acceptodds
Withdrawn from ICLR 2027

QMORE: Robust Quantum Image–Text Retrieval via Encoder Sharing and Diversity-Aware Rank Fusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We present QMORE, a hybrid quantum–classical framework for cross-modal image–text retrieval. QMORE learns a shared image–text geometry by reusing a single variational quantum encoder across both modalities, paired with lightweight modality adapters that preserve modality-specific flexibility while keeping quantum parameters shared. To reduce run-to-run sensitivity, QMORE ensembles independently trained models, weights members by retrieval strength and diversity, and fuses predictions via weighted rank voting. Across three benchmarks, QMORE improves Recall@1 over the strongest parameter-matched baseline by , , and points on SVO Probes, Captioned CIFAR-10, and Flickr30k—a – relative improvement on every dataset—while also achieving the best mean and median rank throughout. The advantage persists against classical encoders with more encoder-core parameters, and under a calibrated noise model of a near-term IBM device QMORE retains of its Recall@1 while the quantum baseline collapses to near-random.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Quantum, QML, Multimodal

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WFK9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes QMORE, a hybrid quantum–classical framework for image–text retrieval. It introduces a shared variational quantum encoder with lightweight modality-specific adapters to improve parameter efficiency.They also use a rank-fusion strategy to enhance retrieval robustness. Moreover, authors test their method on three datasets including SVO Probes, Captioned CIFAR-10, and Flickr30K, which shows the clear improvement over the classical baselines and parameter-matched quantum. And the robustness is further tested under simulated quantum hardware noise.

Critical strengths

1.The paper identifies two practical challenges in quantum multimodal retrieval: parameter redundancy across modalities and performance variability across independent training runs. The overall framework is clearly motivated, and its components are logically connected to the identified challenges. 2.The authors conduct experiments on three datasets, including SVO Probes, Captioned CIFAR-10, and Flickr30k, demonstrating consistent improvements over parameter-matched quantum and classical baselines. The paper also includes parameter-scaling comparisons, component ablation studies, and evaluations of different ensemble strategies. These experiments provide useful evidence for the effectiveness of the proposed framework and help distinguish the contributions of its individual components.

Critical weaknesses

1.The experiments mainly compare against lightweight, parameter-matched baselines without directly evaluating established image–text retrieval models such as CLIP. Furthermore, retrieval is evaluated within 200-candidate subsets rather than standard full-gallery settings, limiting the practical significance and comparability of the reported results. 2.Insufficient evidence of quantum-specific advantages. When applying the same diversity-aware fusion strategy, QMORE improves Flickr30k R@1 only from 13.5% to 14.4% over the strongest classical baseline. Moreover, the noise robustness experiments rely entirely on simulation rather than actual quantum hardware, leaving the practical benefits of the quantum encoder insufficiently established.

Questions for authors

1.In Table 4, QMORE achieves 14.4% R@1 on Flickr30k, compared to 13.5% for the shared classical baseline using the same accuracy–diversity fusion strategy. Could the authors provide additional experiments with stronger parameter-matched classical architectures to better demonstrate that the improvements originate from the quantum encoder rather than parameter sharing or ensemble fusion? 2. The experiments evaluate retrieval within subsets of 200 image–text pairs rather than using a standard full-gallery retrieval setting. Could the authors investigate whether the reported improvements remain consistent with larger candidate pools?

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

This paper proposes QMORE, a hybrid quantum–classical framework for image-to-text retrieval that combines a shared variational quantum encoder with modality-specific phase adapters and accuracy–diversity weighted rank fusion. The proposed architecture reuses a single quantum encoder across image and text modalities to reduce parameter duplication, while an ensemble of independently trained models is employed to improve retrieval robustness.

The authors evaluate QMORE on SVO Probes, Captioned CIFAR-10, and Flickr30k using parameter-matched quantum and classical baselines. The reported results show improvements in Recall@K and retrieval ranks, along with relatively stable performance under simulated quantum hardware noise.

Although the framework is clearly motivated and the empirical results are encouraging, the current evidence does not sufficiently isolate the contribution of the quantum representation from conventional parameter sharing and ensemble learning. Furthermore, several aspects of the baseline design and evaluation protocol limit the strength and generalizability of the conclusions.

Critical strengths
  1. A well-motivated and coherent architecture for parameter-efficient multimodal retrieval.

The combination of a shared encoder, modality-specific adapters, and rank-based ensemble aggregation provides a reasonable architectural approach to constrained multimodal representation learning. The paper presents a clear formulation and includes component ablations demonstrating that both parameter sharing and ensembling contribute to retrieval performance.

  1. Consistent empirical improvements across multiple controlled settings.

The experiments cover three image–text datasets and include comparisons with parameter-matched classical and quantum baselines. The authors also examine parameter scaling, alternative ensemble strategies, and simulated hardware noise. These experiments provide useful initial evidence that the proposed configuration can improve retrieval performance under the evaluated resource constraints.

Critical weaknesses
  1. The quantum-specific contribution is not sufficiently established, and the modality adapter has a potentially restrictive mathematical structure.

The main concern is whether the reported improvement can be attributed to a meaningful quantum-specific inductive bias rather than the combination of parameter sharing, nonlinear feature transformations, and ensemble aggregation.

According to Appendix C.2, the quantum circuit consists of RX data encoding, RX variational rotations, and CNOT entangling gates, followed by modality-specific RZ rotations and Pauli-Y measurements. Under this circuit structure, the expectation value of Pauli-X on each individual qubit is identically zero before the modality adapter. Consequently, the RZ adapter changes the corresponding Pauli-Y expectation only through a multiplicative cosine factor.

This suggests that the proposed modality adapters effectively perform coordinate-wise scaling rather than providing the more general measurement-basis adaptation suggested by the paper. A corresponding classical multiplicative or diagonal-scaling adapter is not included in the reported comparisons.

Moreover, the primary shared classical baseline uses a linear alignment core with additive bias adapters, whereas the quantum model employs a multilayer nonlinear quantum feature map. Matching the number of trainable parameters alone does not establish equivalence in functional expressivity or inductive bias.

The claimed 80-parameter budget also excludes 5,136 trainable parameters in the shared classical projection front-end. Although this exclusion is consistent across the compared models, it limits the interpretation of the claimed parameter efficiency at the level of the complete trainable retrieval architecture.

These issues are consequential because the paper's central contribution depends on demonstrating that the quantum parameterization provides benefits beyond simpler classical alternatives. The current comparisons do not sufficiently establish this distinction.

  1. The retrieval evaluation and statistical evidence are insufficient to substantiate the claimed robustness and generalizability.

The evaluation divides the 2,000 test pairs into ten chunks of 200 candidates and computes retrieval metrics separately within each chunk. This is a substantially easier retrieval setting than searching across the complete held-out gallery, and it does not establish how the method performs under larger candidate pools or conventional image–text retrieval protocols.

Furthermore, the reported mean and standard deviation are computed across evaluation chunks rather than independently repeated training-and-ensemble experiments. Although each ensemble contains ten independently trained members, the study does not quantify performance variability across separately reconstructed ensembles. This is particularly important because reducing run-to-run variability is one of the paper's main motivations.

The comparison in Table 4 also substantially narrows the quantum-specific improvement. When both QMORE and the strongest shared classical baseline use the same accuracy–diversity fusion, their Flickr30k Recall@1 scores are 14.4% and 13.5%, respectively. The absolute difference is only 0.9 percentage points. Although the authors report statistical significance from paired tests across ten chunks, these tests do not account for variability arising from independent ensemble training runs.

Finally, the noise-robustness experiments are conducted using a single simulated hardware-derived noise model, without a broader characterization of variation across repeated noisy evaluations or hardware mappings. The results are promising, but they are not sufficient to establish generalized robustness under realistic NISQ conditions.

Taken together, these limitations make it difficult to determine whether the improvements are stable, practically meaningful, and specifically attributable to the proposed quantum components. They substantially weaken the empirical support for the paper's central claims.

Questions for authors

Q1. Quantum-specific contribution and modality adapters

Given the RX/CNOT circuit structure and Pauli-Y readout, can the authors clarify whether the pre-adapter Pauli-X expectations are identically zero? If so, how does the proposed RZ adapter differ functionally from a classical coordinate-wise scaling operation? A controlled comparison with classical multiplicative adapters and nonlinear alignment heads would materially affect the assessment of the quantum-specific contribution.

Q2. Statistical robustness and baseline comparison

Can the authors report results across independently repeated training-and-ensemble runs, with identical fusion rules applied to quantum and classical models? In particular, does the 0.9-percentage-point Recall@1 improvement on Flickr30k remain consistent across independent ensemble realizations?

Q3. Retrieval evaluation and scalability

Can the authors clarify why retrieval is evaluated over 200-candidate chunks rather than the complete held-out candidate pool? How does the relative performance change when increasing the gallery size or applying a conventional Flickr30k retrieval protocol?

Reviewer_yYkhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QMORE, a hybrid quantum–classical framework for image-to-text retrieval. Frozen ResNet-18 and BERT-tiny features are compressed to 8 dimensions by per-modality linear layers and then passed through a single 8-qubit variational circuit shared by both modalities. Per-modality RZ rotations before the Pauli-Y readout serve as lightweight modality adapters, giving 80 quantum parameters in total. To reduce run-to-run variance, ten independently trained models are combined by weighted rank fusion, where each member's weight mixes its retrieval accuracy on a calibration set and the diversity of its cross-modal distance structure. Experiments on SVO Probes, a self-constructed Captioned CIFAR-10, and Flickr30k compare QMORE with quantum and classical baselines matched in the 80-parameter encoder core, and report R@1 gains of 2–3 points over the best baseline, together with ablations, an ensemble comparison and a noisy simulation under an IBM device noise model. All evaluations are image-to-text only, with 200 candidates per query.

Critical strengths
  1. The design of the shared quantum encoder is supported by experiments. Removing the shared encoder lowers performance on all three datasets (Table 3). The paper also includes a shared classical baseline as a reference.

  2. The experimental conditions are kept consistent across methods. All methods use the same frozen features, the same compressor architecture, the same training budget and the same test chunks, and the appendix gives fairly detailed implementation details.

Critical weaknesses
  1. The source of the performance gain is not yet clear. In the main results (Table 1), QMORE uses its accuracy–diversity rank fusion, while the baselines use accuracy-only rank fusion. When the same fusion is applied to both, the R@1 gap on Flickr30k narrows to 0.9 points (14.4 vs. 13.5, Table 4). Although this difference is statistically significant, it is still unclear whether similar results hold on the other datasets and against stronger classical baselines. Therefore, once the effect of the fusion rule is removed, the current evidence is not sufficient to show that the quantum encoder consistently outperforms the shared classical encoder.

  2. The evaluation scope is narrow, and training stability is not sufficiently verified. Each query is ranked against only 200 candidates, and only image-to-text retrieval is evaluated. This protocol differs from the commonly used Flickr30k 1K test setting, which limits direct comparison with prior retrieval results. In addition, the reported standard deviations are computed over test chunks, so they do not directly show whether the results remain stable when the models are retrained with different initializations.

Questions for authors
  1. Could the stronger classical baselines in Table 2 (the 5× encoder-core variants) be evaluated with the same accuracy–diversity rank fusion as QMORE? Could you also add a uniform-weight rank fusion baseline for comparison?

  2. How does performance change when the modality adapters are removed?

  3. If results under the commonly used Flickr30k 1K protocol are available, could you report them to allow comparison with prior work?

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

Related papers

Discussion (0)

Sign in to comment.