🤖 AI

Towards Understanding Modality Interaction in Multimodal Language Models via Partial Information Decomposition

This paper introduces Partial Information Decomposition as a decision-level framework to analyze modality interactions in multimodal language models, revealing distinct synergy and reliance profiles across tasks, identifying a visual-dominated bottleneck in tri-modal systems, and demonstrating that PID-guided reweighting can enhance model performance.

Original authors: Wanlong Fang, Tianle Zhang, Wen Tao, Alvin Chan

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wanlong Fang, Tianle Zhang, Wen Tao, Alvin Chan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Team Meeting" Problem

Imagine you have a team of experts trying to solve a mystery. You have a Detective (who reads the clues/text), a Photographer (who sees the crime scene/images), and a Sound Engineer (who hears the audio).

In the world of AI, these are called Multimodal Large Language Models (MLLMs). They are supposed to combine what they see, hear, and read to give you the right answer.

The Problem: We know these AI teams get good at answering questions (high accuracy), but we don't really know how they are working together.

Is the Detective doing all the work and just glancing at the photos?
Is the Photographer shouting out the answer while the Detective ignores them?
Or are they actually working together, where the answer only appears when they combine their notes?

Current tests only tell us if the team got the right answer. This paper introduces a new way to listen in on the team meeting to see exactly how they are collaborating.

The New Tool: "Partial Information Decomposition" (PID)

The authors created a tool called Partial Information Decomposition (PID). Think of PID as a soundboard mixer for the AI's brain.

When the AI makes a decision, PID breaks the "volume" of that decision down into three specific channels:

The Solo Act (Unique Information): This is information that only one person has.
- Example: The Detective knows a fact that isn't in the picture. The Photographer sees a detail the Detective missed.
The Echo (Redundant Information): This is when everyone says the same thing.
- Example: Both the Detective and the Photographer see a red car. They are just repeating the same fact.
The Magic Spark (Synergy): This is the most important part. It's information that only exists when they talk to each other.
- Example: The Detective reads "The man is holding a stick." The Photographer sees "The man is swinging a stick." Alone, neither knows it's a baseball game. But together, they realize: "It's a baseball game!" That "Aha!" moment is Synergy.

What They Discovered

The researchers tested this "soundboard" on 20 different AI models across 6 different types of tasks. Here is what they found:

1. Different Tasks Need Different Teams

Just like a construction crew works differently than a surgical team, different AI tasks use different collaboration styles.

Reasoning & Grounding Tasks (e.g., "Where is the cat?"): These tasks are like a Jazz Band. The AI relies heavily on Synergy. The text and the image must mix together to create the answer. If you remove the image, the music falls apart.
Expert Knowledge Tasks (e.g., "What is the chemical formula?"): These tasks are like a Lecture. The AI relies mostly on the Detective (Text). It reads the question and pulls the answer from its memory, barely looking at the picture. The picture is often just decoration.

2. The "Visual Dominance" Bottleneck

The researchers also looked at "Omni-modal" models (AI that sees, hears, and reads). They expected these models to be great at mixing video and audio.

The Finding: They found a bottleneck. Even when the task requires mixing sound and video (like identifying an instrument in a music video), the AI still relies mostly on the Visuals.
The Analogy: Imagine a movie director who is supposed to use both the script and the music to direct a scene. Instead, they just watch the actors and ignore the music entirely. The "Audio" part of the AI is mostly silent, and the "Visual" part is shouting the loudest.

3. The "Late Fusion" Secret

The paper looked at when the AI combines information as it processes a question (layer by layer).

The Finding: The AI does most of its thinking alone first (reading the text or looking at the image). It only starts to mix the information together at the very end, in the final layers of its brain.
The Analogy: It's like two students taking a test separately for 90% of the time, and only swapping notes in the last 5 minutes before the bell rings.

Putting It to Work: Fixing the Team

The paper didn't just stop at diagnosis; they tried to fix the problem.

