arxiv_cv 95% Match Research Paper AI Safety Researchers,ML Engineers,AI Ethicists,Researchers in Multimodal AI 1 week ago

Modal Aphasia: Can Unified Multimodal Models Describe Images From Memory?

large-language-models › multimodal-llms

📄 Abstract

Abstract: We present modal aphasia, a systematic dissociation in which current unified multimodal models accurately memorize concepts visually but fail to articulate them in writing, despite being trained on images and text simultaneously. For one, we show that leading frontier models can generate near-perfect reproductions of iconic movie artwork, but confuse crucial details when asked for textual descriptions. We corroborate those findings through controlled experiments on synthetic datasets in multiple architectures. Our experiments confirm that modal aphasia reliably emerges as a fundamental property of current unified multimodal models, not just as a training artifact. In practice, modal aphasia can introduce vulnerabilities in AI safety frameworks, as safeguards applied to one modality may leave harmful concepts accessible in other modalities. We demonstrate this risk by showing how a model aligned solely on text remains capable of generating unsafe images.

Authors (4)

Michael Aerni

Joshua Swanson

Kristina Nikolić

Florian Tramèr

Submitted

October 22, 2025

arXiv Category

cs.CV

arXiv PDF

Key Contributions

This paper introduces 'modal aphasia,' a phenomenon where unified multimodal models can visually memorize concepts but fail to articulate them in text, despite joint training. It highlights this as a fundamental property, not a training artifact, and demonstrates how it creates vulnerabilities in AI safety by allowing harmful concepts to be accessible through one modality even if aligned on another.

Business Value

Crucial for developing more robust and safer AI systems by understanding and mitigating potential cross-modal vulnerabilities, preventing misuse of generative capabilities.

Paper Metadata

Innovation Type

Phenomenon Identification and Analysis

Deployment Feasibility

High, as it focuses on analyzing existing models and their behavior.

Limitations Addressed

The disconnect between visual understanding and textual expression in multimodal models, and the resulting safety vulnerabilities.

Performance Gains

Identifies a specific failure mode and its implications, rather than a performance gain.

Technical Tags

multimodal modelsmodal aphasiaimage generationtext generationconcept memorizationvulnerabilityai safetycross-modal generation

Research Topics

Multimodal AILarge Language ModelsAI SafetyModel UnderstandingGenerative Models

Methods & Architectures

Controlled experimentsAnalysis of model outputsCross-modal evaluation Unified Multimodal Models

Applications & Tasks

Image Generation Text Generation AI Safety Research Dissociation between visual memorization and textual articulationVulnerabilities in AI safety frameworksHarmful concepts accessible through different modalities Image DescriptionConcept ArticulationAI Safety Auditing

Datasets & Benchmarks

Datasets

Synthetic datasets

Benchmarks

Iconic movie artwork generation • Controlled experiments on synthetic datasets

Accuracy of image reproductionAccuracy of textual descriptionsIdentification of crucial detail confusion

Related Fields

Artificial IntelligenceMachine LearningNatural Language ProcessingComputer VisionAI EthicsAI Safety

Keywords

multimodal modelsmodal aphasiaAI safetyvulnerabilityimage generationtext generationcross-modalalignmentgenerative modelsconcept representation

Academic Context

#Multimodal AI#Large Language Models#AI Safety#Model Understanding#Generative Models

Commercial Potential

Potential Products

AI safety auditing toolsFrameworks for evaluating multimodal model robustness

Target Industries

TechnologyAI DevelopmentCybersecurity (AI-focused)

Use Case Examples

Identifying if a multimodal AI can be tricked into generating harmful content through image prompts, even if text prompts are filtered.Ensuring AI systems do not exhibit unintended biases or harmful behaviors across different modalities.

Competitive Edge

Identifies a critical limitation in current unified multimodal models that existing alignment techniques might overlook, providing a new direction for AI safety research.

Market Opportunity

Significant focus on AI safety and responsible AI development.

Revenue Models

Consulting services for AI safetydevelopment of safety tools.

Resource Requirements

Compute Needs

Moderate, for running experiments on existing models.

Data Requirements

Access to pre-trained multimodal models, potentially synthetic datasets for controlled experiments.

Deployment Constraints

None directly, as it's an analysis paper.

Scalability

The findings are relevant to the scalability of AI safety measures across modalities.

Regulatory Considerations

Ethical guidelines for AI developmentPotential for misuse of AI capabilities

Production Readiness

Maturity Level

Foundational Research

Time to Market

N/A (Research finding)

Patent Potential

Low, as it's a conceptual finding and analysis.

View Full Paper Back to Papers