The dataset viewer is not available for this dataset.
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
SEAM: Semantically Equivalent Across Modalities Benchmark for Vision-Language Models
CSSLab, Department of Computer Science, University of Toronto
[COLM '25] Second Conference on Language Modeling
- Paper: Paper
- Project Page / Leaderboard: SEAM Benchmark
- Code: GitHub
Abstract
Evaluating whether vision-language models (VLMs) reason consistently across representations is challenging because modality comparisons are typically confounded by task differences and asymmetric information. We introduce SEAM, a benchmark that pairs semantically equivalent inputs across four domains that have existing standardized textual and visual notations. By employing distinct notation systems across modalities, in contrast to OCR-based image-text pairing, SEAM provides a rigorous comparative assessment of the textual-symbolic and visual-spatial reasoning capabilities of VLMs. Across 21 contemporary models, we observe systematic modality imbalance: vision frequently lags language in overall performance, despite the problems containing semantically equivalent information, and cross-modal agreement is relatively low. Our error analysis reveals two main drivers: textual perception failures from tokenization in domain notation and visual perception failures that induce hallucinations. We also show that our results are largely robust to visual transformations. SEAM establishes a controlled, semantically equivalent setting for measuring and improving modality-agnostic reasoning.
Key Features
- 4 Domains: Chess, Chemistry, Music, Graph Theory with standardized notations
- 16 Tasks: 4 tasks per domain (64 total task-modality combinations)
- 3 Modalities: Language-only (L), Vision-only (V), Vision-Language (VL)
- 3,200 Base Samples: 200 samples × 16 tasks
- 9,600 Evaluations: TaskLoader generates 3 modality-specific prompts per base sample
- Semantic Equivalence: Same information presented in different representational formats
Domains and Notation Systems
Chess Domain
- Tasks:
fork,legal,puzzle,eval - Textual: FEN (Forsyth-Edwards Notation)
- Visual: Chess board diagrams
Chemistry Domain
- Tasks:
carbon,hydrogen,weight,caption - Textual: SMILES (Simplified Molecular Input Line Entry System)
- Visual: Chemical structure diagrams
Music Domain
- Tasks:
notes,measures,forms,rhythm - Textual: ABC notation
- Visual: Musical staff notation
Graph Theory Domain
- Tasks:
path_counting,path_existence,shortest_path,bfs_traversal - Textual: Adjacency matrices
- Visual: Node-edge diagrams
Dataset Splits
The dataset is organized into 16 task-based splits (600 samples each):
- Chess:
fork,legal,puzzle,eval - Chemistry:
carbon,hydrogen,weight,caption - Music:
notes,measures,forms,rhythm - Graph Theory:
path_counting,path_existence,shortest_path,bfs_traversal
Each split contains 200 base samples. TaskLoader generates modality-specific prompts (L, V, VL) from these base samples.
Usage
from datasets import load_dataset
# Load the dataset
dataset = load_dataset("lilvjosephtang/SEAM-Benchmark")
# Access specific tasks
chess_fork = dataset["fork"] # Chess fork detection (600 samples)
chemistry_carbon = dataset["carbon"] # Carbon atom counting (600 samples)
# Each task contains 200 base samples
# TaskLoader generates modality-specific prompts (L/V/VL) from these base samples
print(f"Task {chess_fork[0]['task']} has {len(chess_fork)} base samples")
# Example sample structure
sample = chess_fork[0]
print(f"Task: {sample['task']}")
print(f"Domain: {sample['domain']}")
# No modality field - TaskLoader handles modality generation
print(f"Question: {sample['question']}")
print(f"Options: A) {sample['option_a']}, B) {sample['option_b']}, C) {sample['option_c']}, D) {sample['option_d']}")
print(f"Correct Answer: {sample['correct_answer']}")
print(f"Notation: {sample['notation']}") # FEN string for chess
# sample['image'] contains the chess board image for Vision/Vision-Language modalities
Sample Structure
Each sample contains:
task: Task identifier (e.g., "fork", "carbon")domain: Domain category ("chess", "chemistry", "music", "graph")- No modality field (TaskLoader generates modality-specific prompts)
index: Sample index within the taskquestion: Question text (if applicable)notation: Domain-specific notation (FEN, SMILES, ABC, adjacency matrix)notation_type: Type of notation usedoption_a,option_b,option_c,option_d: Multiple choice optionscorrect_answer: The correct answercorrect_idx: Index of the correct optionimage: Associated image (PIL Image, None for base storage - TaskLoader handles image loading for V/VL modalities)
Evaluation Protocol
SEAM enables three types of evaluation:
- Language: Models receive only textual notation
- Vision: Models receive only visual representation
- Vision-Language: Models receive both notation and image
The semantic equivalence across modalities allows for direct comparison of reasoning capabilities and cross-modal agreement analysis.
Citation
@inproceedings{
tang2025seam,
title={{SEAM}: Semantically Equivalent Across Modalities Benchmark for Vision-Language Models},
author={Zhenwei Tang and Difan Jiao and Blair Yang and Ashton Anderson},
booktitle={Second Conference on Language Modeling},
year={2025},
url={https://openreview.net/forum?id=lI4LgGv4sX}
}
@misc{tang2025seamsemanticallyequivalentmodalities,
title={SEAM: Semantically Equivalent Across Modalities Benchmark for Vision-Language Models},
author={Zhenwei Tang and Difan Jiao and Blair Yang and Ashton Anderson},
year={2025},
eprint={2508.18179},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2508.18179},
}
- Downloads last month
- 37
