acceptodds
Under review as a conference paper at ICLR 2027

EACache: Entailment-Aware Caching facilitating AI Image Generation

Abstract

Text-to-image generation is computationally expensive, motivating the reuse of previously generated images through caching. Similarity-based cache retrieval is widely used, but image–text similarity may not always accurately indicate whether a cached image contains sufficient content to satisfy a user request. This distinc- tion is especially important for general prompts, where a valid cached image may contain substantial additional content while still satisfying the requested content, which reflects the entailment effect. We introduce EACache, an entailment-aware caching framework that combines similarity-based retrieval with selective entail- ment verification (e-verification). EACache derives two similarity thresholds with the objective of minimizing the expected system cost of serving a query. When the similarity score exceeds the higher threshold, the cached image is returned directly; when the score falls below the lower threshold, a new image is generated; and when the score lies between the two thresholds, e-verification is performed to determine whether the cached image can be reused. The threshold derivation jointly accounts for the costs of e-verification, image generation, transmission, and recovery from an incorrect cache-hit decision. To reduce online e-verification overhead, EACache captions each image when it is added to the cache and performs textual entailment between the stored caption and the incoming query at runtime. Experiments show that similarity remains an informative signal but cannot cleanly separate reusable from non-reusable candidates, while selective e-verification recovers valid cache hits missed by similarity-only reuse without requiring verification for every request. The optimal threshold analysis further provides practical guidance for selecting an appropriate entailment model for caching system design.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Semantic image caching, Image-text similarity, Cost-aware inference, visual entailment

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_aowgOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a novel caching algorithm that uses entailment instead of image similarity for text-to-image generation. While the problem and the idea are interesting, the solution is incomplete.

Critical strengths

S1. Timely and important problem.

Critical weaknesses

Thank you for submitting to ICLR. I think you are working on a very important problem and have a solid amount of work, but the solution is incomplete, unfortunately.

W1. The solution is incomplete.

a. The most important problem with the paper is that it does not describe how to perform entailment… The whole related work and motivation go to a deep understanding of image content beyond computing the trivial similarity score, which makes the problem very interesting. Yet, neither the main body of the paper nor the introduction has a single line on how entailment is done. Figure 2 has a large box for it, but the text does not describe how it is done. Entailment is really at the core of this paper, and it is a highly non-trivial problem to solve. I will talk more about this in my other comments.

b. Section 3.2, Equation 7, “With probability 1 − g(s), the returned image does not satisfy the request.” — this does not make sense. First, Figure 3 does not include any component that decides whether the output satisfies the request. Second, what does “satisfying” mean here? This is just not acceptable. Is there a human in the loop saying “this is OK”? The solution is clearly incomplete. Later in the experimental evaluation, we learn that a GPT is used for that, which really is not acceptable. If something is part of the solution, it has to be part of the solution, not part of the experimental evaluation. Evaluation is for evaluation. Plus, GPT is configured with a specific score of 50, for no clear reason…

Entailment is the place where the decision is made on whether the image is satisfactory or not. Using GPT afterward, unfortunately, shows that the entailment model is not successful. I will talk about this more in my other comments.

Also, what is C_rec here? C_tx and C_g are the recovery costs, aren’t they? What is the extra C_rec?

c. The entailment model ends up being one of the transformers: RoBERTa, DeBERTa, or ViLT. However, it is not clear how they decide whether there is entailment or contradiction… Section 3.3 has no single line on it, which makes the paper incomplete. I will talk about this more in my other comments.

d. OK, the core of the paper is unfortunately one single paragraph in the experimental setup (Models and judgements)… This is not OK, unfortunately. The body of the paper should cover it in detail.

First, there should not be judgments in your evaluation, as judgment is the entailment itself. Why do we have another judgment once the entailment is done? Figure 2, for example, does not have this in its plot… Why not? Motivation also does not talk about it.

Second, entailment uses three categories: entailment, neutral, and contradiction. Figure 2 has two… What is neutral, what to do when it is neutral, and why have it?

Third, deciding entailment when entailment scores the highest is really not a strong metric. Highest might mean anything… It might be just a little above contradiction… This weak entailment model is probably the reason that you need a judge at the end.

W2. There are unclear points.

a. Section 3.2: “C_em, the inference cost of model m” — which model? I reverse-engineer it to be the “entailment model”; it could have been other models, as your pipeline has multiple models. Please specify clearly.

Having said that, “entailment model” is also not clear, as there is only NLI inside the figure, and there is no single line on what an entailment model is. It can be anything. It can be a sophisticated model doing multiple things, etc. NLI does not mean entailment.

b. Image transmission cost C_tx is not clear. What is the end-to-end workflow here? Is it the cost to send the image from the machine to the user, over the network? How do we calculate the network cost? How much bandwidth do you assume? Or is everything running locally?

c. Eq. 7 has C_tx, C_g, and C_rec for failing cases, but Eq. 8 does not have them for (1-e_m(s)). It only has C_g. Why?

d. What are R_G and R_D? Why do we need them? You already defined the costs in the previous sections, haven’t you?

e. The paragraph after line 245 is a repetition. We understand by now how thresholds work. We need to know how you chose it. The paragraph from 253 explains how they are chosen, but would benefit from an expanded description, with more detail on the reasoning behind the cost optimizations, perhaps with some examples.

f. Also, Figure 3 uses a specific dataset, doesn’t it? Which one is it? Isn’t it a limitation/requirement of your method? What if we don’t know the dataset ahead? This is another good item for the limitations and discussion section.

g. Another issue of the paper is that, at the high level, it contradicts its own problem statement. The problem statement refers to the fact that using only similarity is not enough, as it can eliminate images with low similarity scores. Whereas the solution nevertheless relies only on the top-1 similarity score to decide whether an image should be evaluated in the first place. Top-K candidates could, for example, alleviate this problem. I understand that after getting the top-1 candidate, there are more evaluations in the system; nevertheless, if the similarity score is insufficient, the second candidate might actually be a better fit for the query… This could be a good item for a “discussion and limitations” section. Or, potentially, be incorporated into a future version of the solution.

