IPO: Interpretable Prompt Optimization for Vision-Language Models

Pre-trained vision-language models like CLIP have remarkably adapted to various downstream tasks. Nonetheless, their performance heavily depends on the specificity of the input text prompts, which requires skillful prompt template engineering. Instead, current approaches to prompt optimization learn the prompts through gradient descent, where the prompts are treated as adjustable parameters. However, these methods tend to lead to overfitting of the base classes seen during training and produce prompts that are no longer understandable by humans. This paper introduces a simple but interpretable prompt optimizer (IPO), that utilizes large language models (LLMs) to generate textual prompts dynamically. We introduce a Prompt Optimization Prompt that not only guides LLMs in creating effective prompts but also stores past prompts with their performance metrics, providing rich in-context information. Additionally, we incorporate a large multimodal model (LMM) to condition on visual content by generating image descriptions, which enhance the interaction between textual and visual modalities. This allows for thae creation of dataset-specific prompts that improve generalization performance, while maintaining human comprehension. Extensive testing across 11 datasets reveals that IPO not only improves the accuracy of existing gradient-descent-based prompt learning methods but also considerably enhances the interpretability of the generated prompts. By leveraging the strengths of LLMs, our approach ensures that the prompts remain human-understandable, thereby facilitating better transparency and oversight for vision-language models.

Paper

References (66)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer XUaZ6/10 · confidence 4/52024-07-01

Summary

This paper introduces a method called Interpretable Prompt Optimization (IPO) for vision-language models. The goal of IPO is to improve the performance and interpretability of vision-language models by dynamically generating and optimizing text prompts. The paper addresses the limitations of current prompt optimization methods, which often lead to overfitting and produce prompts that are not understandable by humans. IPO leverages large language models (LLMs) to generate effective and human-readable prompts (which is the main advantage of their method). It incorporates a Prompt Optimization Prompt that guides the LLMs in creating prompts and stores past prompts with their performance metrics, providing rich in-context information. Additionally, IPO integrates a large multimodal model (LMM) to generate image descriptions, enhancing the interaction between textual and visual modalities. This allows for the creation of dataset-specific prompts that improve generalization performance while maintaining human comprehension. The paper validates IPO across 11 datasets and demonstrates that it improves the accuracy of existing prompt learning methods and enhances the interpretability of the generated prompts. Overall, IPO ensures that the prompts remain human-understandable, facilitating better transparency and oversight for vision-language models.

Strengths

- The paper tackles one of the most important problems in prompt tuning. Existing prompt tuning methods are not interpretable at all (as they optimize a sequence of vectors with a model specific loss function). This paper mitigates this issue by using a LLM as an optimizer -- which inherently generates text. - The prompt optimization template though is simple, can be scalable towards multiple tasks (as shown in the paper). - Thorough ablations on the language model and the multimodal language model has been provided in the paper.

Weaknesses

- [Major]. While the paper obtains interpretable prompts, the method primarily obtains improved performances for the novel classes, but not for the base classes. In fact for a large number of the datasets, the base classes performance is significantly low when compared to other methods. I believe that there should be a balance in the performance between the novel and base classes. To improve the paper, the authors are suggested to provide strong justifications explaining this. - [Minor]. This area has a multitude of papers so it might be tedious to compare with all the methods. However, I would suggest the authors to add a separate section having a discussion on the comparison of their method with more recent prompt-tuning methods such as LFA, PLOT, DFA etc. - [Minor]. I would also suggest the authors to extend the discussion of the limitations of the method. Currently, as I see, the method is suitable for few-shot scenarios, but what if one has access to a large number of domain samples? How can this method be made scalable?

Questions

Overall, the paper provides a fresh view on prompt-tuning by making the process interpretable. In a way, the paper automates prompt-engineering for a domain by using a LLM as an optimizer. I am going with Borderline Accept, but happy to revisit my scores if the Weaknesses are addressed in the rebuttal.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

See Weaknesses.

Reviewer xRmh6/10 · confidence 5/52024-07-08

Summary

This paper presents an Interpretable Prompt Optimizer (IPO) designed to improve the performance and interpretability of pre-trained visual language models such as CLIP. By dynamically generating textual cues using a large language model (LLM) and combining it with a large multimodal model (LMM) to generate image descriptions, IPO demonstrates better accuracy and interpretability than traditional gradient descent-based cue learning methods on multiple datasets.

Strengths

1)The paper proposes a new prompt optimization framework that combines the advantages of LLM and LMM. 2)Extensive testing on 11 datasets demonstrated the effectiveness of the method. 3)The paper is clearly structured with diagrams and charts to aid illustration. 4)Improved interpretability of visual language models is important for achieving better human-computer collaboration.

Weaknesses

1)Line 150 :(2) -> (3) 2)There are no citations throughout the paper Table 6. 3)Lack of cross-dataset experimental evaluation to validate the generalizability of IPO. 4)The paper mentions the use of large language models such as GPT-3.5 Turbo, but does not discuss in detail the computational cost and efficiency of these models during training and inference. For resource-constrained environments, this may be an important consideration. 5)The importance of each component of the IPO is described in Section 5.2, but it lacks in-depth analysis, e.g., it could be mined for some more intrinsic reasons for the rise in points from a regularization perspective.

Questions

1)Will this parameterless training time be shorter for IPO compared to the gradient descent approach? 2)How does the length of the prompt history affect the stability of training convergence?

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer 4MWY5/10 · confidence 5/52024-07-09

Summary

This paper proposes an interpretable prompt optimizer (IPO) which uses an LLM to iteratively optimize prompt templates that lead to improved zero-shot visual classification performance on CLIP.

Strengths

1) The proposed method outperforms baselines on the novel classes in the evaluation on base-to-novel generalization benchmark

Weaknesses

1) Line 91, “However,to the best of our knowledge, no existing studies have investigated how LLMs could be used to optimize text prompts within vision-language models”. There are actually related work that employed LLMs for optimizing prompts for visual classification in VLMs which are not covered in this paper, e.g. [a] Liu et al. Language Models as Black-Box Optimizers for Vision-Language Models. CVPR’24 [b] Mirza et al. Meta-Prompting for Automating Zero-shot Visual Recognition with LLMs. ECCV’24 Analysis and proper comparison to these works should be conducted. Here [a] also uses an LLM (ChatGPT) to iteratively update the prompt templates for visual classification on CLIP with the good prompts and bad prompts passed as in-context examples. This weakens the novelty of the proposed method. 2) In Table 5, why is the performance of the proposed method consistently worse than baselines on base classes? How does the method perform in a usual one-shot classification setting (instead of base-to-novel setting)? 3) In table 1, the metric “H” should be specified

Questions

1) In the base-to-novel evaluation, 1-shot refers to the setting where one sample of each category in the base classes is provided during training while no sample of the novel classes is used in training? 2) Line 165, “all four models failed to understand our Prompt Optimization Prompt”, here it should be four models or six models (as the experiments are conducted on six models)?

Rating

5

Confidence

5

Soundness

3

Presentation

2

Contribution

2

Limitations

The proposed method has improved performance on novel classes, but suboptimal performance on base classes where the prompts are optimized.

Reviewer VMZQ6/10 · confidence 5/52024-07-13

Summary

The paper addresses the challenge of optimizing text prompts for vision-language models, specifically focusing on the interpretability of these prompts. Traditional methods for prompt optimization rely on gradient descent, which often results in overfitting and produces prompts that are not human-readable. This paper introduces a new approach called Interpretable Prompt Optimizer (IPO), which leverages large language models (LLMs) to generate and optimize prompts in a dynamic and interpretable manner. The paper details the design of the IPO framework and provides extensive experimental results across 11 datasets. The findings demonstrate that IPO not only improves the accuracy of vision-language models compared to traditional gradient-based methods but also significantly enhances the interpretability of the generated prompts.

Strengths

1. The paper introduces IPO, a novel method that uses LLMs to optimize prompts in a way that maintains human readability and interpretability. This approach contrasts with traditional gradient-based methods that often produce opaque prompts. 2. The POP system stores past prompts and their performance metrics, allowing LLMs to generate more effective prompts through iterative refinement. This system enhances the contextual understanding and effectiveness of the generated prompts. 3. The IPO method incorporates LMMs to generate image descriptions, which improves the synergy between textual and visual data. This leads to more accurate and contextually relevant prompts. 4. The paper validates the effectiveness of IPO across 11 different datasets, demonstrating its superiority over traditional methods in terms of both accuracy and interpretability.

Weaknesses

The main weaknesses of this paper lies in the experiment can not support the effectiveness of the LLM optimizer, which is the most important contribution of the authods. IPO incorporates the Large Language Model (LLM) as an optimizer to learn an interpretable prompt pool, and everytime a new instance comes, prompts will be extracted from this prompt. This mainstream is similari to knowledge bank based prompt learning methods such as L2P [1], AttriCLIP [2], DualPrompt [3]. Therefore, the experiments should contains: 1) **A fair comparitive with some of the knowledge bank based prompt learning methods such as L2P [1], AttriCLIP [2], DualPrompt [3]. To exclude the effectivness of the memory retrieval mechanism of IPO.** 2) **The current comparitive experiment merely involves the 1-shot setting in Table 5. Please provides a more comprehensive experiment settings, such as 16-shots and full-sized tuning setting. Besides, the IPO methods do not draw an obvious performance improvement in benmark datasets in the case that IPO has an obviously longer prompt that the learnable method.** 3) What's more, the IPO does not perform well in the large-scale generic datasets such as ImageNet. This is not intuitive, because the LLM should show a better performance in tghe generic datasets. Lastly, I am not agree with your average metric, because the data size of these 11 benchmarks have significant difference. 4) Table 6 are not referred in the author's paper, and the detailed benchmark is also not elaborated. [1] Wang Z, Zhang Z, Lee C Y, et al. Learning to prompt for continual learning[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 139-149. [2] Wang R, Duan X, Kang G, et al. Attriclip: A non-incremental learner for incremental knowledge learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 3654-3663. [3] Wang Z, Zhang Z, Ebrahimi S, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 631-648.

Questions

Authors explor using LLM as an interpretable optimizer for prompt learning, I think this is a contribution worth a weekly accept score, i.e., 6 scores in NeurIPS rating. However, this score should be build upon all below concers are addressed: 1) I can tolerate that the current performance are not that remarkable, but this tolerance should be built upon the experiments are fair. To make a fair comparison in the framework, authors should make comparison with some of the knowledge bank based prompt learning methods such as L2P [1], AttriCLIP [2], DualPrompt [3]. So that we can exculde the effectivness of the memory retrieval mechanism of IPO. 2) Why Table 5 only contains the 1-shot comparive experiments ? Detailed 16-shots or training with more samples setting are needed. I will confirm the effectivness of the your IPO if you draw a remarkable performance. If not, just take a breath and provide enough evidence that LLM optimizer is effectiveness, such as sound losses, sound improvement in the training set, sound improvement in the test sets in different settings. I will agrre with the effectivness of the your IPO if the evidence are sound. [1] Wang Z, Zhang Z, Lee C Y, et al. Learning to prompt for continual learning[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 139-149. [2] Wang R, Duan X, Kang G, et al. Attriclip: A non-incremental learner for incremental knowledge learning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023: 3654-3663. [3] Wang Z, Zhang Z, Ebrahimi S, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 631-648.

Rating

6

Confidence

5

Soundness

1

Presentation

3

Contribution

3

Limitations

Experiments now are not sound and the authors do not provide enough evidnce to prove the effectiveness of IPO.

Reviewer VMZQ2024-08-07

I am happy with your response. However, some concerns are still not solved. 1. Please add these clarifications to the revised manuscript to make the manuscript clearer: In contrast, our IPO primarily focuses on few-shot VLMs, where during training, LLMs are utilized as our prompt bank to generate dataset-specific prompts based on contextual information. During testing, predictions are made directly using the learned prompts without the need for additional retrieval. 2. **Please extend Table 6 to be as detailed as Table 5 because this setting is very important. (If you have space limits, please provide the result of large-scale datasets such as ImageNet)** 3. I can not understand why ImageNet results in a longer LLM input when generating image descriptions for its large samples and categories. Is this because of a larger batch size? In my view, it is not necessary to input all the dataset categories when you processing one training sample. Therefore, I am not satisfied with this answer. Please make more clarification. You are almost close to your 6 rates. Please hold on.

