prokajevo/splice-benchmark

SPLICE: A Human Curated Benchmark for Probing Visual Reasoning in VLMs

Dataset Summary

SPLICE (Sequential Processing for Learning and Inference in Chronological Events) is a human-curated benchmark designed to evaluate the temporal and causal reasoning capabilities of Multimodal Large Language Models (MLLMs). The core task is to reorder a set of shuffled video segments from a single procedural event into their correct chronological sequence. SPLICE dataset

This dataset was created to address the "evaluation deficit" in video understanding, where existing benchmarks often fail to probe deep, structural reasoning and are susceptible to linguistic shortcuts. The dataset is derived from 3,381 instructional videos from the COIN dataset, segmented into 11,423 coherent event clips.

Supported Tasks and Capabilities

The SPLICE benchmark is designed to evaluate a model's core visual reasoning capabilities through a primary task that necessitates a synthesis of multiple underlying reasoning skills.

Primary Task: Video Segment Reordering

The fundamental task is to reconstruct a coherent event sequence from a set of shuffled video clips. This generative challenge requires the model to understand the entire procedural flow and the logical progression of an event.

Probing Multiple Dimensions of Reasoning

As detailed in the paper, the benchmark is explicitly designed to probe five distinct dimensions of event-based reasoning:

• Temporal Reasoning: Grasping the correct linear sequence of events by tracking the changing states of objects over time.
• Causal Reasoning: Recognizing the cause-and-effect relationships that govern why events happen in a particular order.
• Spatial Reasoning: Interpreting the trajectories, orientations, and spatial relationships of objects in 3D space.
• Contextual Reasoning: Using the environment and the logic of a process to infer dependencies.
• General Knowledge Reasoning: Applying prior commonsense or domain-specific world knowledge to infer a plausible sequence.

Dataset Structure

The dataset is structured with one row per video segment (clip). Each row contains a playable video and all its associated metadata, flattened for easy access and analysis. This flat structure, with 11,423 rows in total, is designed for direct use with the Hugging Face Datasets library and the automatic Dataset Viewer. The dataset itself contains all 3,381 validated tasks.

Data Fields

The dataset contains the following columns:

• video: The playable video clip for the segment. This is automatically loaded from the file_name column in the metadata.
• file_name: The relative path to the video clip file (e.g., splice_video_segments/00/oDAY5PMwZgU_part_1.mp4).
• video_id: The unique ID of the original parent video from the COIN dataset. This allows you to group all segments belonging to the same task.
• part: The ground-truth chronological order of this segment within its parent video task (1-indexed). This is the target for the reordering task.
• label: The human-annotated text description of the event in the video segment.
• domain: The high-level category of the task (e.g., 'Electrical Appliance', 'Science and Craft').
• class: The specific name of the task (e.g., 'UseVendingMachine').
• subset: The original split from the COIN dataset (e.g., 'training', 'testing').
• start / end: The start and end timestamps of this segment within the original, full-length video.
• segment_id: The unique ID for this specific segment from the COIN dataset.
• task_duration: The total duration in seconds of the original, full-length parent video.
• video_url: The original YouTube URL of the full-length parent video.

Data Splits

The dataset is provided as a single main split, containing all 3,381 validated tasks.

Dataset Creation

The SPLICE benchmark was created through a rigorous six-stage curation pipeline:

1. Sourcing: Instructional videos were selected from the COIN dataset.
2. Selection: A diverse subset of 3,600 videos was randomly sampled.
3. Segmentation: Videos were temporally segmented into event clips using COIN's ground-truth annotations.
4. Anonymization: All audio tracks, metadata, and original filenames were stripped to isolate visual reasoning for the SPLICE dataset and publication. Clips were standardized to .mp4 format. However, for this dataset, the non-anonyzed version has been presented to encourage versatility, as metadata and audio stripping can be done based on what needs to be probed in future tasks.
5. Human Validation: Each of the 3,600 tasks (a set of shuffled clips) was presented to two independent human annotators. A task was only included in the final benchmark if it was not deemed "unsolvable" by both annotators, ensuring high quality and unambiguous solvability. Solvability does not equal correctness.
6. Final Benchmark: The final dataset comprises 3,381 validated tasks, totaling 11,423 clips.

Sample Data Point

A single row in the dataset looks like this:

{
  "file_name": "splice_video_segments/00/oDAY5PMwZgU_part_1.mp4",
  "video_id": "oDAY5PMwZgU",
  "part": 1,
  "label": "insert money into the vending machine",
  "domain": "Electrical Appliance",
  "class": "UseVendingMachine",
  "subset": "training",
  "start": 2.0,
  "end": 7.0,
  "segment_id": "519",
  "task_duration": 17.205978,
  "video_url": "https://www.youtube.com/embed/oDAY5PMwZgU",
  "video": {
    "path": "splice_video_segments/00/oDAY5PMwZgU_part_1.mp4",
    "bytes": "..."
  }
}

Citation Information

If you use the SPLICE dataset in your research, please cite the following paper:

@inproceedings{
ballout2025can,
title={Can you {SPLICE} it together? A Human Curated Benchmark for Probing Visual Reasoning in {VLM}s},
author={Mohamad Ballout* and Okajevo Wilfred* and Seyedalireza Yaghoubi and Nohayr Muhammad Abdelmoneim and Julius Mayer and Elia Bruni},
booktitle={The 2025 Conference on Empirical Methods in Natural Language Processing},
year={2025},
url={https://openreview.net/forum?id=deFgBHsHxl}
}