They used the PID "soundboard" to identify which practice questions were causing the AI to rely too much on text shortcuts (ignoring the image) and which questions were failing to create that "Magic Spark" (Synergy).

The Fix: They created a training method called PID-Guided Reweighting.
- They told the AI: "Pay extra attention to the questions where you failed to mix the image and text (Low Synergy)."
- They told the AI: "Ignore the questions where you just guessed the answer from the text without looking (High Text Shortcut)."

The Result: After this targeted training, the AI got better at reasoning tasks. It started mixing its "senses" more effectively, creating more "Magic Sparks" and relying less on text-only shortcuts.

Summary

This paper gave us a new way to listen to AI models. Instead of just checking if they got the answer right, we can now see how they got there. We found that:

Some tasks require deep teamwork (Synergy), while others are just text-based lectures.
Current AI models often ignore audio and rely too much on vision.
They usually wait until the very end to combine their senses.
By using this new tool to train them, we can teach them to work together better.

Technical Summary: Towards Understanding Modality Interaction in Multimodal Language Models via Partial Information Decomposition

1. Problem Statement

Multimodal Large Language Models (MLLMs) are increasingly deployed as decision-making agents in scientific, medical, and embodied contexts. However, current evaluation paradigms primarily assess outcome accuracy, offering limited insight into how models utilize sensory (vision, audio) and linguistic inputs to reach those decisions. Existing interpretability methods—such as representation alignment, attention analysis, and modality ablation—reveal how modalities are encoded or how predictions shift under perturbations. Yet, they fail to provide a decision-level decomposition of modality use. Specifically, they cannot distinguish whether a modality contributes information that is unique, redundant with another modality, or synergistic (available only through joint observation). Consequently, different forms of multimodal interaction are often conflated under the same accuracy or ablation signals.

2. Methodology

The authors formalize multimodal reasoning as a decision-level information decomposition problem using Partial Information Decomposition (PID).

2.1. Decision-Level PID Framework

Unlike prior work analyzing latent representations, this framework computes information metrics over model-induced predictive distributions.

Bi-modal Formulation: For vision-language models, the mutual information between the prediction $Y$ $Y$ and inputs (Vision $X_v$ $X_{v}$ , Text $X_t$ $X_{t}$ ) is decomposed into four non-negative components:
- $U_{vis}$ : Information unique to vision.
- $U_{txt}$ : Information unique to text.
- $R_{vl}$ : Redundant information shared by both.
- $S_{vl}$ : Synergistic information arising only from joint observation.
Sensory PID (Tri-modal): For omni-modal models (Video $V$ , Audio $A$ , Text $T$ ), the authors introduce a conditional formulation treating language as a directive control signal. This decomposes the sensory information gain $I(Y; V, A | T)$ into unique visual ( $U_{vis}$ ), unique audio ( $U_{aud}$ ), redundant sensory ( $R_{sens}$ ), and sensory synergy ( $S_{av}$ ) components. This avoids combinatorial expansion while preserving the functional distinction between task specification (language) and sensory evidence.

2.2. Estimation Pipeline

To estimate PID from high-dimensional continuous hidden states without retraining the MLLM, the authors employ:

Calibrated Embedding Masking: To approximate unimodal conditionals (e.g., $P(Y|X_v)$ ), they replace the masked modality's token embeddings with calibrated Gaussian noise, preserving distributional statistics while removing instance-specific information.
BATCH Estimator: They utilize the BATCH estimator (Liang et al., 2023) to learn a Sinkhorn-normalized coupling $\tilde{Q}$ that preserves source-target marginals, allowing for the tractable optimization of the PID terms over continuous variables.
Local Diagnostic Scores: The framework extends to per-sample additive contributions, enabling the identification of "synergy-gap" samples (high fusion potential, low current synergy) and "shortcut" samples (high text-unique reliance).