W3. The presentation can be improved.

a. What is directional semantic coverage? Is this a technical term? If so, can you please cite it? If not, the word “directional” is ambiguous.

b. The right-hand side image of Figure 1 might be disturbing to some people, as it might imply having a violence in it. Let’s please replace it with another one conveying the same message.

c. Contributions (last item of it, generally) should provide, quantitatively, how much improvement the algorithm provides in terms of accuracy/cost.

Questions for authors

Q1. Why do you have a judge at the end of your evaluation, and why is it not in Figure 2?

Q2. Why does your Figure 2 not include "neutral" in your entailment box? What is neutral, and what should we do when it is neutral?

Q3. What is an image transmission cost?

4 comments
AuthorsOpenReview ↗

Response to Reviewer aowg

We thank the reviewer for the detailed feedback and for recognizing the importance of the problem addressed in our work. We acknowledge that several aspects of the entailment verification mechanism, satisfaction evaluation, and cost-aware decision process require clearer explanations. We address these concerns below and clarify the intended scope of EACache.

W1(a, c, d) & W2(a): Entailment Verification Mechanism

We appreciate the reviewer's request for a more detailed explanation of how entailment verification is performed. We would like to clarify that EACache does not aim to train a new entailment model; rather, it introduces a cost-aware framework that determines when pretrained entailment models should be invoked for cache-reuse decisions.

Specifically, we evaluate three pretrained entailment models: RoBERTa, DeBERTa, and ViLT. For RoBERTa and DeBERTa, a vision-language model first generates a detailed caption for each cached image during offline cache construction. The cached caption serves as the premise, and the incoming user query serves as the hypothesis. The entailment model then determines whether the caption semantically supports the requested content. ViLT, in contrast, directly evaluates the image and query without requiring the intermediate caption.

Each verifier predicts three classes: entailment, neutral, and contradiction. We follow a standard three-class NLI inference procedure, where the highest-scoring class determines the prediction. A cached image is accepted only when entailment receives the highest score. Both neutral and contradiction are treated as rejection because neither provides sufficient evidence that the cached image satisfies the query. When verification rejects a candidate, EACache generates a new image.

The inference cost of verifier mm, denoted by Ce,mC_{e,m}, refers specifically to the computational overhead of the selected entailment model. Each model has its own measured inference latency, which is incorporated into the corresponding expected-cost function and threshold derivation.

We acknowledge that the original manuscript did not provide sufficient detail about these inference procedures. We have expanded the system description to explicitly explain the verifier inputs, three-class predictions, acceptance criterion, and respective roles of the textual and multimodal verification models.

W1(b, d) & Q1: Distinction Between Entailment Verification and Satisfaction Evaluation

We thank the reviewer for raising this important concern. We clarify that entailment verification and satisfaction evaluation serve two fundamentally different roles in our framework.

Entailment verification is part of online inference.

EACache uses RoBERTa, DeBERTa, or ViLT to determine whether a retrieved image should be returned to the user. These models provide predictions about semantic entailment but are not assumed to be perfectly accurate.

Satisfaction evaluation in experiments and real deployment.

In real deployment, satisfaction is determined by actual user feedback after receiving a cached image. In our experiments, we use GPT-5.2 as a proxy for user satisfaction to assess whether an image satisfies its corresponding query. These experimental judgments provide the labels required for estimating the probability of successful cache reuse and evaluating the resulting caching decisions.

The distinction is particularly important because a verifier can make incorrect predictions. For example, a verifier may classify an image–query pair as entailed even though the image does not fully satisfy the query. Conversely, it may reject a valid cached image. The independent satisfaction evaluation in our experiments allows us to quantify these errors and account for their consequences in the cost model.

Importantly, GPT-5.2 is used only in our experiments and is not part of the deployed EACache system. It does not determine online cache-reuse decisions or introduce additional online inference overhead. In real deployment, user feedback provides the satisfaction signal.

We selected GPT-5.2 after the authors manually examined its image–query satisfaction judgments and found them consistent with their own assessments. Since our experiments involve more than 10,000 image–query pairs, we use GPT-5.2 as a scalable proxy for manual satisfaction evaluation. We acknowledge that this constitutes qualitative author validation rather than a formally quantified human-agreement study.

We have clarified this separation in the manuscript and revised Figure 2 to distinguish online verification from user-feedback-based satisfaction evaluation.

W1(b) & W2(b, c, d) & Q3: Interpretation of the System Cost Model

We appreciate the opportunity to clarify the different cost components and their roles in EACache.

AuthorsOpenReview ↗

continue from the previous response

Image transmission cost.

The transmission cost represents the latency associated with delivering an image from the service provider to the user. In our experiments, this is modeled as a fixed service-latency parameter and applied consistently across all caching policies.

Generation and recovery costs.

Our objective is to minimize the total time required for a user to obtain a satisfactory image. Image generation cost represents the time required to produce a new image. Recovery cost represents the additional time incurred when a user receives an unsatisfactory cached image, recognizes the failure, and submits another request. This user-side re-prompting delay is separate from the time required to generate and transmit the replacement image. All cost components are expressed in seconds.

Therefore, an unsuccessful cache return incurs the initial transmission time, followed by the additional recovery time and the generation and transmission times required to deliver a satisfactory replacement. These components are explicitly included in our direct-reuse expected-cost formulation.

Why verification rejection incurs only generation cost.

When an entailment verifier rejects a cached candidate, the cached image is never transmitted to the user. Instead, the system immediately generates and transmits a new image. Since the verification cost function already includes the transmission cost for the image eventually delivered, only the additional generation cost is required in the verifier-rejection term.

In contrast, when a verifier incorrectly accepts an unsatisfactory cached image, the system incurs additional regeneration, retransmission, and recovery time.

Expected-cost functions.

The generation, direct-reuse, and verification cost functions represent three alternative actions available to the caching system. Their purpose is to compare the expected total times of these alternatives at a given similarity score, allowing EACache to determine whether a cached image should be returned directly, verified, or replaced through generation.

