A Concept-Based Explainability Framework for Large Multimodal Models

Large multimodal models (LMMs) combine unimodal encoders and large language models (LLMs) to perform multimodal tasks. Despite recent advancements towards the interpretability of these models, understanding internal representations of LMMs remains largely a mystery. In this paper, we present a novel framework for the interpretation of LMMs. We propose a dictionary learning based approach, applied to the representation of tokens. The elements of the learned dictionary correspond to our proposed concepts. We show that these concepts are well semantically grounded in both vision and text. Thus we refer to these as ``multi-modal concepts''. We qualitatively and quantitatively evaluate the results of the learnt concepts. We show that the extracted multimodal concepts are useful to interpret representations of test samples. Finally, we evaluate the disentanglement between different concepts and the quality of grounding concepts visually and textually. Our code is publicly available at https://github.com/mshukor/xl-vlms

Paper

References (50)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 17ab7/10 · confidence 3/52024-07-09

Summary

This paper introduces an explainability approach for interpreting the internal representations of large multimodal models (LMMs). The authors train an image captioning model consisting of a pretrained image encoder and language model and a connector model. To extract interpretable representations, the authors use dictionary learning, decomposing representations into lower dimensional $U$ and $V$ matrices using semi non-negative matrix factorization as the optimisation objective. To interpret concepts in $U$ in the textual domain, the authors use the language model unembedding layer to extract the highest probability tokens associated with a given concept vector $u$. Likewise, to visually interpret a concept vector $u$, the authors find the set of images that maximally activate $u$. The authors provide quantitative evidence demonstrating that their method generates concepts which are well aligned with both the input image and the ground-truth captions. Additionally, they demonstrate their approach qualitatively produces well defined concepts with limited overlap between the tokens represented by different concepts.

Strengths

1. This paper introduces a promising novel approach to performing mechanistic interpretability on multimodal models. Given the increasing ubiquity of multimodal models there exists a clear need for interpretability approaches for this style of model. This work presents a viable dictionary-learning based approach I look forward to seeing other researchers building upon in the future. 2. The evaluation of the framework is comprehensive including both qualitative and quantitative results, both of which are essential for any interpretability method. Additionally, I was impressed to see the authors consider multiple approaches (e.g. PCA/KMeans) and sensible baseline models when evaluating their results, providing more confidence that the final approach taken was an appropriate method. 3. The paper is very clearly articulated, with the rationale well-defined, the approach clearly outlined and the results succinctly and clearly summarised.

Weaknesses

1. The authors provide good evidence that the concepts extracted by their method show minimal overlap with other concepts, however they do not address the alternative possibility, that their extracted concepts might represent more than one distinct concept. This phenomenon of feature “superposition” has been well documented in other model interpretability work (see [1][2][3]) and so it seems plausible that it may arise in the approach taken here. This seems especially likely given that the dimensionality of their concept dictionary is lower than the dimensionality of the internal representations. Though I do not think this should detract from the otherwise excellent contributions presented in this paper, I do think this at least warrants a brief discussion, and perhaps more qualitative analysis of the extracted concepts, to assess whether any evidence of feature superposition is observed. 2. Occasionally the authors make claims that go beyond the scope of this work. For example, in the abstract they state, “we present a novel framework for the interpretation of LMMs”. However, this claim seems too strong given that this approach is only really valid for image captioning models rather than large multimodal models more generally. Additionally, there are a few comments such as “we find the generalization of LLMs to multimodal inputs is an interesting phenomenon to understand” in 3.1 (under “Training”) and “the multimodal structure of internal token representations starts to appear at [later layers]” in 4.2 (under “Layer Ablation”). However these claims do not seem valid as the language model layers are frozen during training. It seems more appropriate to say that these extracted concepts represent language model concepts, and the connector model learns to transform the image representations such that they align with these language model concepts. [1] Elhage, N, et al. "Toy Models of Superposition." arXiv:2209.10652 (2022) [2] Arora, S et al. "Linear Algebraic Structure of Word Senses, with Applications to Polysemy." arXiv:1601.03764 (2016) [3] Elhage, N, et al. "Softmax Linear Units" Transformer Circuits Thread (2022)

Questions

1. Why did the authors choose to use Opt-6.7B as the language model rather than more powerful similar sized models such as Llama-7B? 2. Why do the authors take the absolute activations of $V$ in (5)? Are these activations not guaranteed to be non-negative by the optimisation objective? 3. Did the authors consider trialling gradient-based feature visualisation approaches [1] in addition to taking images with the highest activations? This could be an alternative approach to build additional confidence that the extracted concepts do represent what qualitative analysis of the highest activating samples appears to suggest. I don’t expect the authors to perform analysis of this kind in this manuscript however it could be worth commenting on this as a future avenue of research, or alternatively raising any valid criticisms of this approach? 4. There is a typo in figure 1 (“Captionin”) [1] Olah, C, et al. "Feature Visualization", Distill, 2017.

