GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions

📜 Paper · 🌐 Project Page · 💻 Code · 🤖 Model

GRASP is a large-scale social reasoning dataset that connects high-level social QA with fine-grained gaze and deictic gesture events, organized by a 16-category taxonomy spanning gaze (T1–T6), gesture (G1–G6), and joint gaze–gesture reasoning (J1–J4), together with GRASP-Bench for evaluation.

Structure

train/
├── annotations.jsonl        # all training QA, one row per QA (shown in the Dataset Viewer)
├── data_sft.jsonl           # open-ended QA with structured reasoning traces (SFT, training format)
├── data_rl.jsonl            # MCQ with ground-truth social events (SGR / GRPO, training format)
└── videos/                  # per-QA video clips, tar parts (<=9GB each)
    ├── avsbench_part001.tar         # extracts to avsbench/video/*.mp4
    ├── embody3d_part001.tar ...
    └── ...
test/
├── annotations.jsonl        # GRASP-Bench QA, one row per QA (shown in the Dataset Viewer)
└── grasp_bench.tar          # GRASP-Bench: json/*.json + video/*.mp4 (1,196 items)

annotations.jsonl columns: id, video (path inside the extracted archives), source, category, difficulty, format (mcq / open_ended), question, options, answer, response (structured reasoning target for open-ended training QA).

Each training JSONL line is a list of input specs:

[{"type": "video", "path": "avsbench/video/<clip>.mp4", "fps": 2.0, ...},
 {"type": "text", "content": "<question>", "output": false},
 {"type": "text", "content": "<answer>", "output": true}]

Video path fields are relative to the extracted train/videos/ root. Extract all tars into one directory and point training at it:

for t in train/videos/*.tar; do tar -xf "$t" -C /path/to/video_root; done

GRASP-Bench items are one JSON per QA with clip_file, qa.category (T1–T6 / G1–G6 / C1–C4 for joint), qa.question, qa.options, qa.answer.

Usage

Training and evaluation code: https://github.com/Social-Reaoning/grasp

License

GRASP annotations (QA pairs and social event labels) are released under CC BY-NC 4.0 for non-commercial research use only. Video clips are derived from AVSBench, Embody3D, FriendsMMC, Social-IQ, SocialGesture, TVQA, and Werewolf Among Us, and remain subject to the licenses and terms of use of their original sources. Commercial use is not permitted.

Citation

@article{kim2026grasp,
  title={GRASP: Learning to Ground Social Reasoning in Multi-Person Non-Verbal Interactions},
  author={Kim, Junho and Cao, Xu and Yang, Houze and Boote, Bikram and Jojic, Ana and Ryan, Fiona and Lai, Bolin and Lee, Sangmin and Rehg, James M},
  journal={arXiv preprint arXiv:2605.15764},
  year={2026}
}