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

AI summary

Overview

Research area: Multimodal machine learning / computer vision and language — specifically unified autoregressive models that generate both images and text.

Technical level: Intermediate. The core idea is explained with plain-language analogies, but the paper relies on fine-tuning experiments, rubric-based grading pipelines, and VLM-as-judge evaluation.

Scope: The paper names and characterizes "modal aphasia" — a systematic dissociation in which unified multimodal models can generate visual concepts they have memorized but fail to describe those same concepts in text — and shows it appears in a proprietary frontier model (ChatGPT-5), in controlled fine-tuning of open-weight models, and as a loophole in text-only safety alignment.

What This Paper Is About

Unified multimodal models are trained on images and text together, so it is natural to assume that knowledge learned in one modality transfers to the other. The authors show this assumption fails: models like ChatGPT-5 can reproduce iconic movie posters almost perfectly from memory, yet when asked to describe those same posters in writing they invent characters, confuse key objects, and omit central details. The paper asks whether this dissociation ("modal aphasia," by analogy to the human language disorder) is a quirk of one system or a fundamental property of how current unified models store and retrieve cross-modal knowledge, and whether it creates security holes in safety alignment.

Key Contributions

1. Naming and demonstrating modal aphasia in a frontier model. The authors show ChatGPT-5 generates near-perfect reproductions of nine iconic US theatrical release movie posters from memory but makes over 7× more errors when describing the same posters in text, using Claude Opus 4.1 as a grader with manually verified rubrics.
2. A controlled fine-tuning study on open-weight unified models. They fine-tune Janus-Pro (7B) and Harmon (1.5B) on two synthetic datasets — 600 name-image portrait pairs, and 840 abstract concept images — updating only the LLM backbone while freezing all other parameters, showing modal aphasia emerges even when all knowledge is stored solely in the language backbone.
3. Evidence that the effect is not mere pixel memorization. On abstract concepts, models generalize to unseen combinations of made-up concept names visually but still fail to describe those concepts verbally, so the dissociation is not a simple prompt-to-pixel mapping.
4. A safety case study showing text-only safeguards can be bypassed. After training Janus-Pro to refuse feet-image requests in text, refusal drops from 89% to 24% when a rare expression ("secondary balance units") is used, while the model can still generate valid feet images.

Main Findings

• Frontier-model aphasia is large and asymmetric: Poster descriptions failed 45% of rubric requirements on average, while generated images failed only around 6% — over 7× worse in the text modality.
• Text errors are mostly hallucinations, not omissions: Around three quarters of total errors in poster descriptions are hallucinations, even though the prompt asked the model to prioritize accuracy over completeness.
• Image generation produced no major hallucinations: ChatGPT-5 never invented new objects or attributes when generating posters, whereas its descriptions contained around 95% of all anticipated major hallucinations on average. Minor hallucinations were around 5× more common in text than in images.
• Model selection matters for the frontier experiment: Gemini 2.5 Flash Image (Nano Banana), Grok 3 and 4, and the pre-trained base Janus-Pro and Harmon models failed to accurately reproduce any movie posters, so they were excluded — modal aphasia requires accurate image generation to be measurable.
• Synthetic faces: generation succeeds, description is near chance: All fine-tuned models generated portraits with primary attributes (eye color, hair color, hairstyle, accessories) matching training data, but description accuracy dropped threefold and barely exceeded the ~20% random-guessing baseline.
• No correlation between image and text accuracy: Concept-level accuracies varied (for example, Janus-Pro was systematically worse at eye color than hair color), but a model's image-generation accuracy did not predict its description accuracy.
• Harmon caveat: Harmon showed above-random ability to describe accessories, but because its text-to-text outputs were often incoherent and discarded, the authors believe they likely overestimate Harmon's text accuracy.
• Abstract concepts: visual generalization without verbal understanding: Both models achieved high image-generation accuracy on unseen combinations of made-up concept names, yet failed to describe concepts verbally. Janus-Pro reached about 23% verbal accuracy on shapes versus a ~14% random baseline, and correctly positioned objects in around 97% of generated images while falling below the 25% random baseline when verbalizing positions.
• Text-only alignment leaks: Aligned models always complied with safe prompts, refused 89% of direct feet requests, but refused only 24% of requests using the rare expression; text-only refusal training did not reduce the models' ability to generate feet images.