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

The work presented here is only for a single, relatively small model, trained with a specific objective (image captioning). As such it is not clear that this approach will necessarily generalise to other multimodal models. The authors should touch on this limitation in the discussion or limitations section in the appendix.

Reviewer afLg6/10 · confidence 3/52024-07-12

Summary

This paper proposes a new approach to understand multimodal concepts learned in LLMs with visual prefixes. To do so, the authors propose a dictionary learning-based approach that decomposes the representation of a word token in the product of two low-rank matrices via Semi-NMF, one representing the concepts and the other representing the activations for a given token. The authors evaluate the representation of the DePALM model on 8 common objects (e.g., dog, bus, cat) showing qualitatively good results, as well as better auto-eval metrics than with related baselines constructed by the authors.

Strengths

Originality: The paper proposes a novel method to interpret concepts in a multimodal LLM by grounding them in both text and visual spaces. For a given concept, it creates a representation matrix from image–caption pairs, which is then decomposed into two low-rank matrices using Semi-NMF. Quality: The authors show that extracted concepts for a given token can be interpreted both visually and textually through qualitative examples. Quantitatively, the method outperforms other related baselines introduced by the authors. Clarity: The paper is relatively well written, although some parts (eg, Section 3.4, 4.0 and 4.1) feel very dense and require more time to be processed. Significance: The paper provides interesting insights in the representation of a given word token processed by a frozen LLM that is augmented with image understanding via visual prefix learning.

Weaknesses

1. My main concern is in studying the representation of only eight tokens. It would be interesting to study more words and see if there are any interesting takeaways. These could include the 80 COCO classes, and additional rarer tokens. 2. The use of Semi-NMF in L165-167 is unclear. Later on, we find out that it works better than other decompositions, but I was left wondering how the authors came up with Semi-NMF in Section 3.4. 3. It would have been useful to compare the proposed approach with some current interpretability methods that could apply here (or at least discuss why they might not be used here).

Questions

1. Are there any blockers to perform a larger-scale (wrt tokens) analysis? 2. Interpretability methods like ROME use the last token of a given word – why do you use the first token? 3. Does the method actually work for multi-token words (i.e., words that are split into multiple tokens by the tokenizer)? If so, are any of your 8 words in that category? 4. Please double-check your references, some of them appear more than once.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer XzMc4/10 · confidence 2/52024-07-12

Summary

In this paper, the authors propose a framework for interpreting LMMs. Specifically, they introduce a dictionary learning-based approach applied to the representation of tokens. The elements of the learned dictionary correspond to the proposed concepts. These concepts are semantically well-grounded in both vision and text.

Strengths

Using concepts to interpret large multimodal models is a promising idea.

Weaknesses

1. I'm sorry, I don't fully understand this field, so I will lower my confidence score. 2. The notation system is quite confusing. I suggest the authors reorganize it for better clarity. 3. The paper directly applies the concept activation vector (CAV) to Large Multimodal Models, but it does not explain the benefits of doing so or why this approach is valid.

Questions

The experiments in the paper are relatively few. I suggest conducting further research on the interpretability of multimodal large models such as LLaVA and MiniGPT-4.

Rating

4

Confidence

2

Soundness

2

Presentation

2

Contribution

2

Limitations

yes, the author explains the limitations of their study and potential negative societal impact.

Reviewer Go374/10 · confidence 4/52024-07-13

Summary

The authors propose using dictionary learning to extract concepts from multimodal models and simultaneously ground them in the text and image latent space. They draw on prior work on multimodal neurons and concept activation vectors. The authors provide quantitative results using CLIPScore and BERTScore to measure concept extraction, multimodal grounding, and concept overlap.

Strengths

Appears to be mathematically correct and well-grounded in prior work. Well-written and well-illustrated.

Weaknesses

Evaluates only a few tokens (dog, bus, train, cat) thoroughly, which seems like not enough when prior work (Goh et al.) studies thousands of concepts. Evaluates only one model, DePALM. Uses two automated metrics, CLIPScore and BERTScore, but no human evaluation. A little worrisome that DePALM uses a CLIP-ViT-L14 and CLIPScore is the primary evaluation for the method. Overlaps significantly with prior research (Goh et al, Kim et al); multimodal neurons have been described in work dating back several years using non-negative matrix factorization.

