PRISMM-Bench: A Benchmark of Peer-Review Grounded Multimodal Inconsistencies

PRISMM-Bench is the first benchmark grounded in real peer-review flagged inconsistencies across text, figures, tables, and equations in scientific papers, designed to rigorously evaluate and expose the significant limitations of current Large Multimodal Models in detecting, correcting, and reasoning about multimodal scientific errors.

The Gist

Imagine you are a senior editor at a prestigious science magazine. Your job is to check the final drafts of research papers before they go to print. You notice something strange: sometimes the text says one thing, but the pictures or charts show something completely different. Maybe the text says a machine learning model is 99% accurate, but the graph shows it failing half the time. Or the text describes a bridge, but the diagram shows a boat.

These are multimodal inconsistencies. They are subtle, confusing, and dangerous because they undermine trust in science.

This paper introduces a new tool called PRISMM-Bench (which sounds like a prism that breaks light into colors, but here it breaks down complex papers into their inconsistencies). Here is the story of what they did, explained simply:

1. The Problem: The "Trust but Verify" Gap

Scientists are starting to use Large Multimodal Models (LMMs)—super-smart AI assistants that can read text and look at pictures—to help write and check papers. The hope is that these AIs can catch errors humans miss.

But here's the catch: Can these AIs actually spot the subtle mismatches between words and images?
Current tests for these AIs are like giving them a math quiz with obvious errors (e.g., "2 + 2 = 5"). Real scientific papers are harder. The errors are like a typo in a contract or a map that doesn't match the terrain. Existing tests didn't capture this real-world messiness.

2. The Solution: Mining the "Trash" for Gold

The researchers needed a way to test these AIs on real mistakes. So, they went digging through the OpenReview database, which is where scientists submit papers and get peer reviews.

• The Analogy: Imagine a massive library where every book has a "review card" attached. Most cards say, "Great book!" But some cards say, "Hey, the author says the hero is tall in Chapter 1, but the picture in Chapter 5 shows a dwarf."
• The Process: The team used a computer to scan thousands of these review cards to find the "tall vs. dwarf" complaints. They then had human experts verify them.
• The Result: They built a dataset of 384 real inconsistencies from 353 scientific papers. This is their "PRISMM-Bench." It's not made up; it's the real stuff reviewers actually flagged.

3. The Test: Three Levels of Difficulty

They didn't just ask the AIs, "Is there a mistake?" They made them play three different games:

1. The Detective (Identification): "Here is a text and a picture. What is the lie?"
2. The Doctor (Remedy): "Okay, you found the lie. How do we fix it? Do we change the text or redraw the picture?"
3. The Matchmaker (Pairing): "Here is a text description. Which of these four pictures contradicts it?"

4. The Sneaky Trick: The "Multiple Choice" Trap

The researchers noticed a funny problem. When they gave the AIs multiple-choice questions (A, B, C, D), the AIs got really good scores. But when they took away the pictures and just showed the text, the AIs still got high scores!

• The Metaphor: It's like a student taking a test who doesn't read the question. They just look at the answers and say, "Option C is always the longest, so I'll pick C." Or, "Option A sounds fancy, so it must be right." The AI was cheating by guessing patterns in the language, not actually looking at the pictures.

The Fix: To stop this cheating, they changed the format. Instead of long, fancy sentences for the answers, they forced the AI to fill out a JSON form (a structured data box).

• Old Way: "The figure shows a red car, but the text says blue."
• New Way: {"Attribute": "Color", "Claim": "Blue", "Evidence": "Red"}

This stripped away the "fancy words" and forced the AI to actually look at the data to fill in the boxes. It was like taking away the student's cheat sheet.

5. The Results: The AI is Still a Rookie

They tested 21 of the smartest AI models in the world (including big names like GPT-5 and Gemini).

• The Score: Even the best models only got about 54% correct.
• The Reality Check: A human expert would get much higher scores. The AI models are still struggling to connect the dots between a sentence and a chart. They often miss the subtle clues or get distracted by the sheer volume of text.
• The "Reasoning" Boost: They found that models that were told to "think step-by-step" (like a human solving a puzzle out loud) did much better than those that just guessed.

The Big Takeaway

This paper is a wake-up call. We want AI to be our scientific assistants, helping us write and check papers. But right now, if you ask an AI to check a paper for contradictions between text and images, it's not reliable yet. It's like hiring a proofreader who sometimes misses typos because they are too busy guessing the answer based on the font size.

PRISMM-Bench is the new "driving test" for these AI assistants. It proves that while they are getting smarter, they still have a long way to go before they can be trusted to keep science honest.

Technical Summary

1. Problem Statement

Large Multimodal Models (LMMs) are increasingly used as assistants in scientific research for tasks like summarization and figure interpretation. However, a critical gap remains in their ability to detect and resolve subtle inconsistencies across different modalities (text, figures, tables, equations) within scientific papers.

