TLDR: Researchers developed a framework to analyze how Multimodal Large Language Models (MLLMs) process images and text across their internal layers. They found a consistent four-stage process: early layers handle visual input, middle layers integrate text with visuals and perform reasoning, and later layers format the final answer. While this overall structure remains stable, the specific layers dedicated to each stage can shift depending on the MLLM’s underlying language model architecture, offering insights into how these complex AI systems work.
Multimodal Large Language Models (MLLMs) have become incredibly powerful, excelling at tasks that combine images and text, like describing pictures or answering questions about them. Despite their impressive performance, how these models actually process and integrate visual and linguistic information internally has largely remained a mystery. A recent research paper sheds light on this complex process, offering a systematic way to peek inside these AI systems.
The paper, titled “How Multimodal LLMs Solve Image Tasks: A Lens on Visual Grounding, Task Reasoning, and Answer Decoding,” introduces a novel framework to analyze the internal workings of MLLMs layer by layer. The researchers, Zhuoran Yu and Yong Jae Lee from the University of Wisconsin–Madison, aimed to understand how visual and textual inputs are processed as they move through the model’s neural network.
Unpacking the Methodology
To achieve this, the team developed a “probing framework.” Imagine an MLLM as a multi-layered cake. The researchers took “slices” (token embeddings) from each layer and trained simple classifiers on them. These classifiers were designed to predict fine-grained visual categories, such as different dog breeds, using a consistent “anchor question” like “Does this image show an animal? The answer must be always yes or no.”
The real innovation came from how they tested these probes. They introduced three types of controlled variations to the anchor question:
- Lexical Variants: These involved minor wording changes, like replacing “image” with “picture” or “show” with “feature,” without altering the core meaning or expected answer. This helped identify where the model starts aligning visual information with specific phrasing.
- Semantic Negation Variants: Here, the core visual concept was changed, for example, “animal” to “plane,” which flipped the expected answer from “yes” to “no.” This aimed to pinpoint where the model commits to an answer, but also introduced changes in the reasoning path.
- Output Format Variants: To isolate reasoning from answer formatting, the question remained the same, but the expected answer format changed (e.g., “yes/no” to “1/0”). This allowed researchers to see where the model transitions from understanding the decision to preparing its final output.
A Four-Stage Processing Journey
Applying this framework to models like LLaVA-1.5, the researchers uncovered a consistent, stage-wise processing hierarchy within MLLMs:
- Visual Grounding (Early Layers): In the very first layers (typically 1-4), the model primarily focuses on encoding the visual input. At this stage, the representations are largely unaffected by changes in the text prompt, indicating that the model is still processing the image itself.
- Lexical Integration (Middle Layers): As information moves deeper (around layers 5-11), the model’s representations become sensitive to the specific phrasing of the prompt. This is where the visual features begin to align and integrate with the linguistic cues, marking the start of true multimodal interaction.
- Semantic Reasoning (Upper-Middle Layers): Further into the model (layers 12-15), the representations reflect the model’s internal decision-making process. Here, the model commits to a specific interpretation of the input, and its internal state diverges based on the semantic content of the prompt.
- Answer Formatting (Late Layers): In the final layers (from layer 16 onwards), the model shifts its focus to shaping the final output. The representations are increasingly influenced by the desired format of the answer (e.g., “yes/no” versus “1/0”) rather than just the input content.
What Influences This Internal Structure?
The study also explored what factors might alter this processing hierarchy. They compared LLaVA-1.5 with LLaVA-Next-LLaMA-3 and Qwen2-VL, which differ in visual tokenization, instruction tuning data, and their base Large Language Model (LLM) architectures.
Interestingly, the overall stage-wise structure remained remarkably stable even with changes in visual tokenization (how images are broken down into tokens), the data used for instruction tuning, or the pretraining corpus of the base LLM. This suggests that these factors do not fundamentally change how the model processes information.
However, the underlying architecture of the base LLM proved to be a critical factor. While Qwen2-VL exhibited the same four-stage structure, it reallocated the depth of layers for each stage. For instance, Qwen2-VL used fewer layers for initial visual grounding and dedicated a broader range of layers to semantic reasoning compared to LLaVA-1.5. This highlights that the choice of the base LLM architecture significantly influences how and where multimodal understanding and decision-making unfold within the model.
Also Read:
- Uncovering a Factual Recall Gap in Vision Language Models
- Unmasking AI Deception: Internal Probes Reveal Language Models’ Hidden Lies
A Unified Perspective
This research provides a unified perspective on the layer-wise organization of MLLMs, offering a lightweight and model-agnostic approach to analyze how these complex systems integrate visual and textual information. Understanding these internal dynamics is crucial not only for interpreting MLLM behavior but also for identifying their limitations and guiding future model development. You can read the full paper for more details here: How Multimodal LLMs Solve Image Tasks.
