[NeurIPS 2026] PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop

💻 Code |   📄 Paper

---

Data Loading Guide

1. Field Descriptions

Each annotation item generally contains the following fields:

• name: Video filename.
• source: Data source of the video.
• extract_frames: Number of pre-extracted frames available for the video.
• prefix: Prompt prefix used to guide model output.
• suffix: Prompt suffix used to guide model output.
• question: Question sent to the model.
• options: Candidate answers sent to the model.
• options_answer: Ground-truth answer used only for evaluation. This field must never be included in the model input.
• frame_select: Used only for the Order task. It specifies the indices of the frames provided to the model.

The Comparison task uses separate fields for the two input videos, including video1_name, video2_name, video1_source, video2_source, video1_extract_frames, and video2_extract_frames.

2. Frame Path Format

All tasks directly use pre-extracted frames.

The original MP4 videos do not need to be decoded or re-sampled during inference.

For a standard single-video task, the frame path follows:

/.../{source}/frames/{video_stem}_{frame_index:04d}.jpg

For example:

source = "videophy2_train"
name = "videophy2_21.mp4"
video_stem = "videophy2_21"
frame_index = 0

corresponds to:

/.../videophy2_train/frames/videophy2_21_0000.jpg

3. Task-Specific Frame Selection

Different tasks use different frame-selection strategies:

• Most tasks: use all extract_frames frames.
• Comparison: use all video1_extract_frames frames from the first video and all video2_extract_frames frames from the second video.
• Order: use only the frames specified by frame_select in the given order.
• Prediction: use only the first half of the extracted frames.

The following frame-loading functions are used for different tasks:

• all_video_frames_1: Spatial, Camera, Scale, Uncertainty, Mechanism, Violation, Counterfactual, Critique, Localization, and Score
• all_video_frames_2: Comparison
• all_video_frames_3: Order
• all_video_frames_4: Prediction

import json
from pathlib import Path

FRAME_ROOT = Path("/.../")


def load_gt(gt_path):
    with open(gt_path, "r", encoding="utf-8") as f:
        return json.load(f)


def make_frame_path(source, video_name, frame_index):
    video_stem = Path(video_name).stem
    return (
        FRAME_ROOT
        / source
        / "frames"
        / f"{video_stem}_{frame_index:04d}.jpg"
    )


# Most tasks
def all_video_frames_1(item):
    return [
        make_frame_path(item["source"], item["name"], index)
        for index in range(item["extract_frames"])
    ]


# Comparison
def all_video_frames_2(item):
    video1_frames = [
        make_frame_path(
            item["video1_source"],
            item["video1_name"],
            index,
        )
        for index in range(item["video1_extract_frames"])
    ]

    video2_frames = [
        make_frame_path(
            item["video2_source"],
            item["video2_name"],
            index,
        )
        for index in range(item["video2_extract_frames"])
    ]

    return video1_frames + video2_frames


# Order
def all_video_frames_3(item):
    return [
        make_frame_path(item["source"], item["name"], index)
        for index in item["frame_select"]
    ]


# Prediction
def all_video_frames_4(item):
    return [
        make_frame_path(item["source"], item["name"], index)
        for index in range(item["extract_frames"] // 2)
    ]

4. Model Input Prompt

The model input consists of the following elements in order:

1. Image frames selected according to the task-specific frame-selection rule.
2. prefix
3. question
4. options
5. suffix

The options_answer field must not be included in the model input. It should only be accessed after inference for evaluation.

The textual prompt can be constructed as follows:

prompt = "\n".join(
    item[key].strip()
    for key in ("prefix", "question", "options", "suffix")
    if item[key].strip()
)

Usage

You can load the dataset using Hugging Face datasets:

from datasets import load_dataset

dataset = load_dataset("HelenPeng/PhysVista")

Contact

For questions, please contact: xg.pengv@gmail.com

Citation

@article{physvista2026,
  title   = {PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop},
  author  = {...},
  journal = {Advances in Neural Information Processing Systems},
  year    = {2026}
}