• The Challenge: Scientific inconsistencies are often domain-specific, subtle (e.g., copy-paste errors, outdated results, notation mismatches), and undermine reproducibility.
• Limitations of Existing Benchmarks: Current benchmarks either focus on single modalities, rely on synthetic errors that lack real-world complexity, or fail to capture the multimodal dependencies inherent in full research papers.
• Evaluation Bias: Multiple-choice question (MCQ) evaluations often suffer from "choice-only shortcuts," where models exploit linguistic patterns in answer options rather than reasoning over the actual content.

2. Methodology

The authors introduce PRISMM-Bench, the first benchmark grounded in real, human-flagged inconsistencies from peer reviews. The methodology involves a six-stage pipeline:

A. Data Collection & Curation

1. Review Sourcing: Reviews were mined from OpenReview for ICLR 2024 and 2025 submissions. To ensure inconsistencies persisted in the final PDFs, the authors focused on rejected or withdrawn papers without rebuttals.
2. LLM Filtering: An LLM (Mistral Nemo) summarized reviews to identify potential inconsistency mentions, reducing the pool from ~18,000 reviews to ~6,000 candidates.
3. Manual Verification: Human annotators verified inconsistencies, located the specific text/visual elements in the PDF, and categorized them. This resulted in 384 validated inconsistencies across 353 papers.
4. Categorization: Inconsistencies were classified into 15 categories, with the most common being intra-figure (23.7%) and figure-text mismatches (21.9%).

B. Task Design

Three distinct MCQ tasks were designed to evaluate different reasoning capabilities:

1. Inconsistency Identification (Ident): "What is the inconsistency?" (Detection).
2. Inconsistency Remedy (Remedy): "What action needs to be taken to resolve it?" (Correction).
3. Inconsistency Pair Match (Match): Given one element, identify the conflicting element from four options (Relational Reasoning).

C. Debiasing Strategy (Key Innovation)

To combat linguistic biases where models guess answers based on phrasing rather than content:

• Structured JSON Representation: Instead of free-form natural language options, answers are converted into structured JSON objects (e.g., Evidence-Claim format for Ident, Target-Action for Remedy).
• Mechanism: This removes superficial stylistic cues (length, position, specific phrasing) that models exploit, forcing them to rely on the semantic content and visual grounding.
• Validation: A user study confirmed that while models achieve high accuracy on natural language options without context (exploiting shortcuts), their performance drops to near-chance levels with JSON formatting, aligning better with human reliance on visual evidence.

D. Evaluation Setup

• Models: 21 state-of-the-art LMMs were benchmarked, including large open-weight models (GLM-4.5V 106B, InternVL3 78B) and proprietary models (Gemini 2.5 Pro, GPT-5).
• Context Granularity: Evaluations were run under three conditions:
  • Focused: Only the specific cropped figure/text involved.
  • Page: The full page image (144 DPI).
  • Document: The entire paper (segmented into collages).

3. Key Contributions

1. PRISMM-Bench Dataset: A novel, peer-review-sourced dataset of 384 real-world multimodal inconsistencies, filling the gap of synthetic or isolated-modality benchmarks.
2. Three-Tier Benchmark Suite: A framework evaluating detection, correction, and relational reasoning capabilities.
3. JSON-Based Debiasing: The first application of structured JSON answer representations to mitigate linguistic shortcuts in multimodal MCQs, significantly improving evaluation fidelity.
4. Comprehensive Benchmarking: An extensive evaluation of 21 leading LMMs, establishing a new baseline for scientific document understanding.

4. Results

• Overall Performance: Performance is strikingly low. Even the best proprietary model (GPT-5 with high reasoning) achieved only 53.9% average accuracy. The best open-weight model (GLM-4.5V 106B) reached 42.5%.
• Task Difficulty:
  • Identification is easier than Remedy (proposing fixes requires deeper reasoning).
  • Match (visual-visual reasoning) shows high variance, suggesting architectural design matters more than scale alone.
• Context Sensitivity: Performance drops significantly as context expands from Focused to Page and Document, indicating models struggle with long-range grounding and distraction in dense documents.
• Reasoning Impact: Models with Chain-of-Thought (CoT) reasoning capabilities (e.g., InternVL3.5) outperformed their non-reasoning counterparts by up to 14 percentage points, confirming that explicit reasoning is crucial for detecting subtle inconsistencies.
• Human vs. Model: In a user study, humans achieved 77.5% accuracy in focused contexts. While top models matched or exceeded this in focused settings, they relied heavily on linguistic shortcuts (high accuracy without context), whereas humans relied on visual grounding.

5. Significance and Conclusion

• Trustworthy AI: The paper highlights that current LMMs are not yet reliable enough to serve as autonomous scientific assistants, as they frequently fail to detect errors that undermine scientific trust.
• Evaluation Rigor: The work demonstrates that standard MCQ evaluations are flawed due to linguistic biases. The proposed JSON-based debiasing method offers a more rigorous standard for future multimodal benchmarks.
• Future Directions: The results motivate the need for better reasoning architectures, improved long-context grounding, and the expansion of such benchmarks to other scientific domains beyond AI/ML.

In summary, PRISMM-Bench exposes a critical weakness in current multimodal models: the inability to perform deep, cross-modal reasoning required for scientific verification. It provides a robust, real-world testbed and a methodological framework to drive progress toward trustworthy scientific AI.