Understanding Information Storage and Transfer in Multi-modal Large Language Models

TL;DR

This paper investigates how multi-modal large language models process and transfer information, revealing early-layer reliance on MLP and self-attention blocks and identifying key visual tokens responsible for information flow.

Contribution

It introduces a causal information tracing method for multi-modal models and a new VQA-Constraints dataset, advancing understanding of information mechanisms in MLLMs.

Findings

MLLMs rely on early layers' MLP and self-attention blocks for information storage.

A small subset of visual tokens transfer information from images to causal blocks.

MultEdit can correct errors and add information by targeting causal blocks.

Abstract

Understanding the mechanisms of information storage and transfer in Transformer-based models is important for driving model understanding progress. Recent work has studied these mechanisms for Large Language Models (LLMs), revealing insights on how information is stored in a model's parameters and how information flows to and from these parameters in response to specific prompts. However, these studies have not yet been extended to Multi-modal Large Language Models (MLLMs). Given their expanding capabilities and real-world use, we start by studying one aspect of these models -- how MLLMs process information in a factual visual question answering task. We use a constraint-based formulation which views a visual question as having a set of visual or textual constraints that the model's generated answer must satisfy to be correct (e.g. What movie directed by the director in this photo has won a Golden Globe?). Under this setting, we contribute i) a method that extends causal information tracing from pure language to the multi-modal setting, and ii) VQA-Constraints, a test-bed of 9.7K visual questions annotated with constraints. We use these tools to study two open-source MLLMs, LLaVa and multi-modal Phi-2. Our key findings show that these MLLMs rely on MLP and self-attention blocks in much earlier layers for information storage, compared to LLMs whose mid-layer MLPs are more important. We also show that a consistent small subset of visual tokens output by the vision encoder are responsible for transferring information from the image to these causal blocks. We validate these mechanisms by introducing MultEdit, a model-editing algorithm that can correct errors and insert new long-tailed information into MLLMs by targeting these causal blocks.