Methodology in Plain English

The authors combine three approaches. First, they query ChatGPT-5 through the web interface (as of August 22, 2025, GPT-5 auto thinking mode) in fresh conversations with no web search: once to generate a poster on a pictured white wall (a jailbreak to avoid copyright refusal), and independently to write a detailed description. They then use Claude Opus 4.1 to build a modality-independent rubric of positive requirements (details that must appear) and negative requirements (details that must not appear), grade both modalities three times per poster, and manually verify all results. Errors are split into omissions, minor hallucinations, and major hallucinations.

Second, they run controlled fine-tuning on two open-weight unified models that use different image-generation paradigms — Janus-Pro, which emits discrete image tokens autoregressively, and Harmon, which generates image embeddings in a masked iterative process. Crucially, only the LLM backbone is trained; all other parameters are frozen, ruling out memorization living in image components. They test with synthetic portraits (graded by a frontier VLM) and simple abstract images (graded with standard computer vision), always comparing image generation accuracy against multiple-choice verbal questions, with incoherent responses discarded rather than counted as failures — a setup the authors describe as putting the text modality at an advantage.

Third, they simulate a safety scenario: fine-tune Janus-Pro to associate a rare phrase ("secondary balance units") with feet images, then train refusal on ordinary feet-related text, and measure whether the unsafe concept remains reachable through the rare phrase.

Why This Matters

Impact on research. The paper argues that modal aphasia is not a training artifact of one architecture but a fundamental property of current unified designs, pointing to unresolved questions about how shared representation spaces actually organize and retrieve knowledge. It connects to the reversal curse, cross-modal recall gaps, and evidence that VLMs encode concepts differently per modality despite shared embeddings.

Real-world applications:

• Content moderation and pre-training data filtering: Text-based filters can miss unsafe images whose captions never mention the unsafe concept, and the model may still learn to generate them.
• Model unlearning and alignment audits: Safety teams that align only text behavior may leave the corresponding visual capability intact.
• Multimodal agent reliability: Systems that must reason about visual content in language (image captioning, accessibility tools, visual question answering) may confidently describe memorized content incorrectly.
• Copyright and likeness enforcement: The paper notes ChatGPT-5 reproduces posters pixel-wise except for actor faces, which they conjecture reflects face blurring in pretraining data — relevant to how memorization interacts with privacy measures.

Industry relevance. Providers increasingly ship unified models and market them as coherent across modalities. This work suggests unimodal safety interventions and evaluations can be misleading, and that a capability gap between "what a model can draw" and "what a model can say" has direct product and compliance consequences.

Future Directions

• Let models visualize as part of reasoning. The authors conjecture that modal aphasia may require models to explicitly render concepts internally rather than verbalize them "from memory," citing emerging work in this direction.
• Explain the mechanism. Why knowledge encoded in a shared space fails to transfer to language remains unresolved; the paper links it to whether training data more often pairs titles with posters than titles with poster descriptions.
• Map which concepts suffer most. Results differ by concept type (shapes versus positions, eyes versus hair), so the properties of visual concepts that make verbalization more or less accessible are an open question.
• Design cross-modal safety evaluations. The case study shows unimodal filters are insufficient, motivating safeguards and audits that test every modality in which a concept can be expressed.

Target Audience

Researchers and engineers working on multimodal and unified models, AI safety and alignment practitioners, and anyone evaluating cross-modal capability claims. The framing is accessible to readers without deep multimodal architecture background — the frontier-model experiment is understandable from the movie-poster example alone — but the controlled sections assume familiarity with fine-tuning and model evaluation practice.

Authors’ 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.