We have revised the cost-model description to explicitly distinguish the meanings of generation, transmission, and recovery costs, and explain why the different actions incur different cost components.

W2(e, f): Threshold Selection and Dataset Dependence

We thank the reviewer for requesting additional details regarding the threshold-selection procedure.

Threshold estimation.

We estimate the conditional probabilities using labeled image–query pairs from the threshold-learning dataset. The pairs are grouped into adaptive similarity bins, where we calculate the observed satisfaction rate, verifier acceptance rate, and satisfaction rate among verifier-accepted pairs. We then fit smooth, monotonically increasing logistic probability curves using the observed success and failure counts, giving greater influence to bins containing more observations.

The fitted probability curves are incorporated into our expected-cost model to determine the decision boundaries. The lower threshold is obtained where generation and verification have equal expected costs, while the upper threshold is obtained where verification and direct reuse have equal expected costs. All thresholds are learned exclusively from threshold-learning data and remain fixed during held-out evaluation.

Dataset dependence and Figure 3.

We clarify that Figure 3 is based on the SNLI-VE/Flickr30K dataset, illustrating the expected-cost curves and learned thresholds for the three entailment models under that experimental setting.

The thresholds are data-dependent because the relationships among similarity, verifier acceptance, and user satisfaction can vary across datasets. Accordingly, we independently estimate the probability functions and decision thresholds for each evaluated dataset rather than transferring the thresholds learned on SNLI-VE/Flickr30K to COCO or DiffusionDB.

Importantly, EACache does not require prior knowledge of future incoming queries or the entire deployment dataset. It requires only a representative calibration sample to estimate the probability functions and determine the thresholds before deployment.

If the query or cache distribution changes substantially, the same calibration procedure can be repeated using more recent representative data. We have clarified this requirement and explicitly identified the dataset used in Figure 3 in the revised manuscript.

AuthorsOpenReview ↗

continue from the previous response

W2(g): Top-1 Retrieval, the Role of Similarity, and Extension to Top-K

We agree with the reviewer that restricting retrieval to the highest-similarity candidate may overlook valid images with lower similarity scores.

However, we would like to emphasize that although similarity alone cannot reliably determine whether a cached image satisfies a user query, it remains a highly informative and computationally efficient signal. Our motivation is not to replace similarity-based retrieval with entailment verification, but to combine their complementary strengths.

Similarity allows EACache to efficiently narrow the candidate space and distinguish cases where direct reuse or generation is preferable without invoking an entailment model. Specifically, high-similarity candidates can be returned directly, while sufficiently low-similarity candidates trigger generation. Only candidates in the intermediate similarity region require entailment verification, substantially reducing the need to perform expensive semantic verification on every cached image.

Our current framework uses top-1 retrieval to isolate the benefits of selective entailment verification and maintain a tractable cost-aware decision policy. Nevertheless, we acknowledge that the highest-similarity candidate is not necessarily the most semantically suitable image, and a lower-ranked candidate may better satisfy the query.

Extending EACache to top-K retrieval could further improve cache-reuse opportunities by allowing additional candidates to undergo selective verification. Such an extension would require accounting for additional verification costs and determining when to stop examining candidates and invoke generation. We consider this an important direction for future work and have explicitly acknowledged the top-1 restriction in the manuscript.

W3(a): Clarification of Directional Semantic Coverage

We thank the reviewer for highlighting this terminology issue.

By directional semantic coverage, we mean that the cached image must provide sufficient visual evidence to support the content requested by the incoming query, while allowing the image to contain additional details not specified in the query.

For example, an image depicting two people standing beside several vehicles may satisfy the query "people near cars," even though the image contains additional objects and visual details. The reverse implication does not necessarily hold.

This asymmetric relationship motivates using visual entailment rather than relying exclusively on symmetric image–text similarity. Our formulation follows the premise–hypothesis interpretation of visual entailment, where the image provides the evidence and the user query specifies the content that must be supported.

We have clarified this definition and its relationship to existing visual-entailment formulations.

W3(b): Figure 1 and Ethical Concerns

We appreciate the reviewer for bringing this issue to our attention. We agree that the original example in Figure 1 may be interpreted as depicting a potentially disturbing situation.

We have replaced the right-hand example with a neutral image–query pair that illustrates the same distinction between semantic entailment and image–text similarity, without potentially disturbing visual content.

W3(c): Quantitative Contributions

We agree that the contributions should more clearly communicate the quantitative benefits of EACache.

In response to the concerns about the original similarity-only baseline, we introduced a stronger cost-calibrated baseline using the same expected-cost formulation as EACache.

Under this comparison, EACache reduces realized mean system cost by 5.64–8.10% on SNLI-VE/Flickr30K and 0.67–2.28% on COCO, depending on the selected entailment model.

We additionally evaluated EACache using AI-generated images from DiffusionDB with automatically generated Qwen3-VL-8B captions and generalized incoming queries derived from the original DiffusionDB generation prompts. Under this setting, EACache achieves 0.85–1.51% lower mean system cost than the similarity-only baseline at τ=0.205077\tau=0.205077.

These additional experiments provide a more rigorous assessment of the framework and demonstrate that its cost advantage persists under the evaluated stronger baselines and generated-image settings. We have revised the contribution statements to incorporate these quantitative findings while acknowledging the smaller improvements under stronger comparisons.

AuthorsOpenReview ↗

continue from the previous response

Q1: Why Is There a Judge at the End of the Evaluation, and Why Is It Not Included in Figure 2?

GPT-5.2 is used only in our experiments as a proxy for user satisfaction. Its purpose is to determine whether cached images satisfy their corresponding queries, allowing us to obtain satisfaction labels for threshold learning and experimental evaluation.

In real deployment, actual user feedback determines whether a cached image is satisfactory. When a returned cached image fails to satisfy the user, recovery is triggered.

Figure 2 illustrates the deployed EACache workflow, in which similarity retrieval and selective entailment verification determine whether to return a cached image or generate a new one. User feedback determines satisfaction after a cached image is returned.

Since GPT-5.2 is only used in our experiments and is not invoked during actual online query processing, it is not included in Figure 2. We have revised the figure and its caption to make this distinction explicit.

Q2: Why Does Figure 2 Not Include Neutral in the Entailment Box? What Happens When the Prediction Is Neutral?

The entailment models predict three classes: entailment, neutral, and contradiction. Neutral indicates that the available evidence is insufficient to establish either that the query is supported or that it is contradicted by the image.

For cache reuse, we require positive evidence that the retrieved image satisfies the request. Therefore, a candidate is accepted only when entailment is the highest-scoring class. Both neutral and contradiction lead to rejection, after which the system generates a new image.

We have revised Figure 2 to explicitly show the neutral outcome alongside entailment and contradiction.

Q3: What Is an Image Transmission Cost?

Image transmission cost refers to the time required to deliver an image from the service provider to the user over the network. It is distinct from image-generation time and entailment-verification time.

In our evaluation, transmission is modeled as a fixed latency of 2 seconds per delivered image, applied consistently across all policies. If an unsatisfactory cached image is returned, the system incurs transmission time again when delivering the regenerated image.

Summary

We thank the reviewer for the detailed feedback. We have clarified the entailment-verification mechanism, the distinction between experimental satisfaction judgments and actual user feedback in deployment, system-cost accounting, threshold estimation, and the limitations of top-1 retrieval. Together with the revised figures, stronger similarity-only baselines, and additional AI-generated-image experiments, these revisions strengthen the evaluation and presentation of EACache.

Reviewer_R4evOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EACache, an entailment-aware caching framework for text-to-image generation. The key motivation is that image-text similarity does not necessarily reflect whether a cached image semantically satisfies a user query. To address this limitation, the authors introduce a cost-aware policy that chooses among directly reusing a cached image, performing entailment verification, and generating a new image. The thresholds are derived from expected system costs, incorporating image generation, verification, transmission, and recovery costs. Experiment results demonstrate the reduction in cost compared with similarity-only caching baselines.

Critical strengths
  • Practical motivation. The paper focuses on a efficiency problem in generative AI serving. The discrepancy between semantic similarity and directional entailment is well motivated and examples are provided to illustrate the problems of similarity-based caching.

  • Cost-aware formulation. The paper explicitly considers the trade-off between verification overhead, generation, and cache returns, which provides a comparatively systematic way to select different verification models.

  • Empirical analysis. The paper evaluates multiple entailment models and provide analysis of cache-hit quality, false-turn rates, and system-cost decomposition.

Critical weaknesses
  • Limited methodological novelty. The main technical contribution appears to be the integration of existing image-text similarity, pretrained entailment verification, and cost-aware threshold selection. The each component is well established, which lacks novelty and depth on methodology. The paper would benefit from a clear and in-depth explanation of what is technically novel beyond simply applying these existing components to image generation caching.

  • The baseline is weak. The main experiment considers three thresholds (0.2, 0.25, and 0.27), and reports the best among them as the baseline. The reason why these thresholds are selected is not claimed, which raises the concern that whether the improvements are due to insufficient threshold optimization. More importantly, the paper's own results suggest that a stronger similarity-only baseline can be constructed. For example, using the DeBERTa upper threshold of 0.234591 as shown in Table 1, Tables 3 and 4 imply 3,045 direct returns, including 2,165 successful returns and 880 unsuccessful returns. Under the paper's cost formulation, a similarity-only policy using this threshold would achieve an estimated mean cost of: $ [\frac{4000(0.02)+1835(12)+4880(2)+880(10)}{4000}=10.165.] $ This is substantially lower than the reported best similarity-only cost of 10.894, even without explicitly optimizing the baseline threshold. Although EACache still achieves lower cost under this comparison, the improvement becomes considerably smaller.

  • Lack of experiments on text-to-image models. The experiments are conducted on natural images sampled form Flickr30K and COCO dataset, rather than images generated by text-to-image models, like Stable Diffusion and Flux-Context. The proposed framework should be validated under the deployed serving system, which is close to the author's motivation.

  • The calculation of cost does not capture practical deployment overhead. The cost of generating detailed captions of newly cached images is not explicitly incorporated into the end-to-end cost. The verification latency is measured under batched settings, which can not reflect online per-request latency.

Questions for authors

Refer to the weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer R4ev

We sincerely thank the reviewer for the thoughtful and constructive feedback. We appreciate the recognition of the practical motivation, cost-aware formulation, and empirical analysis of EACache. We have carefully considered the concerns and conducted additional experiments to strengthen our evaluation.

Below, we address the four major concerns regarding (1) methodological novelty, (2) the similarity-only baseline, (3) evaluation on AI-generated images, and (4) practical deployment costs.

W1: Limited Methodological Novelty

We appreciate the reviewer's observation and would like to clarify that the primary technical contribution of EACache is not the development of a new image–text similarity model or entailment verifier, but rather a unified cost-aware decision framework that determines when semantic verification is beneficial for image cache reuse.

Conventional similarity-only caching makes a binary decision between directly returning a cached image and generating a new one. EACache introduces entailment verification as a third possible action, with its use determined by the expected costs of all three alternatives.

Specifically, our formulation jointly considers:

  1. The probability that a cached image satisfies the query.
  2. The probability that an entailment verifier accepts the candidate.
  3. The satisfaction probability conditioned on verifier acceptance.
  4. The costs of verification, generation, transmission, and unsuccessful-return recovery.

These quantities define three expected-cost functions:

RG(s),RD(s),RV,m(s),R_G(s),\qquad R_D(s),\qquad R_{V,m}(s),

which allow the system to select the action with the lowest expected cost at a given similarity score.

Unlike applying a pretrained entailment model to every retrieved image, our formulation explicitly determines whether the additional semantic information provided by verification justifies its inference overhead.

The resulting decision policy depends on both the statistical behavior and inference cost of the selected verifier.

Furthermore, EACache provides a general mechanism for comparing different entailment models under the same system-cost objective.

