Jaehwisong/XVR

Learning Multi-View Spatial Reasoning from Cross-View Relations

CVPR 2026

Project Page | arXiv | Model: Qwen3-VL-2B-XVR

XVR (Cross-View Relations) is a large-scale dataset designed to teach Vision-Language Models (VLMs) spatial reasoning across multiple viewpoints. It comprises 100K vision-question-answer samples derived from 18K diverse 3D scenes (general domain) and 70K robotic manipulation trajectories (robotic domain), spanning three fundamental spatial reasoning tasks: Correspondence, Verification, and Localization.

VLMs fine-tuned on XVR achieve clear improvements on multi-view and robotic spatial reasoning benchmarks (MindCube, RoboSpatial), and when used as backbones in Vision-Language-Action (VLA) models, improve manipulation success rates on RoboCasa by ~13% absolute on average.

Authors

Suchae Jeong*, Jaehwi Song*, Haeone Lee, Hanna Kim, Jian Kim, Dongjun Lee, Dong Kyu Shin, Changyeon Kim, Dongyoon Hahm, Woogyeol Jin, Juheon Choi, Kimin Lee (KAIST · Config · Hanyang University · Yonsei University · Seoul National University)

*Equal contribution

Dataset Structure

XVR/
├── train.jsonl              # 103,576 training samples
├── valid.jsonl              #   3,208 validation samples
├── xvr_eval.jsonl           #   1,866 held-out evaluation samples (XVR-Eval)
└── image_shards/            # 396,320 images packed into ~40 tar shards
    ├── images_000.tar       # ~10K images per shard (~2 GB each)
    ├── images_001.tar
    └── ...

After extraction, the layout becomes:

XVR/
├── train.jsonl
├── valid.jsonl
├── xvr_eval.jsonl
└── images/                  # 396,320 image files (extracted from shards)
    ├── img_0000001.png
    └── ...

The prompt_blocks in each JSONL reference images by the relative path images/img_XXXXXXX.{png,jpg}, so the extracted layout matches exactly.

Splits

Split	# Samples
train	103,576
validation (valid.jsonl)	3,208
test (xvr_eval.jsonl, XVR-Eval)	1,866

xvr_eval.jsonl is the XVR-Eval benchmark, constructed from data sources unseen during XVR creation (MobileAloha trajectories and WildRGB-D boat category scenes).

Sample Format

Each line in the .jsonl files is a JSON object:

{
  "sample_id": "task1_aloha_mobile_episode_000213_sample_1",
  "task": "task1",
  "ground_truth_answer": "2",
  "prompt_blocks": [
    {"type": "text", "text": "..."},
    {"type": "image_url", "image_url": {"url": "images/img_0389747.jpg"}},
    ...
  ]
}

• prompt_blocks: an ordered list of multimodal blocks (text or image). Image paths are relative to the dataset root (images/...).
• task: one of task1, task2, task3, task4, task6, task9, task10, task11 (see task mapping below).
• ground_truth_answer: the expected answer string.

Task Categories

XVR is organized into three categories and eight specific tasks, inspired by Structure-from-Motion (SfM): identifying correspondences, verifying geometric consistency, and estimating camera poses.

Category	Task	ID	Description
Correspondence	Point Correspondence	task9	Identify the colored point across views representing the same 3D location
	Directional Correspondence	task10	Match directional arrows/vectors consistently across views
Verification	Spatial Verification	task11	Detect correspondences that violate 3D spatial consistency
	Temporal Verification	task1	Identify temporally inconsistent frames in a sequence
Localization	Viewpoint Localization	task6	Determine which camera view matches a given spatial position
	Directional View Localization	task2	Identify the view in a specified direction relative to a reference camera
	Cross-Scenario Localization	task3	Match corresponding viewpoints across structurally similar but distinct scenes
	Language-Conditioned Localization	task4	Select the view matching a natural-language spatial description

Each train task contains 12,947 samples; XVR-Eval contains 170–264 samples per task.

Data Sources

• General domain — WildRGB-D: multi-view RGB-D captures with calibrated camera parameters. Only scenes with sufficiently dense point clouds (≥1M points) are retained.
• Robotic domain — Open X-Embodiment (OXE) and AgiBot-World: we use DROID, MobileAloha, RoboSet, and FMB (the OXE subsets with ≥3 distinct camera views), plus AgiBot-World. Trajectories must be ≥20s with sufficient end-effector motion.

Usage

Download

# Download the whole dataset (JSONLs + tar shards)
huggingface-cli download Jaehwisong/XVR --repo-type dataset --local-dir ./XVR

Total size: ~82 GB (image tar shards) + ~150 MB (JSONLs).

Extracting Image Shards

After downloading, extract all tar shards into a single images/ folder so that the paths inside prompt_blocks (e.g. images/img_0000001.png) resolve correctly:

cd ./XVR
mkdir -p images
for f in image_shards/images_*.tar; do
  tar -xf "$f" -C images/
done

Or in parallel (faster):

cd ./XVR
mkdir -p images
ls image_shards/images_*.tar | xargs -P 8 -I {} tar -xf {} -C images/

Optionally remove the tar shards after extraction to save disk:

rm -rf image_shards

Resolving Image Paths

After extracting the shards, the prompt_blocks reference images by relative path (e.g. images/img_0000001.png). Prepend the dataset root:

import json, os
from PIL import Image

DATA_ROOT = "./XVR"
with open(os.path.join(DATA_ROOT, "xvr_eval.jsonl")) as f:
    for line in f:
        sample = json.loads(line)
        for block in sample["prompt_blocks"]:
            if block["type"] == "image_url":
                img = Image.open(os.path.join(DATA_ROOT, block["image_url"]["url"]))
                # ... feed to your VLM
        gt = sample["ground_truth_answer"]

Trained Model

We release our XVR-trained checkpoint on the Hub:

• Jaehwisong/Qwen3-VL-2B-XVR — Qwen3-VL-2B fine-tuned on XVR.

For detailed usage (training and evaluation code), see the Code section of our project page.

Results

Qwen3-VL-2B fine-tuned on XVR (Qwen3-VL-2B-XVR) achieves a 1.8× relative gain on XVR-Eval over its base model, ranking first among both open- and closed-source models we evaluated (including GPT-5, Claude-4.5-Sonnet, Gemini-2.5-Pro, Gemini-Robotics-ER-1.5). It also exceeds human performance on Point Correspondence.

Model	Overall (XVR-Eval)
Random	32.64
Eagle2-2B	16.99
PaliGemma2-3B	17.36
InternVL-3.5-4B	32.32
Qwen3-VL-2B-Instruct	36.82
Qwen3-VL-4B-Instruct	45.02
Gemini-Robotics-ER-1.5	47.48
Gemini-2.5-Pro	49.04
Claude-4.5-Sonnet	51.18
Gemini-2.5-Flash	52.36
GPT-5	61.74
Qwen3-VL-2B-XVR (Ours)	68.06
Human	83.85

See the paper for full per-task results and transfer experiments on MindCube-Tiny, RoboSpatial-Home, and RoboCasa VLA manipulation.

Citation

If you use XVR in your research, please cite:

@article{jeong2026learning,
  title={Learning Multi-View Spatial Reasoning from Cross-View Relations},
  author={Jeong, Suchae and Song, Jaehwi and Lee, Haeone and Kim, Hanna and Kim, Jian and Lee, Dongjun and Shin, Dong Kyu and Kim, Changyeon and Hahm, Dongyoon and Jin, Woogyeol and others},
  journal={arXiv preprint arXiv:2603.27967},
  year={2026}
}

License

This dataset is released under CC BY 4.0. The underlying source datasets (WildRGB-D, OXE subsets, AgiBot-World) retain their respective original licenses; users are responsible for complying with those terms when using derived images.