CaST-Bench: Benchmarking Causal Chain-Grounded Spatio-Temporal Reasoning for Video Question Answering

This is the official repository for the CaST-Bench dataset, introduced in the paper "CaST-Bench: Benchmarking Causal Chain-Grounded Spatio-Temporal Reasoning for Video Question Answering".

CaST-Bench is the first benchmark to evaluate Vision-Language Models (VLMs) on causal chain reasoning grounded in fine-grained spatio-temporal evidence. Given a video and a causal question, a model must not only select the correct answer, but should also identify and spatiotemporally localize the chain of visual evidence that grounds the answer—covering both the causal trigger and its downstream effects.

This release provides a carefully curated set of human-verified QA pairs, intended for rapid and reliable evaluation of VLMs on causal video understanding.

---

News

• [2026-06-17] Leaderboard Launched — The CaST-Bench Leaderboard is now live on Hugging Face Spaces! Track and compare model performance on CaST-Bench v1.
• [2026-05-28] v1 Released — We publicly release CaST-Bench v1, a carefully selected subset of the full benchmark. This partial release is intended as a reference version for fast and reliable evaluation of VLMs.
• [Planning] v2 Release — We plan to release CaST-Bench v2 with a larger number of QA pairs.

---

Dataset Structure

.
├── castbench_hf.jsonl   # Test split (human-verified QA pairs)
└── videos/              # MP4 video clips sourced from SAV

Data Fields

Each line in castbench_hf.jsonl is a JSON object with the following top-level fields:

Field	Type	Description
video	string	MP4 filename (e.g. "sav_000267.mp4")
question	string	Natural-language causal question about the video
question_type	string	Question subcategory name (see taxonomy below)
scene_category	string	Short free-text label describing the scene
options	object	Six answer choices with keys "A" through "F"
answer	string	Key of the correct answer (one of "A"–"F")
evidence	string	JSON-encoded array of spatio-temporal evidence objects

Evidence Object Fields

The evidence field is a JSON-encoded list. Each element has:

Field	Type	Description
evidence_start_time	string	Segment start time in "mm:ss" format
evidence_end_time	string	Segment end time in "mm:ss" format
evidence_instance_id	string	Tracked instance identifier (e.g. "person_FnH", "car_lEu")
evidence_rationale	string	Natural-language description of what this evidence contributes causally
SAV_instance_id	string	Reference to the source SAV tracking file
bboxes_in_range	object	Per-second bounding boxes; keys are integer seconds as strings, values are "[x_min, y_min, x_max, y_max]" strings

Answer Options Structure

Each QA item has exactly six answer options designed to probe distinct failure modes:

Option	Type
1 of 6	Correct answer
1 of 6	Near-miss answer (same structure, one key attribute altered)
1 of 6	Text-based distractor (plausible without watching the video)
1 of 6	Near-miss text-based distractor
1 of 6	Video-based distractor (based on causally irrelevant instances present in the video)
1 of 6	Near-miss video-based distractor

The option keys A–F are randomly shuffled per item, so the correct answer is not always the same letter.

---

Question Type Taxonomy

CaST-Bench organizes questions into four top-level categories covering the breadth of causal reasoning:

Category	Subcategory
1. Causal Explanation	Why Questions (Causal Reasons)
	How Questions (Mechanisms)
2. Counterfactual Reasoning	Physical Counterfactual
	Social Counterfactual
3. Predictive Anticipation	Physical Prediction
	Behavioral Prediction
4. Inferential Description	Identity / Role Inference
	Interpersonal Relationship Inference
	Goal / Intent Inference (non-immediate)
	Emotion / Motivation Inference
	Skill / Expertise Inference
	Ownership / Belonging Inference
	Risk / Safety Assessment
	Success / Failure Inference
	Environmental Condition Inference
	Other Contextual Inference

---

Data Splits

Split	File	Description
test	castbench_hf.jsonl	Human-verified QA pairs for rapid evaluation

This release contains a rigorously quality-controlled subset of the full CaST-Bench benchmark. Each QA pair has been manually reviewed by expert annotators and passed a multi-stage automated filtering pipeline (see Dataset Construction below).

---

Dataset Construction

CaST-Bench was built via a multi-stage Human-AI collaborative pipeline:

1. Video Selection — Videos were drawn from the SAV (Segment Anything Video) dataset, which provides densely annotated instance segmentation tracks. A VLM automatically filtered for videos with multiple people engaged in distinct activities in stable, high-quality shots.

2. Spatio-Temporal Instance Description — For each tracked instance, a 3-step pipeline generated timestamped action logs: (i) instance isolation into a cropped clip, (ii) static description using highlighted frames, (iii) dynamic description over the cropped clip. Seven professional native-English-speaking annotators then reviewed and rewrote every AI-generated caption using a custom web annotation tool, yielding over 10,000 high-quality instance descriptions.

3. QA and Causal Chain Generation — A VLM "causal thinking agent" took the instance descriptions, video, and the question taxonomy as input and generated structured QA pairs with causal evidence chains in a prescribed JSON schema.

4. Multi-Choice Distractor Generation — Text-based and video-based distractors, along with near-miss variants for all three base options, were generated and verified for semantic exclusivity.

5. Multi-Stage Filtering — (i) A text-only LLM filter removed questions answerable without video. (ii) A video-mask filter blacked out all causal-chain bounding boxes; items answerable on the masked video were discarded. (iii) Human annotators performed a final quality review and revise the QA and evidences.

---

Citation

@inproceedings{zhang2026castbench,
  title     = {CaST-Bench: Benchmarking Causal Chain-Grounded Spatio-Temporal Reasoning for Video Question Answering},
  author    = {Zhang, Mingfang and Pan, Jingjing and Kumar, Ashutosh and Saini, Rajat and Erdogan, Mustafa and Yang, Hsuan-Kung and Kang, Caixin and Huang, Yifei and Sato, Yoichi and Kong, Quan},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

---

License

The benchmark annotations are released under CC BY 4.0. The accompanying video clips are sourced from the SAV dataset and are subject to SAV's original license terms.