As demonstrated by our experiments, the verifier with the highest entailment precision does not necessarily produce the lowest overall service cost. This distinction motivates evaluating verification models based on their downstream caching decisions rather than standalone classification performance.

We agree that the individual components are established techniques, and we will more clearly position our contribution as their cost-aware integration and selective decision optimization for text-to-image caching, rather than as a new entailment-learning architecture.

W2: Stronger Similarity-Only Baseline

We thank the reviewer for identifying the weakness in our original similarity-only baseline and for providing a concrete numerical example.

We agree that the original thresholds {0.20,0.25,0.27}\{0.20,0.25,0.27\} did not constitute a sufficiently optimized comparison.

Cost-calibrated similarity-only baseline.

To address this concern, we introduced a principled similarity-only threshold-selection procedure using the same expected-cost formulation as EACache.

For similarity-only caching, direct reuse is preferable to generating a new image when

RD(s)≤RG(s),R_D(s)\leq R_G(s),

which gives

g^(s)≥Ctx+CrecCg+Ctx+Crec=0.5.\hat{g}(s)\geq\frac{C_{\mathrm{tx}}+C_{\mathrm{rec}}}{C_g+C_{\mathrm{tx}}+C_{\mathrm{rec}}}=0.5.

We locate this crossing using the satisfaction-probability curve fitted exclusively on the threshold-learning data, without accessing held-out evaluation labels.

This produces development-calibrated similarity-only thresholds of 0.220482 for SNLI-VE/Flickr30K and 0.230615 for COCO.

Table 1. Realized mean system cost per incoming query across datasets. Lower is better.

PolicySNLI-VE/Flickr30KCOCO
Always generate14.02014.020
Similarity-only (τ=0.20\tau=0.20)10.8948.239
Similarity-only (cost-calibrated)10.4117.858
EACache–RoBERTa9.7377.693
EACache–DeBERTa9.8247.679
EACache–ViLT9.5687.805

Table 2. Cost reduction relative to the cost-calibrated similarity-only baseline.

EACache variantSNLI-VE/Flickr30KCOCO
RoBERTa6.47%2.10%
DeBERTa5.64%2.28%
ViLT8.10%0.67%

The calibrated similarity-only baseline improves upon the original baseline on both datasets. Nevertheless, all three EACache variants retain lower realized costs under the same cost assumptions.

AuthorsOpenReview ↗

continue response from the previous

Reproduction of the reviewer's proposed comparison.

We also reproduced the reviewer's suggested similarity-only threshold of τ=0.234591\tau=0.234591, which corresponds to the independently learned DeBERTa upper threshold.

On SNLI-VE/Flickr30K, applying this threshold as a similarity-only cutoff produces 3,045 direct cache returns, including 2,165 satisfactory returns and 880 unsatisfactory returns.

The resulting mean system cost is

Cˉ=4000(0.02)+1835(12)+4880(2)+880(10)4000=10.165.\bar{C}=\frac{4000(0.02)+1835(12)+4880(2)+880(10)}{4000}=\boxed{10.165}.

This exactly matches the reviewer's calculation.

We acknowledge that this is lower than the 10.411 obtained using our development-calibrated similarity-only threshold.

The difference arises because the calibrated threshold is selected from the fitted development-set satisfaction-probability curve, rather than from the realized evaluation outcomes.

Specifically, among the 582 held-out queries with similarity scores in the interval

0.220482≤s<0.234591,0.220482\leq s<0.234591,

only 250 (42.96%) are satisfactory, while 332 are unsatisfactory.

Although the fitted probability curve predicts satisfaction probabilities of at least 50% within this interval, the observed satisfaction rate is below this break-even point.

Consequently, the more conservative threshold avoids 332 unsuccessful cache returns at the expense of 250 successful returns, reducing the realized mean cost by

(332−250)(12)4000=0.246.\frac{(332-250)(12)}{4000}=0.246.

We therefore distinguish the development-calibrated threshold from the stronger observed threshold identified by the reviewer, rather than claiming that our calibrated threshold is the empirical optimum on the held-out test set.

Importantly, EACache continues to outperform the reviewer's stronger baseline.

Compared with the mean cost of 10.165, RoBERTa-, DeBERTa-, and ViLT-based EACache achieve costs of 9.737, 9.824, and 9.568, corresponding to reductions of 4.21%, 3.35%, and 5.87%, respectively.

We acknowledge that these improvements are smaller than originally reported. The revised analysis more accurately characterizes the magnitude of EACache's advantage under stronger similarity-only comparisons.

W3: Evaluation on Text-to-Image Generated Images

We agree that experiments using only natural images from Flickr30K and COCO do not fully represent the intended deployment setting of EACache.

We initially used these datasets to evaluate the general image–query caching mechanism, which does not inherently depend on whether cached images are naturally sourced or AI-generated.

To address the deployment concern, we conducted an additional experiment using AI-generated images with detailed captions automatically produced by Qwen3-VL-8B.

We construct generalized incoming queries based on the original generation prompts associated with DiffusionDB images. We evaluate EACache on 4,000 such incoming queries to investigate whether previously generated images can satisfy related requests.

Unlike the original benchmark evaluation, this experiment directly considers cached AI-generated images and automatically generated captions, two components central to the proposed caching framework.

Table 3. Realized mean system cost per incoming query on AI-generated images with generalized queries derived from DiffusionDB generation prompts. Lower is better.

PolicyMean cost
Always generate14.020
Similarity-only (τ=0.205077\tau=0.205077)7.741
EACache–RoBERTa7.645
EACache–DeBERTa7.624
EACache–ViLT7.675

Table 4. Cost reduction relative to the similarity-only baseline at τ=0.205077\tau=0.205077 on AI-generated images.

EACache variantCost reduction
RoBERTa1.24%
DeBERTa1.51%
ViLT0.85%

The evaluated similarity-only threshold is τ=0.205077\tau=0.205077.

All three EACache variants achieve lower realized costs than the similarity-only baseline.

Although the improvements are modest, these results provide additional evidence that selective entailment verification remains beneficial when the cache contains AI-generated images with automatically generated captions.

We acknowledge that the performance advantage depends on the characteristics of the incoming queries and the reliability of similarity-based retrieval.

The additional experiment therefore strengthens the applicability evidence for EACache without implying a universal improvement across all deployment conditions.

W4: Practical Deployment Cost and Verification Latency

We thank the reviewer for raising the important question of practical deployment overhead.

AuthorsOpenReview ↗

continue response from the previous

Cost-aware theoretical formulation.

We emphasize that the proposed theoretical framework explicitly captures the trade-offs among image generation, entailment verification, image transmission, and unsuccessful-return recovery.

Rather than assuming verification always improves efficiency, EACache evaluates the expected cost of each available action:

a∗(s)=arg⁡min⁡a∈{G,V,D}Ra(s).a^*(s)=\arg\min_{a\in\{G,V,D\}}R_a(s).

The decision thresholds consequently depend on both the system-cost parameters and the estimated probabilities of successful cache reuse.

This formulation allows the decision policy to adapt to different cost configurations.

For instance, an increase in verification latency makes verification less attractive unless it avoids sufficient downstream generation or recovery cost.

Conversely, higher generation or recovery costs can increase the value of accurate verification.

Measured model-specific verification latency.

Our experiments incorporate the measured inference latency of each entailment model rather than assuming a uniform verification overhead.

The measured verification costs for RoBERTa, DeBERTa, and ViLT are included in the system-cost calculation and used to derive their model-specific decision thresholds.

We acknowledge that these verification latencies were measured under batched inference conditions and may differ from online per-request latency.

Our experimental results therefore characterize performance under the stated inference configuration, rather than claiming deployment-independent end-to-end latency reductions.

Automatic caption-generation overhead.

We clarify that detailed image captions are generated offline when images are added to the cache and stored for subsequent reuse. Therefore, caption generation does not introduce additional inference overhead during online query processing.

Summary

We thank the reviewer for the constructive feedback. The stronger similarity-only baselines, reproduction of the reviewer's numerical comparison, and additional AI-generated-image experiments strengthen our evaluation. We have also clarified EACache's methodological contribution and practical deployment assumptions, while acknowledging the limitations of the current cost measurements.

Thank you for the detailed rebuttal and additional experiments. I appreciate the author's efforts.

However, my main concerns remain. The methodological novelty is limited, as the approach integrates existing techniques within a cost-aware decision framework. Moreover, the improvements over the stronger baseline are relatively modest (3.35%), and become smaller on AI-generated images (0.85%). The practical benefits also remain uncertain without an end-to-end evaluation of online inference latency and deployment costs.

The paper does not meet the criteria of ICLR. I maintain my original rating and recommendation.

Reviewer_NSwzOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces EACache, a semantic caching framework for text-to-image services. It retrieves the most similar cached image and uses two cost-derived thresholds to choose among direct reuse, entailment verification, and new image generation. Text-based verifiers operate on stored image captions, while ViLT directly verifies the image-query pair. On SNLI-VE/Flickr30K and COCO, EACache reports lower modeled service cost than similarity-only caching and always generating. A controlled SNLI-VE experiment further shows that entailment verification can recover valid cache hits rejected by conservative similarity thresholds.

Critical strengths
  1. Appropriate formulation of cache reuse as directional entailment. Distinguishing whether an image satisfies all requested content from symmetric image-text similarity is well motivated. Selective verification is a sensible way to recover valid low-similarity cache hits without verifying every request.
  2. Recovery-aware decision model. The analysis incorporates verification, generation, transmission, and failed-return recovery costs. Evaluating text-based and multimodal verifiers also illustrates that verifier accuracy alone does not determine end-to-end system cost.
Critical weaknesses
  1. The main evaluation does not represent the claimed deployment setting. Cached candidates are natural Flickr30K and COCO images, while incoming requests are benchmark hypotheses or captions rather than representative user prompts. The paper does not clearly specify how cache captions are generated, and satisfaction in the main system-cost evaluation is assessed by GPT-5.2 without reported human validation of the judge. Consequently, the experiments do not establish that EACache works reliably on generated images with automatic captions and realistic requests.
  2. The reported cost advantage is not established under matched or robust baselines. The similarity-only baseline tests only three thresholds rather than optimizing its threshold on the same threshold-learning data. Moreover, generation, transmission, search, and recovery costs are largely assumed, with no sensitivity analysis or measured end-to-end deployment latency. Since the improvement over the best evaluated similarity-only baseline is approximately 12.2% on SNLI-VE/Flickr30K and 6.8% on COCO, these choices could materially change the conclusion.
Questions for authors
  1. Clarification and additional validation of the evaluation setting. How are the stored captions used by RoBERTa and DeBERTa obtained? Are they human dataset annotations? Results using generated images, automatic captions, representative user prompts, and human validation of the GPT-5.2 judge would be necessary to establish applicability to the target setting.
  2. Matched baseline and cost robustness. What are the results when the similarity-only threshold is optimized on the same threshold-learning split? Does EACache remain preferable across reasonable ranges of generation, transmission, and recovery costs, and under measured end-to-end latency?
  3. Clarification of threshold estimation. How are (g(s)), (e_m(s)), and (r_m(s)) fitted? Are the three minimum-cost regions guaranteed to be ordered and contiguous, and how are multiple or uncertain cost-curve intersections handled?
3 comments
AuthorsOpenReview ↗

Response to Reviewer NSwz

We thank the reviewer for the thoughtful and constructive feedback. We appreciate the recognition of our motivation for directional entailment-aware caching and our recovery-aware cost formulation.

We have carefully considered the concerns and conducted additional experiments to strengthen the evaluation. Below, we address the major points regarding (1) deployment realism and evaluation reliability, (2) stronger similarity-only baselines and cost modeling, and (3) probability estimation and threshold derivation.

W1 & Q1: Evaluation on Generated Images and Reliability of Satisfaction Judgments

Evaluation on AI-generated images.

We agree that experiments based solely on Flickr30K and COCO do not fully reflect the intended text-to-image deployment setting.

To address this concern, we conducted an additional evaluation based on DiffusionDB using AI-generated images with detailed captions automatically produced by Qwen3-VL-8B. The incoming queries are generalized from the original text-to-image generation prompts associated with DiffusionDB images, rather than taken from human-written natural-image captions. We use 4,000 such incoming queries to investigate whether previously generated images can satisfy related user requests.

The results are summarized below.

Table 1. Realized mean system cost per incoming query on AI-generated images with prompt-derived generalized queries. Lower is better.

PolicyMean cost
Always generate14.020
Similarity-only (τ = 0.205077)7.741
EACache–RoBERTa7.645
EACache–DeBERTa7.624
EACache–ViLT7.675

Table 2. Cost reduction relative to the similarity-only baseline at τ = 0.205077 on AI-generated images.

EACache variantCost reduction
RoBERTa1.24%
DeBERTa1.51%
ViLT0.85%

The evaluated similarity-only threshold is τ = 0.205077.

All three EACache variants continue to outperform the similarity-only baseline. Although the improvements are modest, these results provide additional evidence that the effectiveness of selective entailment verification extends beyond natural-image benchmarks.

Automatic caption generation.

We clarify that EACache does not require human-written captions in deployment. Each image is captioned by a vision-language model when added to the cache.

In the new generated-image experiments, we use Qwen3-VL-8B for this purpose. RoBERTa and DeBERTa subsequently perform textual entailment verification using the generated caption as the premise and the incoming query as the hypothesis, whereas ViLT directly evaluates the image–query pair.

Reliability of the GPT-5.2 satisfaction judge.

We also clarify that GPT-5.2 was selected after the authors manually examined its image–query satisfaction judgments and found them consistent with their own assessments.

Given the scale of our experiments, involving more than 10,000 image–query pairs, GPT-5.2 provides a practical proxy for manually evaluating every pair.

Importantly, GPT-5.2 is not part of the online cache-reuse decision process. It provides satisfaction labels for offline threshold learning and evaluation, whereas online verification decisions are performed by RoBERTa, DeBERTa, or ViLT.

We acknowledge that our manual examination constitutes qualitative validation rather than a formally quantified human-agreement study, and will clarify this distinction in the manuscript.

W2 & Q2: Stronger Similarity-Only Baseline and Cost Robustness

Cost-calibrated similarity-only baseline.

We agree that the original thresholds {0.20, 0.25, 0.27} did not provide a sufficiently optimized similarity-only baseline.

To address this concern, we derived a similarity-only decision threshold using the same expected-cost formulation as EACache.

Specifically, direct reuse is preferred over generation when

RD(s)≤RG(s)R_D(s) \leq R_G(s)

We identify this crossing using the satisfaction-probability curve fitted exclusively on the threshold-learning data.

The resulting development-calibrated thresholds are 0.220482 for SNLI-VE/Flickr30K and 0.230615 for COCO.

Table 3. Realized mean system cost per incoming query across datasets. All policies are evaluated on 4,000 held-out queries per dataset. Lower is better.

PolicySNLI-VE/Flickr30KCOCO
Always generate14.02014.020
Similarity-only (τ = 0.20)10.8948.239
Similarity-only (cost-calibrated)10.4117.858
EACache–RoBERTa9.7377.693
EACache–DeBERTa9.8247.679
EACache–ViLT9.5687.805

Table 4. Cost reduction relative to the cost-calibrated similarity-only baseline.

EACache variantSNLI-VE/Flickr30KCOCO
RoBERTa6.47%2.10%
DeBERTa5.64%2.28%
ViLT8.10%0.67%
AuthorsOpenReview ↗

continue from the previous response

These results demonstrate that EACache maintains lower realized costs than the development-calibrated similarity-only baseline across both datasets and all three entailment models.

We acknowledge that the improvements are smaller than those originally reported and have adjusted our claims accordingly.

Cost modeling and robustness.

We emphasize that our proposed theoretical formulation explicitly accounts for the trade-offs among image generation, entailment verification, transmission, and unsuccessful-return recovery.

Rather than assuming verification is always beneficial, EACache compares the expected costs of three actions and selects the minimum-cost action under the fitted model.

Furthermore, our experiments incorporate measured inference latency for each entailment model instead of assigning a uniform verification cost.

Specifically, RoBERTa, DeBERTa, and ViLT have different measured inference costs, which are incorporated into the realized cost evaluation and corresponding threshold derivation.

The remaining service-cost parameters are held consistent across all policies.

Because the decision thresholds are derived from the expected-cost functions, the theoretical framework naturally accommodates different cost configurations by recalculating the action costs and their decision boundaries.

Thus, the proposed policy is not tied to one fixed set of generation, transmission, or recovery costs.

We note that the current results use measured model-specific verification latency together with specified service-latency parameters, rather than a fully measured end-to-end deployment.

The theoretical model captures sensitivity to these parameters, while a comprehensive empirical sensitivity analysis across deployment configurations remains beyond the reported experiments.

Q3: Probability Estimation and Threshold Derivation

Probability estimation.

We estimate the conditional probabilities using labeled image–query pairs from the threshold-learning dataset. The pairs are grouped into adaptive similarity bins, where we compute the observed user satisfaction rate, verifier acceptance rate, and satisfaction rate among verifier-accepted pairs.

We then fit smooth, monotonically increasing logistic curves using the observed success and failure counts, giving greater weight to bins with more observations.

Threshold derivation.

The fitted probability curves are incorporated into our expected-cost model to determine the decision boundaries. Specifically, the lower threshold is obtained where generation and verification have equal expected costs, while the upper threshold is obtained where verification and direct reuse have equal expected costs.

Both thresholds are learned exclusively from the threshold-learning data and remain fixed during held-out evaluation.

Ordering and contiguity of the decision regions.

We thank the reviewer for raising this question.

Our two-threshold formulation is motivated by the observed relationship between image–query similarity and the estimated conditional probabilities.

