Seeing the Image: Prioritizing Visual Correlation by Contrastive Alignment

Existing image-text modality alignment in Vision Language Models (VLMs) treats each text token equally in an autoregressive manner. Despite being simple and effective, this method results in sub-optimal cross-modal alignment by over-emphasizing the text tokens that are less correlated with or even contradictory with the input images. In this paper, we advocate for assigning distinct contributions for each text token based on its visual correlation. Specifically, we present by contrasting image inputs, the difference in prediction logits on each text token provides strong guidance of visual correlation. We therefore introduce Contrastive ALignment (CAL), a simple yet effective re-weighting strategy that prioritizes training visually correlated tokens. Our experimental results demonstrate that CAL consistently improves different types of VLMs across different resolutions and model sizes on various benchmark datasets. Importantly, our method incurs minimal additional computational overhead, rendering it highly efficient compared to alternative data scaling strategies. Codes are available at https://github.com/foundation-multimodal-models/CAL.

Paper

Similar papers

Peer review

Reviewer mF2S5/10 · confidence 4/52024-07-08

Summary

The paper proposes a new re-weighting strategy for training VLMs to improve visual-language alignment. The motivation is to assign higher weight to visually relevant tokens and lower weight to visually irrelevant and visually contradictory tokens. The re-weighting factor relies on the logit difference with and without image inputs for each token. Using the proposed CAL, the paper demonstrates general improvement in VQA and captioning tasks.

Strengths

The motivation to dynamically assign weights to different types of tokens is intuitive and well-studied in the language domain. Several studies have attempted to apply this method to vision-language tasks in a zero-shot manner, which incurs high inference costs. This paper proposes training the model with this method instead. The approach is simple and effective, demonstrating general improvements over existing state-of-the-art models across benchmarks. The analysis further validates the effectiveness of the proposed method.

Weaknesses

1. **Lack of Clarification in Presentation**: Some content is confusing. For example, equation 2 is averaged across the entire sentence with length $l$, yet the description in lines 88-89 refers to the loss objective for the $i^{th}$ sample at token $t_j$, which seems more like the loss at a single token. Additionally, equation 4 lacks clarification; it's unclear how to apply average pooling with a window size from this equation. Some experiments also lack necessary context. For instance, it's unclear which model is evaluated in Figure 2a, and the ablation study in section 3.3 doesn't specify which model is used for evaluation. The main result in section 3.3 doesn't disucuss how to apply CAL into model training, at which stage for example. The annotation is inconsistent; in equation 4, it's $\tilde{w}^{i,t_j}$, while in equation 5, it's $\tilde{w}(i.t_k)$ 2. **More Experiments Are Required**: The paper presents decent improvements over some baselines, mainly in vision-language tasks. However, since the training recipe has changed, it would be better to also evaluate other tasks, such as language-only tasks, to see how the CAL affects the model's ability in standard language understanding. Additionally, the method seems quite relevant to hallucination tasks, so it would be beneficial to report results on those as well. 3. **Lack of Originality**: The motivation and methodology are already well-studied in language models [1,2], and prior studies have also applied changes to logits based on the presence or absence of images to measure the relevance of visual and language inputs. In this context, the paper simply applies the method during the training stage, instead of the inference stage as in [3]. Therefore, the novelty is somewhat limited. [1] Jiang S, Wolf T, Monz C, et al. TLDR: token loss dynamic reweighting for reducing repetitive utterance generation[J]. arXiv preprint arXiv:2003.11963, 2020. [2] Lin Z, Gou Z, Gong Y, et al. Rho-1: Not all tokens are what you need[J]. arXiv preprint arXiv:2404.07965, 2024. [3] Zhu L, Ji D, Chen T, et al. Ibd: Alleviating hallucinations in large vision-language models via image-biased decoding[J]. arXiv preprint arXiv:2402.18476, 2024.

Questions

1. The trend shown in Figure 2a is based on which model? Is this a common phenomenon across existing VLMs? 2. How do you deal with cases where a semantic word has been split into two subparts, each with a different value, such as "tr" and "uck" in Figure 2a? 3. The CAL seems to rely heavily on the model's logits with and without image input. However, can the model accurately predict changes in logits, as shown in Section 2, during the initial alignment stages? 4. The training requires an additional forward pass of the LLM. Why is the training described as "lightweight" in the paper? Doesn't this double the cost?

Rating

5

Confidence

4

Soundness

3

Presentation

1

Contribution

2

Limitations

not applied

Reviewer ewhK6/10 · confidence 4/52024-07-09

Summary

This paper introduces Contrastive Alignment (CAL), a straightforward yet effective re-weighting strategy designed to improve multimodal alignment. Specifically, the authors propose contrasting image inputs and calculating the differences in prediction logits for each text token to determine its training weights. Extensive experiments are conducted to demonstrate the effectiveness of CAL across various tasks, providing a robust validation of the proposed approach.

Strengths

1. The Contrastive Alignment (CAL) proposed in this paper is an efficient method for multimodal alignment, particularly beneficial when training data is heavily contaminated with noise. The dynamic adjustment of training weights represents a lightweight yet innovative solution to this problem. 2. The experimental section of this paper is thorough. It includes extensive quantitative analysis and demonstrates the performance of CAL across a broad range of scenarios. These results not only allow for a deep understanding of the CAL method but also provide valuable insights into its practical applications.

Weaknesses

1. While CAL is innovative for aligning multimodal data, its robustness and comprehensiveness in complex task settings, such as multimodal question answering (VQA), appear limited. The method ensures correct alignment but may inadvertently impair other model capabilities, such as reasoning ability or retention of original world knowledge. This issue is critical in VQA scenarios where data often include elements not directly related to image content, potentially conflicting with it and incorporating extensive external knowledge. The application of CAL in such contexts might suppress training on these critical aspects, introducing uncontrollable biases and worsening model performance. Although CAL shows promise in simpler tasks like captioning and grounding, its performance in VQA tasks is inconsistent, and its effectiveness in addressing hallucination issues, such as in the POPE dataset, is negligible. 2. CAL’s effectiveness appears contingent upon the pre-existing performance of the underlying multimodal model. Since it calculates visually correlated weights based on the model itself, any hallucinations or errors present in the base model are not only perpetuated but potentially exacerbated by the CAL method. This limitation restricts CAL’s applicability and caps its performance, as there is no mechanism within CAL to correct these amplified errors.

Questions

1. Why MGM-HD-13B perform worse that -7B model in caption tasks as shown in Tab-2 in the paper ? 2. The visually correlation weights in CAL are smoothed using average pooling with a window size of W. How is W determined? Is this parameter set consistently across all datasets, or does it require adjustment for each training run to optimize performance? 3. The paper mentions that CAL adds minimal computational overhead compared to other data scaling strategies. However, could the authors provide more quantitative examples or an intuitive comparison? It seems that processing very long texts could significantly increase the overhead.

Rating

6

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

N/A

Reviewer tsLt7/10 · confidence 4/52024-07-11

Summary

This paper points out that parts of samples in the broadly used datasets contain visually contradictory text tokens. To mitigate the sub-optimal cross-modal alignment in VLMs, this proposed method is to assign distinct contributions for each text token based on its visual correlation.

Strengths

1. The proposed method is easy to implement and effective. The idea of assigning distinct weights for each token is intuitive. 2. This paper is well organized and provides clear preliminaries, helping readers understand the method. 3. CAL achieves consistent and solid performance across different benchmarks.

Weaknesses

Please see the Questions

Questions

1. It’s unclear how to calculate the prediction logit distribution o without input I for the VLM like CLIP. Is this method not applicable to the pure image-text matching VLM like CLIP? 2. There’s no l in Equation 4. Does it mean W in Sec. 3.1 implementation details “we set l in Equation 4 to 3 for all experiments”? 3. For Figure 3, it is better to describe what the dashed line represents as well.

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

N/A

Area Chair om212024-08-13

Hi Reviewer tsLt, Could you take a look at the authors' rebuttal and finalize your rating? Thanks, AC

Reviewer DwVB6/10 · confidence 3/52024-07-13

Summary

This study proposes a reweighting strategy, namely contrastive alignment, to enhance model learning on visually correlated tokens. Specifically, the authors divided the tokens into three sub-groups, virtually related, non-related, and contradictory, and assigned different weights to each group. The weight is calculated based on the predicted contrastive logit (w/o visuals) and further post-processed by clamping and average pooling. This study is evaluated on vision question answering, image caption, and grounding tasks using two foundation model structures, LLaVA and Mini-Gemini.

Strengths

-- The paper is well-written. -- Deploying a reweighting strategy on different types of tokens seems feasible for performance improvement. -- The division of token groups is straightforward yet reasonable.

Weaknesses

-- This study does not compare with any baseline models within the topic of resisting noisy tokens, such as [*] and [**]. -- In [*], a study with similar aims was presented. In particular, they also studied the image-token consistency. Thus, detailed baseline comparisons, including overall performance and token consistency evaluation/comparison, are expected. -- The authors claim that the proposed strategy is simple yet efficient. However, it may rely upon several strong assumptions, such as high capability requirements of the pre-trained model. These assumptions are not presented, discussed, and validated in the paper. -- Weight distribution on different token sub-groups is not presented. This would contribute to the validation of weighting. [*] Gou, Yunhao, et al. "Leveraging per image-token consistency for vision-language pre-training." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. [**] Wu, Cheng-En, et al. "Why Is Prompt Tuning for Vision-Language Models Robust to Noisy Labels?." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.

Questions

See above weakness.

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

As I can see, this paper has no potential negative societal impact.

Area Chair om212024-08-13

Hi Reviewer DwVB, Could you take a look at the authors' rebuttal and finalize your rating? Thanks, AC

Reviewer ewhK2024-08-10

Thanks for the response, most of my concerns have been addressed. However, for an algorithm that introduce additional computational costs, mere performance enhancements in "most" scenarios are insufficient, particularly as we observe that the CAL method results in performance degradation in the LLaVA-Next model. This phenomenon could be anticipated given the lack of interpretability of CAL in complex settings, or as discussed in W2, the introduction of uncontrollable biases stemming from inherent limitations of the base model. Despite its flaws, this work challenges the dominant paradigm of pure supervised learning in multimodal training, making it a valuable contribution to the field. In recognition of its potential, I will raise my score. I urge the authors to provide more quantitative analyses in future version of the paper, such as exploring the weight distribution highlighted by Reviewer DwVB, or presenting scenarios where CAL fails to enhance performance, to deepen our understanding of these issues.

Authorsrebuttal2024-08-14

Thank you for your detailed feedback and for raising your score. We appreciate your recognition of our work's potential and your valuable suggestions. Regarding the computational cost, we can preprocess and store the CAL weights using the trained model offline, which will not increase the cost in subsequent training. This is one of our future improvement directions. In future versions of our paper, we will add more analyses, including exploring weight distribution and scenarios where CAL may not enhance performance. In our subsequent work, we will strive to address the uncertainties introduced by the base model.

Reviewer mF2S2024-08-12

I appreciate the authors' response in addressing my questions and concerns. The explanation effectively clarifies the concern regarding training costs, particularly noting that the second forward pass does not require storing intermediate activations or performing backward propagation. However, the claim of a 20% increase in time is not scientifically grounded, as it can vary depending on the dataset used for training. My third question was not fully understood by the authors. The inquiry was not about the training recipe for MLLMs like LLaVA. In the early stages of training, one might assume the model lacks the ability to generate captions from images but the LLMs can perfectly generate text from a partial caption. The question is whether the gap in predicted logits with and without the image is accurate in the initial stage where the model lacks good captioning capacity. It might be more appropriate to apply this training during the instruction-tuning stage, where it is assumed the model can generate captions after pretraining. Another way to frame my question is: considering that language models typically use RoPE as a positional embedding, how does prepending image tokens affect logits given the altered positional embedding of text tokens in this scenario? Overall, while some concerns remain, particularly regarding the limited originality of the methodology in light of prior works, I find the work interesting and believe it may inspire future research. As such, I have raised my score to 5.

Authorsrebuttal2024-08-14

Thank you for your detailed feedback and for raising your score. We appreciate your recognition of our work's potential and your valuable suggestions. Regarding the computational cost, we can preprocess and store the CAL weights using the trained model offline, which will not increase the cost in subsequent training. This is one of our future improvement directions. In the pretraining stage, since the number of trainable parameters (only projector) is limited and there is a minimum weight for all text tokens, inaccuracies in predicting logits in the early stages do not significantly impact the overall performance. As shown in Table 3, applying CAL in both pretrainig and finetuning stages yields the best performance. Regarding the issue of positional embeddings, in the current version, we ensure that the positional embeddings for text tokens are consistent in both two scenarios. Once again, thank you for your valuable insights and support. We are committed to addressing these issues in our future work.

Reviewer tsLt2024-08-13

Response to rebuttal by authors

Thanks for the response. The Answer for Q1 is still a vague idea in its infancy. In future work, the authors may implement the CAL on CLIP by using "noise-added images" to decide the token scores and see if it works. I choose to keep my original positive rating.

Reviewer DwVB2024-08-13

The authors partially addressed my concerns. So, I increased my rate.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC