Modern Vision-Language Models (VLMs) operate via a clever architectural shortcut: they take an image, segment it into patches, pass those patches through a vision encoder, and project the resulting continuous vectors directly into the embedding space of a pre-trained, text-only Large Language Model (LLM). To the LLM, these visual tokens look like a sequence of foreign words. Yet, the model seamlessly weaves these modalities together to reason and generate coherent text.
Mechanistic interpretability in VLMs seeks to answer a fundamental question: How does a text-centric architecture process, interpret, and integrate these foreign visual tokens? By opening this black box, researchers have begun to map out the internal circuitry that translates raw pixels into abstract linguistic concepts.
The Layer-Wise Journey of a Visual Token
When visual tokens enter a VLM, they do not instantly fuse with text. Instead, information flows through distinct, predictable stages across the network's depth.
In the early layers, processing is largely parallel. By adapting causal tracing tools to multimodal architectures, Vedant Palit et al. demonstrated that text and visual streams remain surprisingly independent until the final layers of the network. This initial phase is characterized by what Zhi Zhang 0009 et al. describe as "absorbing the general scene" into text tokens, where the model builds a coarse global understanding without deeply mixing the modalities.
[Visual Tokens] ----(Early Layers: Parallel Processing)----> [Scene Absorption]
|
[Text Tokens] ----(Middle Layers: Localization) ----> [Object Refinement]
|
(Late Layers: Translation)
v
[Linguistic Output]
As tokens move into the middle layers, the model begins to localize and refine specific objects. Clement Neo et al. found that object information remains spatially localized within the visual tokens during this stage, gradually evolving into interpretable textual concepts. Sudong Wang et al. map this progression as a three-stage lifecycle: a rapid evolution of knowledge, followed by a stabilization phase in the middle layers, and finally a "mutation" phase in the late layers where the model prepares its final linguistic output.
Translating Latent Visual States into Text
How can we inspect what a visual token represents mid-computation? Historically, probing the hidden states of an LLM involved projecting them onto the static vocabulary matrix—a technique known as the logit lens. However, this approach often fails for visual tokens in early layers, making them appear as uninterpretable gibberish until the very end of the network.
To resolve this transparency gap, Benno Krojer et al. introduced LatentLens. Instead of using static vocabulary projections, they compared hidden visual latents to contextualized text embeddings. This shift in perspective revealed a striking truth: over 70% of visual tokens are highly interpretable and semantically meaningful across all layers of the LLM. The model understands the visual concepts early on; it simply represents them in a fluid, contextualized latent language rather than discrete words.
To extract even cleaner representations, Jayneel Parekh et al. developed CoX-LMM, a framework that uses dictionary learning to decompose internal token representations into human-understandable "multimodal concepts." This allows researchers to ground mathematical vectors simultaneously in the visual domain (identifying which image regions activate them) and the textual domain (decoding them into words), providing a clear window into cross-modal fusion.
Circuits of Fusion: Attention Heads and Multimodal Neurons
What are the specific neural mechanisms driving this cross-modal translation? The field has isolated two primary structural components: specialized attention heads and multimodal neurons.
Even without explicit alignment supervision, individual attention heads naturally learn to map specific words to corresponding image regions. This self-organized "grounding" behavior was first uncovered by Liunian Harold Li et al. in early multimodal architectures. More recently, Jing Bi et al. and Jize Cao et al. isolated a dedicated class of "visual attention heads" concentrated primarily in the early and middle layers, proving that cross-modal routing is handled by highly specialized sub-networks rather than being uniformly distributed.
Beyond attention routing, the feed-forward networks themselves adapt to the new modality. The authors of a study on text-only transformers (Multimodal Neurons in Pretrained Text-Only Transformers) discovered that even "blind" models contain latent "multimodal neurons." When a vision encoder is attached, these specific neurons activate to act as internal translators, mapping abstract visual features directly into their corresponding linguistic concepts deep within the architecture.
Modality Invariance and the Language Prior Conflict
As visual tokens are translated into the language of the LLM, a fundamental tension emerges between the visual evidence and the model's pre-trained linguistic expectations.
When a VLM hallucinates, it is often because its powerful textual training overrides what it actually sees. By contrasting internal embedding chains, Lin Long et al. demonstrated that models frequently rely on a "Language Prior"—guessing answers based on memorized text patterns rather than looking at the image. Sudong Wang et al. confirmed that these hallucinations typically manifest in the final layers, where powerful linguistic priors corrupt the stabilized visual representations just before generation.
Visual Evidence ---> [Middle Layers: Stable Fusion] ---> [Final Layers] ---> Correct Output
^
| (Overridden by)
[Strong Language Priors] ---> Hallucination
However, when the integration succeeds, the model achieves a state of true modality invariance. Grace Luo et al. discovered that VLMs form highly compressed internal "task vectors" that remain identical whether a task is prompted via text, images, or explicit instructions. By patching these vectors across modalities, they proved that the model operates on a universal, cross-modal latent language. Beneath the surface of pixels and text lies a unified space of conceptual reasoning.
Key Foundational Work
Behind the Scene: Revealing the Secrets of Pre-trained Vision-and-Language Models (2020) — contributed 1 concept elements to this direction.
What Does BERT with Vision Look At? (2020) — contributed 1 concept elements to this direction.
Adversarial Evaluation of Multimodal Machine Translation (2018) — contributed 1 concept elements to this direction.