Authorsrebuttal2024-08-09

Thank you for the response.

Thank you very much for your encouragement. 1. We will clarify this point in the related work section of our revised manuscript, including a discussion on knowledge bank-based prompt learning methods. Additionally, we will incorporate the comparison experiment with VPT + L2P. 2. We present the 16-shot performance of our IPO method across 11 datasets, showing detailed results for Base/Novel/H metrics in the table below. Our IPO outperforms all other methods on the Novel classes and the H metric, demonstrating its ability to mitigate overfitting. However, due to the word limit in the rebuttal, we cannot include comparisons with all methods here. In the revised manuscript, we will include detailed comparisons with CLIP, CoOP, CoCoOp, MaPLe, PromptSRC, CoPrompt, LFA, PLOT, and VPT + L2P. | Model | ImageNet | Caltech101 | OxfordPets | StanfordCars | Flowers102 | Food101 | |------------|----------|------------|------------|--------------|------------|---------| | CoOp | 76.47/67.88/71.92 | 98.00/89.81/93.73 | 93.67/95.29/94.47 | 78.12/60.40/68.13 |97.60/59.67/74.06 | 88.33/82.26/85.19 | | CoCoOp | 75.98/70.43/73.10 | 97.96/93.81/95.84 | 95.20/97.69/96.43 | 70.49/73.59/72.01 |94.87/71.75/81.71 |90.70/91.29/90.99 | | MaPLe | 76.66/70.54/73.47 | 97.74/94.36/96.02 | **95.43**/97.76/96.58 | 72.94/74.00/73.47 |95.92/72.46/82.56 | 90.71/92.05/91.38 | | PromptSRC | 77.60/70.73/74.01 | **98.10**/94.03/96.02 | 95.33/97.30/96.30 | **78.27**/74.97/**76.58** |**98.07**/76.50/85.95 | 90.67/91.53/91.10 | | IPO | **77.83**/**72.45**/**75.04** | 97.32/**95.23**/**96.26** | 95.21/**98.23**/**96.70** | 73.42/**75.71**/74.55 | 96.78/**78.32**/**86.58** | **90.92**/**93.08**/**91.99** | | Model | Aircraft | SUN397 | DTD | EuroSAT | UCF101 | Average | |------------|----------|------------|------------|--------------|------------|---------| | CoOp | 40.44/22.30/28.75 | 80.60/65.89/72.51 |79.44/41.18/54.24 | 92.19/54.74/68.69 | 84.69/56.05/67.46 |82.69/63.22/71.66 | | CoCoOp | 33.41/23.71/27.74 | 79.74/76.86/78.27 | 77.01/56.00/64.85 | 87.49/60.04/71.21 | 82.33/73.45/77.64 |80.47/71.69/75.83 | | MaPLe | 37.44/35.61/36.50 | 80.82/78.70/79.75 | 80.36/59.18/68.16 | 94.07/73.23/82.35 | 83.00/78.66/80.77 | 82.28/75.14/78.55 | | PromptSRC | **42.73**/37.87/40.15 | **82.67**/78.47/80.52 | **83.37**/62.97/71.75 | 92.90/73.90/82.32 | **87.10**/78.80/82.74 | **84.26**/76.10/79.97 | | IPO | 41.21/**41.42**/**41.31** | 81.25/**80.92**/**81.08** | 82.14/**66.81**/**73.69** | **94.25**/**80.11**/**86.61**| 85.32/**80.92**/**83.06**|79.92/**80.51**/**80.21**| 3. You are correct in your understanding. When we use a larger LLM like GPT-4o, its ability to handle longer text inputs allows us to increase the batch size, leading to higher-quality prompt generation. Even when using the same batch size, GPT-4o's text understanding capability is better than GPT-3.5 turbo, resulting in good performance. For ImageNet, batch training is necessary due to its large number of classes and samples, which require processing multiple instances simultaneously to generate effective prompts. However, this need for batch processing does not apply to other datasets with fewer classes, where single-instance processing suffices. The table below compares the performance of GPT-3.5 turbo and GPT-4o at different batch sizes. We observed that when the batch size increases to 128, the GPT-3.5 turbo's performance starts to decline due to its limited ability to process longer input texts effectively. However, GPT-4o maintains strong performance even at larger batch sizes. That said, using very large batch sizes with GPT-4o becomes cost-prohibitive, so we selected a batch size of 128 for our experiments. We found that even larger batch sizes could further improve performance, but the cost becomes a key factor. We will include this experimental comparison in the revised manuscript. | Model | Batch size | Base | Novel | H | |------------------|------------|--------|--------|--------| | IPO w/ GPT-3.5 | 4 | 73.11 | 68.08 | 70.51 | | IPO w/ GPT-4o | 4 | 74.32 | 67.98 | 70.55 | | IPO w/ GPT-3.5 | 16 | 73.42 | 68.43 | 70.82 | | IPO w/ GPT-4o | 16 | 74.94 | 70.75 | 72.78 | | IPO w/ GPT-3.5 | 32 | 73.79 | 68.72 | 71.16 | | IPO w/ GPT-4o | 32 | 75.01 | 70.93 | 72.91 | | IPO w/ GPT-3.5 | 64 | *74.09* | *69.17* | *71.54* | | IPO w/ GPT-4o | 64 | 75.34 | 71.23 | 73.45 | | IPO w/ GPT-3.5 | 128 | 73.67 | 68.07 | 70.75 | | IPO w/ GPT-4o | 128 | 76.14 | 72.13 | 74.09 | | IPO w/ GPT-3.5 | 256 | 73.11 | 67.81 | 70.36 | | IPO w/ GPT-4o | 256 | **76.81** | **72.73** | **74.71** |

