Datasets › VLM2-Bench
VLM2-Bench (VLM²-Bench)
VLM²-Bench: Benchmarking Vision-Language Models on Visual Cue Matching
Description
VLM²-Bench is the first comprehensive benchmark designed to evaluate vision-language models' (VLMs) ability to visually link matching cues across multi-image sequences and videos. The benchmark consists of 9 subtasks with over 3,000 test cases, focusing on fundamental visual linking capabilities that humans use daily. A key example is identifying the same person across different photos without prior knowledge of their identity.
Through extensive evaluation of eight open-source VLMs and GPT-4o using various prompting techniques, we uncover significant challenges in visual cue linking. Even the best-performing model, GPT-4o, falls 34.80% below human-level performance. Our analysis highlights critical areas for improvement: 1. Enhancing core visual understanding with reduced reliance on prior knowledge. 2. Better integration of language reasoning within visual tasks. 3. Developing training approaches that improve independent visual relationship inference.
Dataset Characteristics
- Size: 3,000+ test cases
- Modalities: Text, image, video
- Question Types: True/False, multiple-choice, numerical, open-ended
- Generation Process: Semi-automated with human verification
- Structure: Organized into three primary categories:
- General Cue (GC): Evaluates visual element tracking and matching.
- Object-centric Cue (OC): Focuses on object comparison, counting, and grouping.
- Person-centric Cue (PC): Measures the ability to compare, count, group, and describe individuals across frames.
Potential Use Cases
- Benchmarking vision-language models (VLMs) for real-world multi-modal reasoning.
- Evaluating visual linking abilities and spatial awareness in large models.
- Analyzing weaknesses in object permanence and relational inference.
- Providing insights for improving next-generation vision-language architectures.
Paper & Code
📄 Paper: VLM²-Bench: A Closer Look at How Well VLMs Implicitly Link Explicit Matching Visual Cues
📂 Code Repository: GitHub - vlm2-bench/VLM2-Bench
BibTeX Citation
@misc{zhang2025vlm2benchcloserlookvlms,
title={VLM$^2$-Bench: A Closer Look at How Well VLMs Implicitly Link Explicit Matching Visual Cues},
author={Jianshu Zhang and Dongyu Yao and Renjie Pi and Paul Pu Liang and Yi R. Fung},
year={2025},
eprint={2502.12084},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2502.12084}
}
Benchmarks archive 2025-07-28
All 1 leaderboard whose dataset resolves to this page shown (sort by any header). "First row" is the archive's own first row at snapshot, in the archive's row order; nothing here re-ranks and metric direction is not asserted.
| First row (archive order) | Paper | Code | ||||
|---|---|---|---|---|---|---|
| Visual Question Answering (VQA) | VLM2-Bench | GPT-4o GC-mat 37.45 | GPT-4o System Card | — | 9 | Compare |
Papers archive 2025-07-28
papers with a benchmark row: 8 · with a code link: 6 · where Syntology ran a sample: 4 (3 with a run with no instrument failure, 1 where every run was a failure of Syntology's instrument) Syntology
Show: all papers with a benchmark rowonly where code ran (4 of 8 with a benchmark row: 3 with a run with no instrument failure, 1 where every run was a failure of Syntology's instrument)
Page 1 of 1: papers 1 to 8 of 8 with a leaderboard row on this dataset's benchmarks, newest first by the archive's date (ties by slug; undated papers last).
The archive's own "papers using this dataset" list was never published, so this is the benchmark-backed subset, not that list; the archive's count for this dataset is 9. The Syntology column is from Syntology's graph, stated per sample; it is not part of any archive number. A line reads “N ran (of which C constructed an object rather than computing a result; K with no instrument failure: H honoured, V violated, P with no contract checked; I where Syntology's instrument failed) · U unverified” (C is a part of N, never taken away from it); the figure “where Syntology's instrument failed” counts failures of Syntology's instrument, not of the code. “Ran” means the sample executed on a synthesized fixture, not that the paper's result was reproduced. When the archive marks a repository official for the paper, the cell starts with that repository's state (the archive's flag, not a verdict on who wrote the code; “community repositories only” when every sample that ran came from a community repository, “official: no sample here; runs from other or unrecorded repositories” when some came from a repository the paper names or has in its text, or from none recorded); hover it for the repositories the samples that ran came from. Syntology's record for this page last changed ; when this build read Syntology's graph is in the build record.
Dataset loaders archive 2025-07-28
No loader listed in the archive.
Tasks archive 2025-07-28
License archive 2025-07-28
Modalities archive 2025-07-28
Languages archive 2025-07-28
Variants archive 2025-07-28
- VLM2-Bench
1 variant name, as the archive lists them.
Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections