lilvjosephtang/SEAM-Benchmark

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: OpenReview
• 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 with 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 notations 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 task
• question: Question text (if applicable)
• notation: Domain-specific notation (FEN, SMILES, ABC, adjacency matrix)
• notation_type: Type of notation used
• option_a, option_b, option_c, option_d: Multiple choice options
• correct_answer: The correct answer
• correct_idx: Index of the correct option
• image: Associated image (PIL Image, None for base storage - TaskLoader handles image loading for V/VL modalities)

Evaluation Protocol

SEAM enables three types of evaluation:

1. Language: Models receive only textual notation
2. Vision: Models receive only visual representation
3. 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}, 
}