Questions

In table 2, why is PCA by far the best method for minimizing overlap between learned concepts? Presumably because it’s PCA on only these five concepts, meaning that variance is maximized between only the five concepts and not between the dictionary more broadly? Do you observe any evidence of polysemanticity, as reported in prior work including Goh et al.?

Rating

4

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

I don’t see many negative societal implications of this research.

Reviewer 8A1z8/10 · confidence 4/52024-07-15

Summary

Authors propose to look at the representation of chosen concepts, in multimodal models. They test different automated methods to learn decompositions of a token representations (which they then linearise in a dictionary of concepts). They also provide quantitative and qualitative analysis of a few examples, showing that this method clusters different human-interpretable meaning or usages of a same concept, for both visual and textual modalities.

Strengths

Great qualitative analysis Multiple representation decomposition algorithms are tested, including more well known PCA or very specific but better suited Semi NMF, which is an adaptation to practical realities. Evaluation of the method is quite extensive, using a wide range of relevant state of the art interpretability tools to verify their work theoretically and practically. Effort is made to put quantitative metrics on concepts and to evaluate very abstract semantic analysis. Provided examples are clear and strike curiosity.

Weaknesses

I worry there is a circularity when evaluating with ClipScore a frozen model with a CLIP component. (and even with BERT scoring, where you evaluate an interpretability metric with a non interpretable system of somewhat similar complexity). (Minor) Very recent works are relevant to the interpretability of LLM / Multimodal models (Templeton, et al., "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet", Transformer Circuits Thread, 2024 AND Gao, Leo, et al. "Scaling and evaluating sparse autoencoders." arXiv preprint arXiv:2406.04093, 2024) and could be cited. Line 223 "the correspondance between a the image" is probably a typo and should be fixed

Questions

Mostly a personal curiosity question: Are there any intuitions from this work allowing to disentangle the data effect of the apparition of the studied concepts from the model/training effect?

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

Sample experiments are done on very simple words and concepts, available at pre-training for the tested model. How well this method would adapt to more complex concepts and images is not studied, but is where a lot of interpretability becomes necessary. This is nonetheless a necessary first step in an exciting direction. Do you have an intuition on how well this method would scale to this?

Reviewer 17ab2024-08-09

Thank-you for addressing all my comments, the additional analyses are very interesting. I’m happy with all the responses to my queries and look forward to the final version of the manuscript. I have one additional comment I’ve made below: **Qualitative analysis for feature superposition/specificity of concept vectors** The polysemanticity analysis is very interesting. Thank-you for taking the time to conduct this analysis, I think it would be a useful discussion point in the appendix. The analysis you have presented considers the specificity of concepts related to a given token (“dog” in this instance). I would be interested in understanding the extent to which the concepts for one token are polysemantic for other tokens e.g. is the “Hot dog” concept active for tokens other than “dog”. I don’t expect you to conduct any additional analyses or to edit the manuscript but just wanted to raise this point as a potential future avenue of research.

Authorsrebuttal2024-08-11

Thanks for the rebuttal acknowledgement! Happy to address all your questions! For future development, we'll certainly keep your point about polysemanticity of concept vectors to other tokens under consideration .

Reviewer afLg2024-08-09

Thank you for answering my questions and clarifying my doubts. I will keep my positive evaluation of the paper, and follow on the discussion with the other reviewers.

Authorsrebuttal2024-08-11

Thanks for the rebuttal acknowledgement! We're glad to address your doubts.

Authorsrebuttal2024-08-11

Thank you for the rebuttal acknowledgement!

Authorsrebuttal2024-08-13

Dear reviewer, We have incorporated your complete feedback in our rebuttal and sincerely hope it addresses all the concerns.

Authorsrebuttal2024-08-13

Dear reviewer, We have incorporated your complete feedback in our rebuttal and sincerely hope it addresses all the concerns.

Reviewer XzMc2024-08-13

I thank the authors for their endeavor in preparing the rebuttal. I have a new question: why are the results of MiniGPT4 missing in global response?

Authorsrebuttal2024-08-13

Thank you for the rebuttal acknowledgement. Due to the limited rebuttal time frame it was not possible for us to experiment with MiniGPT-4. We prioritized experiments on LLaVA because of its popularity and the competitiveness of its newer versions with state-of-the-art models like Gemini and GPT-4. Another point to note is that MiniGPT-4 has a similar architecture to LLaVA and thus there is no particular reason to expect its observations to be fundamentally different from LLaVA. In regard to experiments with other LMMs, we would also like to note our experiments in global rebuttal on other DePALM models with non CLIP visual encoders, where we obtain consistent results. We will certainly consider evaluating more LMMs, including MiniGPT4, for future works. We hope this clarifies your doubt.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC