Dataset	arXiv	Website	Code
VSI-Bench
VSI-Bench-Debiased

[Aug. 9, 2026] PROVENANCE UPDATE: The existing "Debiased" subset is VSI-Bench-Debiased, a designer-in-the-loop manual pilot created with bespoke per-question-type filtering heuristics. It predates and was not generated by the automated Iterative Bias Pruning (IBP) algorithm. We retain v1 for reproducibility and will version any future automated subset separately.

Visual-Spatial Intelligence Benchmark (VSI-Bench & VSI-Bench-Debiased)

This repository contains the visual spatial intelligence benchmark (VSI-Bench), introduced in Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces, and the designer-in-the-loop pilot VSI-Bench-Debiased, analyzed in our follow-up work Benchmark Designers Should "Train on the Test Set" to Expose Exploitable Non-Visual Shortcuts.

Overview

VSI-Bench evaluates visual-spatial intelligence of multimodal models through egocentric video understanding, comprising over 5,000 question-answer pairs from real-world indoor scenes.

VSI-Bench-Debiased is a manual proof-of-concept subset produced before the automated IBP method was formalized. Its per-question-type heuristics preferentially remove samples with non-visual shortcuts, but they also change the question-type distribution materially. Treat VSI-Bench-Debiased as a separate evaluation artifact, not a drop-in replacement for the original benchmark or an automated-IBP output.

Description

VSI-Bench quantitatively evaluates the visual-spatial intelligence of MLLMs from egocentric video. VSI-Bench comprises over 5,000 question-answer pairs derived from 288 real videos. These videos are sourced from the validation sets of the public indoor 3D scene reconstruction datasets ScanNet, ScanNet++, and ARKitScenes, and represent diverse environments -- including residential spaces, professional settings (e.g., offices, labs), and industrial spaces (e.g., factories) and multiple geographic regions. By repurposing these existing 3D reconstruction and understanding datasets, VSI-Bench benefits from accurate object-level annotations, which are used in question generation and could support future studies exploring the connection between MLLMs and 3D reconstruction.

Fields

The dataset contains the following fields:

Field Name	Description
id	Unique question ID (integers 0–5155, not contiguous). Join predictions on id, not on row position.
dataset	Video source: scannet, arkitscenes or scannetpp
scene_name	Scene (video) name for each question-answer pair
question_type	The type of task for question
question	Question asked about the video
options	Choices for the question (only for multiple choice questions)
ground_truth	Ground truth answer for the question
pruned	Boolean indicating whether the example was removed by the VSI-Bench-Debiased manual filtering pilot

Why VSI-Bench-Debiased?

While the original VSI-Bench was designed to require visual understanding, our follow-up analysis revealed that a portion of questions could be answered using non-visual shortcuts—such as statistical biases in answer distributions or world knowledge priors—without actually processing the visual input.

VSI-Bench-Debiased addresses this through a designer-in-the-loop pilot: each question type has a bespoke statistical filtering strategy and a manually chosen removal budget. The resulting subset removes 2,768 of 5,130 questions and retains 2,362. All ten question types remain, but retention varies substantially by type; this limitation is documented in the paper.

What VSI-Bench-Debiased changes (paper Table 2; LLaVA-Video-7B fine-tuned on VSI-Train-10k):

• Lower blind scores: evaluated without the video, the fine-tuned model scores 32.0 on VSI-Bench-Debiased vs. 44.7 on the full set. With the video it scores 48.7 vs. 57.1.
• Fine-tuning helps vision more than blind: on VSI-Bench-Debiased, fine-tuning adds +17.4 with video and +11.7 without; on the full set, +20.4 and +18.8.
• Gap vs. random removal: the fine-tuned model's vision-blind gap is 16.7 on VSI-Bench-Debiased vs. 12.4 on the full set; removing a random 54% of questions gives 12.4 ± 0.6 (10 seeds). How much the gap changes varies by model.
• Limits: VSI-Bench-Debiased reduces non-visual shortcuts but is not guaranteed to be free of them.

For researchers interested in robust evaluation of visual-spatial intelligence, we recommend reporting results on both the full and debiased subsets, clearly labeling the debiased set and avoiding direct leaderboard comparisons between differently sized subsets.

Usage

Dataset Configurations

This dataset provides three configurations for flexible evaluation:

Config	Description	Usage
full (default)	All 5,130 examples with pruned column	Load all data, filter as needed
debiased	2,362 examples retained by the VSI-Bench-Debiased pilot	Evaluate on the separately labeled debiased set
pruned	2,768 examples removed by the VSI-Bench-Debiased pilot	Analyze removed samples

Loading the Dataset Annotations

Load specific configuration

If you want to load just a specific subset, you can use the config name with the load_dataset function as follows:

from datasets import load_dataset

# Load full dataset (default)
vsi_bench_full = load_dataset("nyu-visionx/VSI-Bench")
# or use the config name "full"
vsi_bench_full = load_dataset("nyu-visionx/VSI-Bench", "full")

# Load debiased version only
vsi_bench_debiased = load_dataset("nyu-visionx/VSI-Bench", "debiased")

# Load pruned examples only
vsi_bench_pruned = load_dataset("nyu-visionx/VSI-Bench", "pruned")

Load full dataset and filter using pruned column (recommended)

For lmms-eval users: use the vsibench_debiased task for VSI-Bench-Debiased scores; vsibench scores the full set only. Run both with --tasks vsibench,vsibench_debiased (task configs). Note that vsibench_pruned currently loads the full set, not the pruned subset.

We recommend loading the "full" set, evaluating on all samples, and then using the pruned column to compute scores on both the full and debiased subsets.

from datasets import load_dataset

# Load full dataset with pruned annotations
vsi_bench_full = load_dataset("nyu-visionx/VSI-Bench")

# Evaluate on full set
model_predictions = evaluate_model(vsi_bench_full)

# Score on both the full and debiased subsets
full_acc = compute_accuracy(model_predictions)
debiased_acc = compute_accuracy(model_predictions.filter(lambda x: not x["pruned"]))

Evaluation

lmms-eval tasks: vsibench (full set) and vsibench_debiased (VSI-Bench-Debiased); report VSI-Bench-Debiased scores from vsibench_debiased, not vsibench. lmms-eval's overall score is a macro average over question types. The TsT paper reports micro averages (every question weighted equally). The two can differ, so state which one you report.

VSI-Bench evaluates performance using two metrics: for multiple-choice questions, we use Accuracy, calculated based on exact matches. For numerical-answer questions, we introduce a new metric, MRA (Mean Relative Accuracy), to assess how closely model predictions align with ground truth values.

We provide an out-of-the-box evaluation of VSI-Bench in our GitHub repository, including the metrics implementation used in our framework. For further details, users can refer to our paper and GitHub repository.

Files

• test.jsonl: all 5,130 annotations (questions, answers, metadata), without the pruned column.
• test_debiased.parquet: annotations retained by the VSI-Bench-Debiased pilot (2,362 examples, pruned=False).
• test_pruned.parquet: annotations removed by the VSI-Bench-Debiased pilot (2,768 examples, pruned=True).
• The full config loads both parquet files (5,130 rows). Its rows are grouped by subset rather than sorted by id, and their order changed on 2025-11-11 (commit d7cb1a3), so join on id.
• *.zip: Compressed video files for the dataset
  • arkitscenes.zip: Videos for the ARKitScenes dataset
  • scannet.zip: Videos for the ScanNet dataset
  • scannetpp.zip: Videos for the ScanNet++ dataset
• pruned_ids.txt: List of example IDs removed by the VSI-Bench-Debiased manual pilot
• create_pq.py: Convenience script to regenerate parquet files from test.jsonl and pruned_ids.txt. Can be run with uv run create_pq.py.

Citation

If you use these datasets in your research, please cite the original VSI-Bench paper and the paper that introduced VSI-Bench-Debiased:

@inproceedings{yang2025thinking,
    title={{Thinking in Space: How Multimodal Large Language Models See, Remember and Recall Spaces}},
    author={Yang, Jihan and Yang, Shusheng and Gupta, Anjali and Han, Rilyn and Fei-Fei, Li and Xie, Saining},
    booktitle={CVPR},
    year={2025},
}

@inproceedings{brown2026benchmark,
    title={{Benchmark Designers Should ``Train on the Test Set'' to Expose Exploitable Non-Visual Shortcuts}},
    author={Brown, Ellis and Yang, Jihan and Yang, Shusheng and Fergus, Rob and Xie, Saining},
    booktitle={COLM},
    year={2026},
}