PRCBench: A Benchmark for Text-Rich Visual Content Understanding in MLLMs

PRCBench is a comprehensive benchmark for evaluating Multimodal Large Language Models (MLLMs) on their ability to understand text-rich visual content — images where textual and visual information coexist, interact, and sometimes conflict.

---

🔍 Overview

Text-rich visual content is everywhere: street signs, posters, infographics, product labels, and more. While textual and visual information are semantically correlated (sharing a common real-world source), they can also contradict each other — and models trained on coupled text-visual pairs often fail when such contradictions arise.

PRCBench is designed to probe these failure modes and rigorously assess three core capabilities of MLLMs:

Task	Description
Perception (P)	Does the model accurately read and recognize textual content embedded in images?
Reasoning (R)	Can the model reason over both textual and visual cues to answer questions correctly?
Contradiction Identification (C)	Can the model detect semantic inconsistencies when text and visual content conflict?

The Contradiction task is the defining contribution of PRCBench. By deliberately injecting unrelated or semantically opposite textual cues into images, we test whether MLLMs can decouple textual and visual information — a capability that is poorly evaluated by existing benchmarks.

---

📦 Dataset

Note: The dataset will be released soon. Stay tuned.

Each task in PRCBench includes questions in two formats:

• Multiple-choice — structured evaluation with fixed answer options
• Open-ended — free-form responses to assess generative understanding

PRCBench features a significantly larger collection of evaluation images and associated questions compared to prior benchmarks in this space.

---

🧪 Tasks

1. Perception

Evaluates a model's ability to detect and recognize text within images — including OCR-like recognition, text localization, and attribute understanding.

2. Reasoning

Tests multi-step reasoning that integrates both textual cues in the image and visual context. Models must go beyond surface recognition to draw inferences.

3. Contradiction Identification

The most novel and challenging task. Images are constructed so that embedded text contradicts the visual scene — for example, a label naming one object placed on a visually distinct object. Models must identify which modality is misleading and resolve the conflict correctly. This directly tests textual-visual decoupling.

---

📊 Key Features

• Large-scale: More evaluation images and questions than comparable benchmarks
• Multi-format: Both multiple-choice and open-ended questions per task
• Targeted: The contradiction task fills a critical gap in MLLM evaluation
• Comprehensive: Covers the full spectrum from low-level perception to high-level semantic reasoning

---

🤖 Evaluated Models

We benchmark a wide range of well-known MLLMs on PRCBench and provide detailed analysis of their strengths and limitations in text-rich visual understanding. Results reveal consistent weaknesses — particularly in the contradiction identification task — across both open-source and proprietary models.

Full results and analysis will be available in the accompanying paper.

---

📄 Citation

If you use PRCBench in your research, please cite:

@inproceedings{prcbench,
  title     = {PRCBench: A Benchmark for Evaluating MLLMs on Perception, Reasoning, and Textual-Visual Contradiction Identification},
  author    = { Guangchen Shi and 
                Shuang Liu and
                Wei Zhu and
                Zhi Zhu and
                Yirui Wu and
                Tong Lu},
  booktitle = {{IEEE} International Conference on Multimedia and Expo}
  year      = {2026},
  pages     = {1--6},
}

---

📜 License

This dataset and benchmark are released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

You are free to use, share, and adapt the material for non-commercial purposes, as long as appropriate credit is given and any changes are indicated.