benchmark with a progressive 7-question (image) and 6-question (video) QA hierarchy, fully human-authored CoT rationales, and manual bounding boxes for explicit visual grounding of rule infractions, plus a SFT+GRPO training recipe on the image split.

Key results

• GPT-5 reaches only 5.70 IoU on image grounding (Q5), and zero-shot baselines stay below 7% IoU.
• Qwen-VL-7B SFT+RL raises Q2 foul classification from 14.43 to 51.54 and grounding IoU from 4.61 to 9.94.
• Qwen2.5-VL-7B trained only on images improves video infraction identification (Q8) from 25.49 to 59.52, surpassing GPT-5 (59.17).

Surprising findings

• A model trained only on image data improves video infraction identification, suggesting cross-modal transfer.
• GPT-5's domain knowledge gap rises from 20% on video to 36% on images, indicating rule mapping rather than perception is the bottleneck on static images.

Limitations

• Grounding is evaluated only on static images; no video localization.
• Explanation task (Q4) excluded from RL because no suitable sports-specific metric exists.
• LLM-as-judge correlation with humans is moderate (Pearson about 0.65 for image, 0.70 for video).

Experimental setup

Datasets: SportsImage SportsVideo

Benchmarks: SportR (SportsImage) SportR (SportsVideo)

Metrics: accuracy via LLM-as-judge score IoU

Baselines: GPT-5 Claude 4 Sonnet Gemini 2.5 Pro Qwen2.5-VL-72B GLM-4.5V MiniCPM-V-4.5 LLaVA-Next DeepSeek-VL2 Video-R1-7B

Builds on: GRPO SPORTU

Closest papers by meaning

• SPORTU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models (ICLR 2025 · similarity 0.95)
• SoccerMaster: A Vision Foundation Model for Soccer Understanding (CVPR 2026 · similarity 0.88)
• Towards Universal Soccer Video Understanding (CVPR 2025 · similarity 0.88)
• MVU-Eval: Towards Multi-Video Understanding Evaluation for Multimodal LLMs (NeurIPS 2025 · similarity 0.88)
• ViSTec: Video Modeling for Sports Technique Recognition and Tactical Analysis (AAAI 2024 · similarity 0.87)
• Spot The Ball: A Benchmark for Visual Social Inference (CVPR 2026 · similarity 0.87)