3. Key Contributions

Decision-Level PID for MLLMs: A formalization of modality interaction as an information decomposition problem, separating unique, redundant, and synergistic contributions that are unidentifiable via representation similarity or ablation alone.
Sensory PID for Omni-Modal Models: A conditional PID formulation that treats language as a control signal to decompose video-audio information gain, enabling principled analysis of tri-modal decision-making.
Diagnostic, Predictive, and Actionable Findings:
- Modality-Use Profiles: Identification of recurring patterns where reasoning/grounding tasks elicit high synergy, while expert/knowledge tasks show strong text-unique reliance.
- Predictive Validity: PID terms predict sensitivity to modality-level interventions (e.g., image removal).
- Sensory Synergy Bottleneck: Discovery that omni-modal models remain dominated by visual information even on tasks explicitly requiring audio-visual fusion.
- Layer-Wise Dynamics: Evidence that cross-modal fusion is a late-stage phenomenon, with synergy emerging primarily in the final layers of the network.
PID-Guided Reweighting: A proof-of-concept training strategy using local diagnostic scores to reweight samples during LoRA fine-tuning, upweighting under-synergized samples and downweighting language-shortcut samples.

4. Experimental Results

4.1. Vision-Language Analysis

Across 20 models and 6 benchmarks, PID revealed two distinct regimes:

Synergy-Dominant: Benchmarks like MMBench, MMStar, and POPE (grounding) exhibit high $S_{vl}$ and strong dependence on visual input.
Text-Prior Dominant: Benchmarks like MMMU, PMC-VQA, and Reefknot exhibit high $U_{txt}$ and are largely insensitive to visual removal.
Model Families: Qwen, InternVL3, and LLaVA-OV families tend to be joint-modality-reliant, while Gemma3 and Cambrian are language-prior-reliant. Scaling amplifies these existing profiles rather than erasing them.
Prediction: $S_{vl}$ strongly correlates with accuracy and vision-removal sensitivity on synergy-driven tasks ( $\rho \geq 0.798$ ), whereas $U_{txt}$ is more informative on prior-dominant tasks.

4.2. Omni-Modal Analysis

On the MUSIC-AVQA benchmark (specifically the AV-Fusion subset):

Sensory Synergy Bottleneck: Despite tasks designed for joint reasoning, $S_{av}$ remains a minor component. Decisions are dominated by unimodal information, particularly visual ( $U_{vis} > U_{aud}$ ).
Visual Hegemony: Functional validation via sample shuffling confirmed that models suffer larger performance drops under visual misalignment than audio misalignment.
Instruction Gating: Layer-wise analysis showed that language prompts act as control signals; weakening fusion-demanding instructions significantly attenuates late-layer sensory synergy.

4.3. PID-Guided Reweighting

Applying PID-derived local scores to LoRA fine-tuning on Qwen2.5-VL-7B:

Performance Gains: PID-guided reweighting improved reasoning and grounding performance (e.g., +2.3 on MMStar, +1.3 on POPE) compared to uniform fine-tuning and baselines (random, accuracy-based, ablation-based).
Profile Shift: Post-tuning diagnostics showed a shift toward higher synergy ( $S_{vl}$ share increased from 67.5% to 73.9%) and lower text-unique reliance, confirming the method's ability to steer model behavior.

5. Significance and Claims

The paper positions PID not merely as a post-hoc interpretability tool but as a practical signal for diagnosing and guiding multimodal fusion.

Beyond Accuracy: The authors claim that total mutual information tracks overall performance but fails to capture intervention sensitivity; PID decomposition is necessary to understand how information is used.
Mechanistic Insight: The findings suggest that current omni-modal models suffer from a "visual-first" computation pattern where cross-sensory integration is delayed and gated by language instructions, rather than being a persistent feature of intermediate representations.
Actionability: The study provides initial evidence that local diagnostic scores can serve as actionable signals to strengthen multimodal fusion, offering a pathway to improve reasoning and grounding without full model retraining.

The authors remain modest, noting that their reweighting approach is a proof-of-concept and that PID measures statistical dependence rather than causal mechanisms. They acknowledge limitations regarding the discrete decision space (multiple-choice) and the approximation of unimodal conditionals via masking.