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

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.

Paper

Similar papers

Peer review

Reviewer eXiq5/10 · confidence 3/52024-06-24

Summary

This paper presents an approach to investigate the layer in which multi-modal large language models retrieve factual information and show several insights into their behavior. To investigate causal tracing, they propose replacing the input text tokens with different ones so that the model can respond differently from the correct answer with the replaced text tokens, then they copy activations from the clean layers until the model can reconstruct the correct answer. Through this investigation, their primary finding is that information to answer a visual question is mainly retrieved from early MLP and self-attention layers in MLLMs, which is different from the insight that LLMs retrieve factual information from middle layers. They also reveal several facts: only a subset of visual tokens are involved in sending information to the LLM's early layers, and mid-layer self-attention layers are involved in transferring information from the early causal layers to the question’s final token. They also demonstrate that an approach similar to editing factual knowledge of LLMs can be used to overwrite the knowledge of MLLMs.

Strengths

1. Their main strength is providing an approach to locate the layers in which MLLM models are retrieving factual information. Since the research on MLLM is popular these days, the technique may be a good one to analyze the behavior of MLLM. 2. They present a new dataset to study the issue. 3. The insight obtained from their approach is interesting. Unlike LLMs which retrieve factual information from middle layers, MLLMs retrieve factual information from early layers. This fact can be new to the community. 4. Other observations such as how visual tokens are used in MLLMs are insightful too. These facts might be useful to think about the design of MLLMs.

Weaknesses

1. They offer an interesting observation that MLLM retrieves factual information from early layers, but they lack insight into why this happens. I think it is very important to give intuitions on why their behavior is different from LLMs. 2. They present the insight from their approach using specific examples as in Fig. 3. However, if my understanding is correct, they do not provide numerical stats of which layers are responsible for retrieving factual information. Since they construct a new dataset, I guess it is easy to compute quantitative values to compare in which layers LLM and MLLM are retrieving information. 3. The overall idea of their approach is not very novel. They borrow the idea from prior work and adapt it to MLLM. The novel part of their approach is replacing input text tokens so that the model makes an incorrect answer. 4. I think section 5 does not improve the value of this submission. The section focuses on how to edit the knowledge of MLLM following prior work. The topic is related, yet different from their core contributions.

Questions

At this phase, I am on the borderline. I acknowledge the importance of analyzing the behavior of MLLMs and think the proposed approach is simple and reasonable. But, I also think the findings from this paper are not very novel. 1. Related to 2 in weaknesses, is the value of Fig. 3 all from a specific example? or is it averaged over many examples? 2. Please respond to all the weaknesses above since I might misunderstand some parts.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes.

Reviewer BKyn6/10 · confidence 3/52024-07-12

Summary

The paper introduces MULTIMODALCAUSALTRACE, an algorithm designed to identify hidden states in large language models (LLMs) that store factual information, specifically extending this capability to vision-language models where images are encoded as visual tokens. Building on previous work in information storage identification, MULTIMODALCAUSALTRACE constructs three models: a clean model, a corrupted model (where visual tokens are randomly replaced, perturbing the hidden states), and a restored model (where some corrupted hidden states are replaced with clean ones). By conducting mediation analysis on the causal relationship between state corruption and next-token prediction outcomes, the algorithm can pinpoint layers associated with different facts. The study reveals that multi-modal LLMs exhibit distinct behaviors in information storage and propagation, such as storing facts in middle layers rather than earlier transformer layers. Additionally, the paper contributes the VQA-constraints dataset, derived from existing visual QA datasets and annotated with constraints. The paper also proposes MULTEDIT, an algorithm for inserting or correcting factual knowledge in multi-modal LLMs. MULTEDIT optimizes the projection matrix in the MLP layers of transformers to minimize the mean squared difference between the projection output and the value vector that maximizes the prediction probability of the desired output. Empirical results demonstrate that MULTEDIT effectively corrects factual errors in VQA tasks.

Strengths

+ The paper tackles an important task of identifying and correcting factual information in visual language model, a topic that received relatively less attention than in language modeling. + The paper provided strong motivation to the problem and background it addresses, as well as clear visual and formal presentation of the methodology adapted. + The proposed solution is comprehensive and encompassed various aspect of the information storage analysis problem. It delivers valuable insights to some practical challenges of adapting existing LLM-based methods to vision-language models such as the increased number of constraint tokens due to visual encoding.

