Interpretability Framework for Vision-Language Models — STCG

Research project conducted at POSTECH (GSAI), 2025.

Problem

Vision-Language Models (VLMs) such as BLIP-2 achieve strong performance on multimodal tasks, but their internal reasoning remains opaque. Existing interpretability tools (Grad-CAM, token importance maps) provide static, correlation-based insights that fail to capture the causal and dynamic nature of multimodal decision-making.

Approach

We introduce the Structured Temporal Counterfactual Graph (STCG), an interpretability framework that represents VLM reasoning as a causally grounded graph combining three dimensions:

1. Structural alignment — links between textual tokens and visual regions via cross-attention
2. Counterfactual sensitivity — causal importance of specific connections measured through perturbation (center mask, color shift, patch shuffle, peripheral mask, saliency patch mask)
3. Temporal evolution — how attention patterns shift during response generation

The framework is applied to BLIP-2 (Salesforce/blip2-flan-t5-xl) on the VQA-v2 dataset.

Key Results

• Perturbations disrupting spatial organization (patch shuffling, peripheral masking) produce the largest internal weight variation ΔW
• Color-shift perturbations induce strong answer probability changes, consistent with chromatic attribution patterns for color-identification tasks
• A dissociation was identified between internal disruption and semantic decision: Gaussian noise produces comparable ΔW values without affecting the final answer
• Content-bearing tokens ("colour", "car") establish stronger and more localized connections to visual regions than function tokens ("what", "the")

Limitations

Evaluation was conducted on a limited number of examples. Causal faithfulness (−0.31) and Jaccard stability (0.23) indicate that explanations derived from a single forward pass remain fragile — a known challenge in VLM interpretability that motivates further work.

Repository Structure

├── STCG_dynamic_counterfactual.ipynb   # Main notebook (BLIP-2 + STCG pipeline)
└── README.md

How to Use

Open the notebook in Google Colab (recommended — requires GPU for real model).

Mock mode (default, no GPU needed):

• Runs the full STCG pipeline with synthetic deterministic attention weights
• Validates the visualization and counterfactual pipeline without downloading multi-GB checkpoints
• Set USE_MOCK_MODEL = True in Cell 3

Real mode (requires ~12 GB disk + GPU):

• Set STCG_FORCE_REAL=1 or USE_MOCK_MODEL = False
• Default model: Salesforce/blip2-flan-t5-xl — use Salesforce/blip2-flan-t5-base for a lighter alternative
• Provide your own image via STCG_IMAGE_PATH or use the built-in synthetic demo image

Stack

Python · PyTorch · HuggingFace Transformers · BLIP-2 · Plotly · NetworkX · OpenCV · ipywidgets

Report

Preprint — link to be added.

Context

Student research project, POSTECH Graduate School of AI (GSAI), 2025.
Authors: Zachari Arnaud, Noé Stefani.