In our experiments, we observe that both the probability of user satisfaction g(s)g(s) and the probability of verifier acceptance em(s)e_m(s) increase with similarity.

These trends are captured by our monotonically increasing fitted probability curves and support the observed transition from generation to verification and subsequently to direct cache reuse.

Importantly, our framework naturally reduces to a single-threshold similarity-only policy when the entailment verifier provides no additional reuse opportunities.

Specifically, if the verifier rejects all candidate pairs, then em(s)=0e_m(s)=0, and the expected verification cost becomes

=K+C_{e,m}+C_{\mathrm{tx}}+C_g =R_G(s)+C_{e,m}.$$ Since $C_{e,m}>0$, verification is strictly more expensive than direct generation and is therefore never selected by the minimum-cost policy. Consequently, the decision reduces to comparing direct reuse with generation: $$a^*(s)=\arg\min_{a\in\{G,D\}}R_a(s).$$ The corresponding decision boundary satisfies $$g(\tau_{\mathrm{sim}})=\frac{C_{\mathrm{tx}}+C_{\mathrm{rec}}}{C_g+C_{\mathrm{tx}}+C_{\mathrm{rec}}}.$$ Thus, when verification provides no accepted candidates, EACache reduces to the cost-calibrated similarity-only policy under the fitted probability model. More generally, the three-region structure is supported by the fitted cost curves in our evaluated settings. While the observed monotonic trends provide empirical support for this ordering, they do not constitute a universal mathematical guarantee for arbitrary probability distributions or cost configurations. The underlying decision principle remains selecting the action with the minimum expected cost at each similarity value.

Thank you for the additional evaluations. They address some concerns, but the gains remain modest and lack uncertainty estimates. Human validation of the satisfaction judge remains qualitative, while end-to-end performance and cost sensitivity remain unvalidated. I therefore maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Responses to All Reviewers

Dear ACs, SACs, and PCs,

We sincerely thank all three reviewers for their constructive feedback. We have addressed their concerns through additional experiments, stronger baselines, methodological clarifications, and revised figures. Below is a consolidated summary of our responses.

Reviewer NSwz (Initial Score: 1)

  1. Evaluation on AI-generated images. We initially used SNLI-VE/Flickr30K and COCO to evaluate the general image–query caching mechanism, which is applicable regardless of whether cached images are naturally sourced or AI-generated. We additionally evaluated DiffusionDB using AI-generated images, Qwen3-VL-8B-generated captions, and 4,000 queries derived from original generation prompts. EACache achieves 0.85–1.51% lower mean cost than the optimal similarity-only baseline (τ = 0.205077), confirming its effectiveness on generated images (Response W1 & Q1).

  2. Stronger baselines. We introduced cost-calibrated similarity-only thresholds using the same expected-cost formulation as EACache. EACache maintains 5.64–8.10% cost reductions on SNLI-VE/Flickr30K and 0.67–2.28% on COCO (Response W2 & Q2).

  3. Evaluation reliability. We clarified that GPT-5.2 serves as an offline satisfaction proxy, selected after qualitative examination of its judgments by the authors. Actual user feedback determines satisfaction in deployment (Response W1 & Q1).

  4. Cost modeling. We clarified generation, verification, transmission, and recovery costs, including measured model-specific verification latencies (Response W2 & Q2).

  5. Threshold derivation. We explained adaptive similarity binning, count-weighted logistic fitting, and cost-based threshold selection, including the conditions for a three-region policy and its reduction to similarity-only caching when verification provides no accepted candidates (Response Q3).

Reviewer R4ev (Initial Score: 2)

  1. Methodological novelty. We clarified that EACache's contribution is a recovery-aware, three-action decision framework that determines when entailment verification is economically beneficial, rather than a new entailment model (Response W1).

  2. Reviewer's stronger baseline. We reproduced the suggested threshold τ = 0.234591, confirming a mean cost of 10.165 on SNLI-VE/Flickr30K. EACache maintains cost reductions of 4.21%, 3.35%, and 5.87% for RoBERTa, DeBERTa, and ViLT, respectively (Response W2).

  3. Independent baseline calibration. We derived additional similarity-only thresholds using development data and the same cost formulation, without accessing held-out evaluation labels. We revised our claims to reflect the smaller improvements against these stronger baselines (Response W2).

  4. Generated-image evaluation. We added DiffusionDB experiments with automatically generated captions and prompt-derived incoming queries. All EACache variants outperform the evaluated similarity-only baseline, providing evidence beyond natural-image benchmarks (Response W3).

  5. Deployment considerations. We clarified verifier inference latencies, system-cost assumptions, and offline caption generation during cache construction (Response W4).

Reviewer aowg (Initial Score: 1)

  1. Entailment verification. We clarified how RoBERTa and DeBERTa evaluate caption–query pairs, while ViLT evaluates image–query pairs directly. Only entailment leads to acceptance; neutral and contradiction lead to rejection (Responses W1, W2 & Q2).

  2. Online versus offline evaluation. We distinguished online entailment verification from offline GPT-5.2 satisfaction assessment. In deployment, actual user feedback determines satisfaction, as illustrated in revised Figure 2 (Responses W1 & Q1).

  3. System costs. We clarified that costs are expressed in seconds, including transmission and user-side re-prompting time after unsuccessful returns. Verification rejection triggers generation without the additional recovery cost associated with an unsuccessful cache return (Responses W1, W2 & Q3).

  4. Threshold learning. We identified SNLI-VE/Flickr30K as the source of Figure 3 and explained that thresholds are estimated from representative development data and can be recalibrated when query distributions change (Response W2).

  5. Selective verification and top-1 retrieval. We clarified why similarity remains an efficient filtering signal, why verification is applied selectively, and the limitations and potential extensions of top-1 retrieval (Response W2).

  6. Presentation improvements. We clarified directional semantic coverage, replaced the potentially disturbing example in Figure 1, revised Figure 2, and updated quantitative comparisons using stronger baselines and AI-generated images (Response W3).

Sincerely,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.