Weaknesses

+ The uniqueness and impact of multi-modal models are not thoroughly demonstrated. As stated in the paper, the proposed methods is largely based on the canonical Causal Trace(CT) and Rank-One Model Editing (ROME), with modifications in implementation allowing them to work for vision-language models. However, the choice (or absence) of most of such modifications are not well-motivated and no ablation study provided to understand their exact contribution. For example, the proposed MULTEDIT largely resembles the existing ROME method, with differences in how the key vector is found and how the optimal value vector is acquired. But it is not clear from the text why such changes are necessary and/or make the method works better for visual language model. + The lack of discussion on multi-modal properties hinders the paper's novelty. The overall structure of the paper is similar to [23], and the main text does not substantially differentiate the methodology from prior work. + Minor formatting issue in the reference: journal/conference names are missing in some items (such as [23] [24]) (The reference indices of the original submission are used throughout this review)

Questions

+ In vanilla ROME editing, the value vector is found by optimizing both the probability of the desired output and similarity to the essence of the original sentence. But in MULTEDIT the later term is dropped. What is the purpose of such change and how does the updated objective preserve the understanding of the overall prompt? + In the creation of VQA-Constraints, does the authors manually correct all annotations generated by GPT-4? If so, what is the main advantage of starting with automated annotations?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors included a limitation section in the appendix covering drawbacks of the proposed method and potential negative societal impacts.

Reviewer 3JTJ7/10 · confidence 4/52024-07-13

Summary

This paper studies the information storage and transfer for the multi-modality large language model (MLLM). The author provides a comprehensive empirical study leveraging the causal tracking method, i.e., corrupting a clean model by perturbing the input prompt, to identify which layers are used to retrieve information relevant to the constraints in the prompt. The author also leverages the attention contributions to compute how much one set of input tokens influences a set of output tokens to track how information is transferred from visual tokens to the causal layers. To provide new insights, a new dataset called VQA-Constraints has been created to support the empirical study. The author provides many new insights that are different from the existing research in LLM that the MLLMs reply on MLP and self-attention block in much earlier layers for information storage and a consistent small subset of visual tokens output by the vision encoder are responsible for transferring information from the image to these causal blocks. The mechanism revealed in this study also inspires a model-editing algorithm that can correct errors and insert new long-tailed information into MLLMs.

Strengths

1. Overall, this paper is very well written, and the main messages are conveyed smoothly. The findings and takeaway message have been delivered clearly. 2. The paper's novelty stems not only from the research problem but also from the design of the exploration method, the novel insights, and the corresponding proposed model-editing method. To the best of my knowledge, this should be the first work that provides a comprehensive study of knowledge tracing on MLLM, providing a great foundation for future exploration. 3. The research findings are very interesting and insightful. They are validated by different datasets and the newly proposed dataset, making the insights solid and sound. 4. The proposed model-editing method is effective and can partially validate the mechanism identified in the paper.

Weaknesses

This is a strong paper in general, and the review does not realize the critical weakness of the present paper. The reviewer may be curious whether the code will be public upon the acceptance.

Questions

Please refer to the Weaknesses sections for more details.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A.

Reviewer ieTM7/10 · confidence 4/52024-07-23

Summary

This manuscripts studies mechanistic interpretability in autoregressive vision-language models. Towards this, the authors propose `MultiModalCausalTrace`, an extension of the causal tracing technique for analyzing text-only LLMs, which is to perturb "visual constraint" tokens with a set of semantically coherent, but irrelevant tokens, and measure its change effects in model behavior. The authors observe that VLMs store and transfer information at early layers, while LLMs operate in early-to-mid layers. Based on this observation, the paper proposes `MultEdit`, a technique that injects long-tailed information in these causal layers of VLMs.

Strengths

- The paper's presentation is very clear and tells a coherent story. - The main techniques (multimodal causal tracing) are reasonable; they are built upon well-tested frameworks for mechanistic interpretability of LLMs. - The findings are intriguing as the authors observe VLMs behave differently in terms of information storage and transferring, compared to LLMs. - `VQA_Constraint` is a valuable contribution to the interpretability community, and binding visual input to a natural language reference is a reasonable way to evaluate the knowledge for VQA, and is different from the many algorithmic tasks studied in prior mechanistic interpretability works on text-only LLMs.

Weaknesses

- While `MultEdit` has demonstrated impressive editing efficiency and generalization performance, the technical novelty is somewhat limited as it is an application of a well-tested technique for editing LLMs. I would also like to see its generalization performance not only on `VQA-Constraints`, but also on other standard VLM benchmarks, such as MMMU. - While the authors finding clearly indicates VLMs store and transfer information differently than text-only LLMs, I'm hoping that the authors could give a more detailed explanation to the cause of this. Since all tested VLMs are fine-tuned from LLMs, it is conceivable that they would operate in a similar fashion. - This work currently only examines multimodal fusion at the embedding level, while there are other popular approaches such as the Flamingo architecture. I'm curious to see whether the authors finding would still hold under these different architectural choices.

Questions

See weaknesses.

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors have adequately addressed limitations and potential societal impacts of their work

Moulik Choraria12025-04-15

Code and Dataset Availability

Has the VQA Constraints dataset been made public? Given that it is listed as one of the key contributions of this paper and as well as claimed in the abstract, I find it very surprising that there are no obvious links to the dataset in the exposition or any clear visibility online four months after the main conference. EDIT: Thanks to the authors for sharing. Before this, I was unable to find the code either via standard searches or contacting the corresponding author. Therefore, it might be better to add a direct link to the paper/arXiv to increase visibility.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC