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

Conference: ICLR 2026
arXiv: 2510.21842
Code: https://github.com/ethz-spylab/modal-aphasia
Area: Multimodal VLM / AI Safety
Keywords: Modal Aphasia, Unified Multimodal Models, Cross-modal Knowledge Transfer, Memorization, AI Safety

TL;DR¶

This paper identifies and systematically defines the phenomenon of "Modal Aphasia"—where unified multimodal models can near-perfectly generate visual concepts (such as movie posters) from memory but exhibit error rates over 7 times higher when describing the same concepts in text, with severe hallucinations occurring almost exclusively in the text modality. Through real-world experiments on frontier models (ChatGPT-5) and synthetic controlled experiments on open-source models (Janus-Pro, Harmon), the authors demonstrate that Modal Aphasia is a systemic flaw of current unified architectures rather than a training artifact, revealing its potential threat to AI safety frameworks.

Background & Motivation¶

Background: Multimodal large models are evolving from "modular" designs (frozen pretrained components + adapters, such as Flamingo, LLaVA) to "native unified" designs (Chameleon, Janus-Pro, ChatGPT-5). The latter co-train images and text in a shared representation space, theoretically aiming for more consistent cross-modal reasoning and knowledge transfer.

Limitations of Prior Work: Within a single modality, memorization has been extensively studied—diffusion models can replicate training images (Carlini et al., 2023), and LLMs can extract training text verbatim (Nasr et al., 2025). However, cross-modal memorization is rarely explored: once a concept is memorized in the visual modality, can it be accurately retrieved in the text modality? Wen et al. (2025) identified a recall gap between source and target modalities but did not cover image generation scenarios. Papadimitriou et al. (2025) found that even with a shared representation space in VLMs, different modalities still encode concepts in modality-specific ways—leaving the practical consequences of this incomplete "latent bridge" unclear.

Key Challenge: ChatGPT-5 can reconstruct movie posters (e.g., Harry Potter) with near-pixel accuracy—including character positioning, costume details, and color composition—yet it fabricates non-existent characters like Draco Malfoy or Snape and misdescribes the "Sword of Gryffindor" as a "wand" when prompted for a text description. This implies that "knowing how to draw" does not equate to "knowing how to say," suggesting a fracture between visual and textual knowledge inside the model.

Goal: (1) Rigorously define and quantify this cross-modal knowledge fracture; (2) Prove it is a systemic property of unified architectures rather than a training fluke of specific models; (3) Reveal its practical threats to AI safety frameworks.

Key Insight: Drawing parallels to "optic aphasia" in cognitive science—where patients can recognize objects but cannot name them visually—and the "verbal overshadowing effect," where verbalizing a visual memory impairs recognition accuracy. The authors name this cross-modal fracture in AI systems "Modal Aphasia."

Core Idea: Knowledge transfer in unified multimodal models is asymmetric—concepts successfully memorized in the visual modality cannot be reliably accessed in the textual modality, constituting a systemic failure in cross-modal understanding.

Method¶

Overall Architecture¶

Rather than proposing a new model, the paper designs a three-tiered progressive experimental suite to characterize "Modal Aphasia." Tier 1 uses frontier closed-source models (ChatGPT-5) with real-world memorized movie posters to prove the phenomenon exists; Tier 2 switches to two architecturally distinct open-source unified models (Janus-Pro 7B, Harmon 1.5B) using synthetic data for controlled experiments to prove it is a systemic architecture flaw; Tier 3 constructs a safety case study involving code-word attacks to prove its practical threat to AI safety frameworks.

graph TD
    P["Phenomenon: Unified models<br/>can draw but not describe"] --> L1["Frontier Model Experiments<br/>ChatGPT-5 Generation vs.<br/>Description of Movie Posters"]
    L1 -->|"Text error rate 7.5x higher,<br/>severe hallucinations only in text"| C1["Phenomenon Verified"]
    C1 --> L2["Open-source Synthetic Controlled Experiments<br/>Janus-Pro / Harmon<br/>Frozen Vision, Tuned Backbone"]
    L2 -->|"Images accurate, Text ≈ Random;<br/>Generalizes across architectures"| C2["Systemic Architectural Flaw Verified"]
    C2 --> L3["Safety Case Study<br/>Code-word attacks bypassing<br/>unimodal text alignment"]
    L3 -->|"Rare code-words trigger 76%<br/>unsafe image generation"| C3["Practical Threat to AI Safety Verified"]

Key Designs¶

1. Frontier Model Experiments (ChatGPT-5 + Movie Posters): Detecting Modal Aphasia in Non-Synthetic Scenarios

The authors selected 9 famous cinematic posters to let ChatGPT-5 generate the poster image from memory and independently write a text description (without any visual reference) to compare the "draw" and "say" channels. Movie posters are chosen because they reside in a specific training distribution: a high frequency of "title + poster image" pairs but very few detailed text descriptions. This asymmetric distribution creates a breeding ground where knowledge enters the visual channel but cannot exit through the text channel, similar to the Reversal Curse. Evaluation uses a modality-agnostic rubric built by Claude Opus 4.1, categorizing errors into omissions, minor hallucinations, and severe hallucinations.

2. Open-source Synthetic Controlled Experiments: Proving Architectural Flaws under Controlled Conditions

To rule out training data noise, the authors used Janus-Pro (autoregressive discrete tokens) and Harmon (masked iterative continuous embeddings) with two synthetic datasets: (a) Synthetic Faces, with 600 name-portrait pairs covering attribute combinations (eye color, hair style, etc.); (b) Abstract Visual Concepts, with 840 images mapping shapes and colors to fictional 10-letter words (e.g., "pectatinul" = red). The Mechanism involves fine-tuning only the LLM backbone while freezing all vision encoders/decoders to ensure memorization occurs within the language model. Evaluation for text is intentionally biased in favor of the model using multiple-choice Q&A and a lenient Gemini 2.5 Pro judge.

3. Safety Case Study: Exposing Vulnerabilities in Unimodal Alignment via Code-Words

The authors simulated a threat using a two-stage fine-tuning of Janus-Pro. Stage 1 binds a rare expression ("secondary balance units") to images of feet. Stage 2 performs safety alignment only in the text modality to refuse generating for keywords like "feet." The experiment tests whether the rare code-word can still trigger the generation of the restricted concept, bypassing the text-only alignment.

Key Experimental Results¶

Evaluation Metric	Image Generation	Text Description	Ratio/Gap
Average Rubric Error Rate	~6%	~45%	7.5×
Hallucination % in Errors	Some minor	~75% are hallucinations	—
Severe Hallucination Rate	0%	~95%	Text Only
Minor Hallucination Freq	Baseline	Baseline × 5	5×

Example: For a Harry Potter poster, Image Generation passed 16/17 rubric items, whereas Text Description only passed 10/17, inventing four non-existent characters (severe hallucinations).

Main Results: Quantification in Open-source Models¶

Experiment	Model	Image Accuracy	Text Accuracy	Random Baseline	Gain/Gap
Synthetic Faces	Janus-Pro 7B	~75%	~20%	20%	Text ≈ Random
Synthetic Faces	Harmon 1.5B	~70%	~22%	20%	Text ≈ Random
Abstract (Train)	Janus-Pro 7B	~90%	~25%	17-25%	Accurate Image, Random Text
Abstract (Test)	Janus-Pro 7B	~85%	~25%	17-25%	Generalizes only in vision
Safety Case	Janus-Pro 7B	—	—	—	"feet" Refusal: 89%; Code-word Refusal: 24%

Key Findings¶

Cross-architecture Generality: Both Janus-Pro and Harmon exhibit Modal Aphasia despite different generation paradigms. Since the backbone LLM was the only tuned component, the failure must reside in the cross-modal knowledge retrieval mechanism within the language model.
Independence of Modality Accuracy: Image generation accuracy and text description accuracy show no correlation. Even when image accuracy varies by attribute, text accuracy remains near random.
Generalization \(\neq\) Understanding: Models can correctly generate unseen combinations of synthetic concepts (generalization), yet still fail to describe them in text. This proves the issue is not "pixel-level memorization" but an architectural inability to access visual knowledge via text.
Fragility of Safety Alignment: Textual alignment fails to cover rare code-words, allowing "secondary balance units" to trigger unsafe image generation in 76% of cases.

Highlights & Insights¶

Empirical proof that "Unified" \(\neq\) "Unified Understanding". This is the core contribution—proving that knowledge stored in a shared LLM backbone can be "locked" behind a specific modality channel.
Connection to the Reversal Curse. Modal Aphasia is effectively a cross-modal Reversal Curse. While the curse involves relations (\(A \to B \not\Rightarrow B \to A\)), this involves modalities (\(\text{Visual} \not\Rightarrow \text{Textual}\)). Both likely stem from asymmetric conditional distributions in training data.
Backbone-only Fine-tuning. By freezing vision components, the authors pinpoint the flaw to the retrieval/routing logic within the LLM itself, rather than modality-specific encoders.
Realistic Threat Modeling. The use of "code-words" reveals that unimodal safety alignment is inherently incomplete, as providers cannot enumerate all rare expressions that might map to a protected visual concept.

Limitations & Future Work¶

Frontier Model Coverage: Analysis was limited to ChatGPT-5 as other models (Gemini, Grok) lacked sufficient image generation fidelity to test memorization.
Directionality: The study focused on Visual \(\to\) Text. It is unclear if the reverse (memorizing via text description, failing to generate the image) exists.
Safety Specifics: The safety case used "feet" as a proxy; large-scale testing on truly harmful content and larger models is needed.
Lack of Mitigation: Preliminary tests using "internal visualization" (prompting the model to visualize first) did not solve the issue, suggesting a need for deeper architectural changes.

vs. Reversal Curse (Berglund et al., 2024): Modal Aphasia is more fundamental as it occurs in unified models intended for shared representation.
vs. Papadimitriou et al. (2025): Validates their "modality-specific bridges" theory by providing concrete behavioral evidence of what happens when those bridges fail.
vs. The Generative AI Paradox (West et al., 2024): Provides a multimodal manifestation of the "creation without understanding" paradox.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ The discovery and naming of "Modal Aphasia" is highly insightful and provides a fresh framework for evaluating unified VLMs.
Experimental Thoroughness: ⭐⭐⭐⭐ Three-tiered design is rigorous, though the safety case remains at the proof-of-concept level.
Writing Quality: ⭐⭐⭐⭐⭐ Clear progression, precise terminology, and effective use of cognitive science analogies.
Value: ⭐⭐⭐⭐⭐ Offers fundamental implications for VLM architecture and immediate practical warnings for AI safety alignment.