Reviewer VMZQ2024-08-09

OK, most concerns have been addressed. But why does the highest performance of "IPO w/GPT-40, batchsize=256" is (76.81/72.73/74.41) for ImageNet, while in Table 6, the reported performance is (77.83/72.45/75.04). If any different experiment setting is used, please discuss them in detail.

Authorsrebuttal2024-08-09

Thank you for the response.

We are very grateful to hear that most of our issues have been resolved. “IPO w/GPT-4o, batch size=256” with a performance of (76.81/72.73/74.41) for ImageNet refers to the **1-shot** setting, using **GPT-4o** as our LLM. In contrast, the performance reported in Table 6 (77.83/72.45/75.04) is for the **16-shot** setting, using **GPT-3.5 turbo**. We will clarify these detailed settings in the revised manuscript. If you have any further questions, please feel free to let us know.

Reviewer VMZQ2024-08-10

We appreciate your thorough response. The authors have resolved most of my concerns, and the paper can be improved to be 6 rates

Reviewer U8995/10 · confidence 3/52024-07-13

Summary

The paper introduces a method named IPO for VLMs which uses LLMs to dynamically generate and refine text prompts. The method is to improve the accuracy and interpretability of prompts. Experiments show that it can address issues like overfitting and lack of human comprehension in traditional gradient descent-based methods.

Strengths

The prompts generated using IPO enhance the performance of vision-language models across various vision task related datasets. The process of generating prompts are interpretable which brings better transparency and controllability.

Weaknesses

LMM seems contribute little to th final performance. Is the reason that used LMM is not strong enough to generate high-quality image caption? Also for different LLMs, the performance gap is not obvious and the author did not test on GPT4 or stronger models. The overall performance is not much better than optimized methods like CLIP. Could the authors provide more examples to prove the iterpretability advantage of this method?

Questions

Will this method be generalized to other vision related tasks?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

n/a

Reviewer xRmh2024-08-08

We appreciate your thorough response. The authors have resolved most of the issues highlighted in the initial review, and we wish to retain the current score.

Authorsrebuttal2024-08-09

Thank you for the response.

We sincerely appreciate the reviewer's encouragement and suggestions. We will ensure that the revised manuscript includes all the experiments and typos mentioned in our response.

Reviewer XUaZ2024-08-10

Response to Authors

Thanks for the experimental results. The weaknesses have been addressed and as promised, I increase my score.

Reviewer 4MWY2024-08-12

Thanks for the efforts in the response! My concerns are addressed and I increase my rating accordingly.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC