Pith. sign in

REVIEW 4 major objections 6 minor 58 references

InfoCIR: Multimedia Analysis for Composed Image Retrieval

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read InfoCIR ties retrieval, explainability, and prompt refinement into one dashboard, and a user study reports top-3 success rising from 37.5% to 87.5%.

desk verdict Real systems contribution; the user study's headline claim is circular and underpowered, but the dashboard, code, and honest limitations make it worth a serious look. read the letter →

arxiv 2602.13402 v1 pith:4OSNVBKL submitted 2026-02-13 cs.HC cs.IRcs.MM

Ioannis Dravilas , Ioannis Kapetangeorgis , Anastasios Latsoudis , Conor McCarthy , Gonçalo Marcelino , Marcel Worring This is my paper · ORCID
classification cs.HCcs.IRcs.MM
keywords composedimageretrievalvisualanalyticspromptenhancementexplainabilitysaliencymapstokenattributionembeddingvisualizationuserstudy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Composed image retrieval (CIR) — searching images by combining a reference photo with a text description of a change — is hard to debug because developers tweak prompts without seeing why the model behaves as it does. InfoCIR is a visual analytics dashboard that joins retrieval results, a 2D embedding view, saliency maps, token attributions, and an LLM-based prompt enhancer, letting users diagnose failures, compare prompt variants, and refine queries. The paper's controlled study with eight participants reports that the full interface raised the chance of getting a target into the top three from 37.5% to 87.5%, cut average time from 277 to 133 seconds, and reduced average unique queries from 7.4 to 3.3. The system is modular, so other CIR models and attribution methods can be plugged in, and the authors position it as the first to jointly support analysis of embedding space, query dynamics, and model behavior.

What carries the argument

The core object is the InfoCIR dashboard, a six-panel visual analytics interface built on a web-based interactive framework and wrapped around the SEARLE zero-shot CIR model, which uses CLIP embeddings. The load-bearing component is the explainable prompt-enhancement loop: after an initial search, the user selects an ideal target image I*; a large language model generates alternative prompts conditioned on the reference image, the original text, and the visual features of I*; these variants are re-ranked and their effect shown in an Ideal-anchored Rank-Δ heatmap, with gradient-based saliency maps and token-attribution bars tracing which words and image regions drive the match. A second key p

What would settle it

Re-run the study with the prompt enhancer blinded to the ideal image's visual features (only reference image and original text, no I* grounding). If success rate stays near 87.5%, the widgets are doing the work; if it drops toward the baseline's 37.5%, the enhancement loop is effectively using privileged access to the target.

Watch

Extended reading notes

Core claim

The claim is that integrating retrieval, embedding-space visualization, and explainable prompt enhancement in one coordinated interface materially improves a user's ability to drive a composed image retrieval model to the intended target. The paper identifies six linked panels—query input, ranked results, histogram/word cloud, a debiased UMAP embedding view, a prompt enhancement panel with an Ideal-anchored Rank-Δ heatmap, and an explanation panel with saliency map and token attribution—as the mechanism. The study reports success rate rising from 37.50% to 87.50%, average time dropping from 277 to 133 seconds, and average unique queries from 7.4 to 3.3. The authors also claim this is the fir

Load-bearing premise

The user study assumes that letting the LLM generate prompts conditioned on the ideal target image, and then measuring success by that same image's rank, is a fair test of the dashboard's diagnostic value rather than an oracle-assisted shortcut.

Editorial extensions

If this is right

  • CIR developers can replace blind trial-and-error with a diagnosis-comparison-enhancement feedback loop that pinpoints which words or image regions drive retrieval outcomes.
  • Prompt refinement becomes a measurable, local optimization: fixing an ideal image as an anchor lets users see exactly how each rephrasing moves a target's rank.
  • The Rank-Δ heatmap and token/saliency explanations provide a verification layer for LLM-generated prompt suggestions, letting users discard ineffective variants based on evidence.
  • The modular architecture supports swapping in new CIR models and attribution methods, turning the dashboard into a reusable comparison and evaluation platform.
  • The reported improvements—higher success, fewer queries, less time—suggest that interactive explainability can change user performance in multimodal retrieval workflows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A likely untested dependency: part of the gain may come simply from LLM-generated prompt suggestions rather than the embedding view, saliency maps, or token bars; an ablation that hides individual panels would reveal which widgets actually carry the benefit.
  • Anchoring enhancement on the ideal image risks overfitting to that specific target; a follow-up could check whether an 'enhanced' prompt generalizes to retrieving other examples of the same class rather than just the selected anchor.
  • The cold-start limitation means the system only pays off once a relevant image appears in the initial results; automatically generating candidate anchors in zero-hit scenarios would extend the tool's usefulness.
  • Because the Rank-Δ heatmap fixes the initial top-k as the reference set, it visualizes only local reordering; a wider view that tracks changes across more positions could tell whether improvements are local or truly global.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. InfoCIR is a web-based visual analytics system for composed image retrieval. It integrates the SEARLE retrieval backend with six linked panels: query composition, top-k results, histogram/word cloud, UMAP embedding view, LLM-based prompt enhancement, and explanation views (saliency maps, token attribution, Rank-Δ heatmap). A three-stage projection pipeline (PCA style debiasing, contrastive debiasing, ICA) is proposed before UMAP. The main evaluation is a within-subjects user study (n=8) comparing full InfoCIR to a baseline with only retrieval panels; the paper reports success rates of 87.5% vs. 37.5%, lower completion time, fewer queries, and qualitative Likert/feedback. The paper claims a controlled user study demonstrating significant improvements and the first joint support of embedding space, query dynamics, and model behavior.

Significance. If the empirical claims could be supported, InfoCIR would be a useful open-source diagnostic environment for CIR developers, combining relevance feedback with explanation and visualization. Strengths include the modular Plotly-Dash implementation, use of a standard CIR backend, explicit discussion of limitations, and a reproducible demo/code repository. The central evaluation, however, is not adequate as submitted: the prompt-enhancement loop is conditioned on the target image used for scoring, and the study lacks statistical support and condition counterbalancing. The paper is better positioned as a system prototype with formative evaluation than as a validated demonstration of significant retrieval improvement.

major comments (4)
  1. [§4.6 and §5.1] The evaluation of prompt enhancement is circular. In §4.6, candidate prompts are generated by R(I_ref, T, I*), i.e., conditioned on the user-selected ideal image, and in §5.1 the success criterion is whether that same I* reaches the top-3 after re-ranking. The LLM is therefore given the target before the rank is scored. The 87.5% vs. 37.5% difference is compatible with the system acting as a relevance-feedback oracle rather than demonstrating that the embedding view, saliency maps, word cloud, or other diagnostic widgets improve retrieval. The paper's own 'Cold Start Vulnerability' and 'Generalization vs. Overfitting' limitations (and P6's failure) show that enhancement only works when a target is already present in the initial results and may overfit to that anchor. The headline 'significantly improves' claim is not supported by this design.
  2. [§5.1.2, Table 1] No significance test or confidence interval is reported; for the success-rate contrast 3/8 vs. 7/8, a two-sided Fisher exact test gives p≈0.12, far from conventional significance. Moreover, condition order is not counterbalanced: every participant completed the baseline scenario first and the InfoCIR scenario second; only the task order (A/B) is reversed across groups. Learning effects and interface acclimation are therefore confounded with condition. Task comparability (porcupine vs. Boston Terrier) is asserted but not validated. The 'Average Time on Task' includes failures capped at 300s although the text defines it as 'mean time to successful retrieval'; medians/IQR and per-task breakdowns would be more appropriate with n=8.
  3. [§4.4, Fig. 3] The claimed improvement from the 'debiased UMAP' / Class Projection Preservation Pipeline is shown only in a conceptual figure. No quantitative separability or projection-quality metric (e.g., silhouette, trustworthiness/continuity, class purity), no ablation of the three stages, and no hyperparameter settings for PCA/ICA/UMAP are given. Since the pipeline is a listed contribution ('supervised metric projection pipeline utilizing a debiased UMAP'), it needs an offline evaluation on a labeled set with the free parameters (n_neighbors, min_dist, metric, component counts, contrastive strength) specified and ideally swept. Without this, the reader cannot assess whether the projection actually reduces style bias or merely rearranges clusters.
  4. [§5, intro and §5.2.2] The study is framed as validating the whole six-panel system via a 'holistic proxy' (prompt enhancement), but the baseline removes all widgets at once. The qualitative results show that participants mainly relied on the Prompt Enhancement Panel and the histogram/word cloud, while the Embedding View and Saliency Map were 'not used quickly' (P5) or 'takes too much time' (P3). This undercuts the claim that the experiment validates the diagnostic widgets collectively. A component-level ablation or separate analysis of which panels contributed to successful refinements is needed to support the 'holistic proxy' argument.
minor comments (6)
  1. [Abstract and §5.1.2] The word 'significantly' is used in the contribution bullet and outcomes, but no inferential statistics are provided. Rephrase to 'observed improvements' or add appropriate tests and confidence intervals.
  2. [Table 1] Use median and IQR for time and query counts; clarify whether the success rate denominator is participants or trials, since each participant contributes one trial per condition.
  3. [§4.2] The ideal-image input is described as optional but 'strongly recommended'. In the evaluation it is always used and is central to the success metric. State more clearly how the system behaves without it and what the evaluation actually covers.
  4. [§2.5] The claim of being 'the first system to jointly support analysis of embedding space, query dynamics, and model behavior' is not supported by a systematic comparison against the cited systems. A feature-comparison table or a more precise scoping of 'jointly' would help.
  5. [§4.4, Fig. 3] Define 'style debiasing' and 'contrastive debiasing' precisely; specify how PCA dimensions and ICA components are selected, and whether class labels are used in the pipeline. If labels are used, the 'unsupervised' nature of the projection should be clarified.
  6. [§4.5, F3] The Rank-Δ heatmap is said to fix the initial top-k as a stable reference set. Clarify how new images that enter top-k after re-ranking are represented, since the row set is not closed under re-ranking.

Circularity Check

1 steps flagged · score 6.0 of 10

Prompt-enhancement evaluation is anchored to the target it scores: R uses I* and success is rank(I*); the 87.5% success rate is largely by construction.

  1. fitted input called prediction [Sec. 3.1 (Eq. T'=R(I_ref,T,I*)), Sec. 4.6 (Explainable Prompt Enhancement), Sec. 5.1 (Quantitative Findings)]
    "T ′ =R(I ref ,T,I ∗ ) ... the user selects one or more ideal target images I ∗ . These chosen images serve as the anchors for enhancement: the system leverages ... LLM to generate a set of alternative prompts T ′ grounded on the content of I ref , T, and the visual features of I ∗ . For each proposed variant T ′ , the system re-ranks the results and visualizes the rank change of I ∗ via an Ideal-anchored Rank-∆ heatmap. ... Each task required participants to refine a text prompt until the target appeared in the top-3 retrieved images."

    R is parameterized by I* and the success metric is the rank of that same I*. The LLM is asked to write prompts grounded on the visual features of the target, then that target's rank improvement is reported as success. This is relevance feedback with oracle access: the outcome measures how well the model exploited the injected I*, not the diagnostic value of the dashboard. The baseline has no such anchor, so the comparison conflates oracle access with InfoCIR's UI. The paper's own 'Cold Start Vulnerability' and 'Generalization vs. Overfitting' limitations admit the loop requires I* in the initial top-k and may overfit to it (P6 failed precisely because no target appeared in the top-10), which is exactly the condition under which the 'success' is defined.

full rationale

The central load-bearing claim — 'A controlled user study demonstrating that InfoCIR significantly improves retrieval efficiency and success rates' — is evaluated with a loop that feeds the target into the enhancer and then scores that same target. The paper's formal definition T' = R(I_ref, T, I*) makes the conditioning explicit, and the quantitative study operationalizes success as the target reaching top-3. Thus the reported 87.5% success rate largely measures the LLM's exploitation of the I* anchor, not the embedding view, saliency maps, histogram, word cloud, or token attribution, which the study claims to validate. The paper itself concedes the dependency via the Cold Start and Overfitting limitations. No other significant circularity was found: the backend (SEARLE) is external, the projection pipeline is heuristic, and self-citations (Worring et al.) are not load-bearing for the central claims. Because the evaluation's central 'prediction' reduces by construction to the engineered input I*, a score of 6 is appropriate; the system itself may have independent content, but the headline empirical support is partial.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical, mathematical, or ontological entities are introduced. InfoCIR is a software artifact; the debiased UMAP pipeline and Rank-Δ heatmap are computational/visual designs, not invented postulates with independent falsifiable handles.

free parameters (2)
  • UMAP hyperparameters (n_neighbors, min_dist, metric)
    Not reported; chosen by hand for the Embedding View. These parameters directly control cluster geometry and the perceived 'style bias' separation.
  • PCA/ICA component counts and contrastive-debiasing strength
    The three stages of the Class Projection Preservation Pipeline (Fig. 3) are heuristic and their hyperparameters are unspecified. They shape the claimed class-separability improvement.
assumptions (5)
  • domain assumption CLIP cosine similarity in the joint embedding space is a faithful measure of CIR relevance.
    Retrieval ranking (Sec. 3.1) and all similarity-based explanations rest on this. The paper's own style-bias discussion acknowledges its limits.
  • domain assumption Gradient of the similarity score with respect to tokens/pixels is a valid local attribution.
    Token attribution bars and saliency maps use gradients. The Discussion concedes gradient approximations may not capture transformer attention mechanisms.
  • domain assumption UMAP projection preserves enough semantic neighborhood structure for the claimed diagnoses.
    The Embedding View is central to DG1, and the paper itself cautions against metric reading of the projection while still relying on it for cluster verification.
  • ad hoc to paper LLM-generated prompt variants conditioned on the ideal image are a fair test of retrieval improvement.
    Section 4.6 grounds prompts in I*, and Section 5.1 scores the rank of I*. This may be a self-fulfilling loop rather than evidence of general system value.
  • ad hoc to paper The two tasks are of comparable difficulty and counterbalancing removes order effects.
    Each participant gives one baseline and one InfoCIR data point; task-specific differences or learning transfer could explain part of the observed improvement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InfoCIR: Multimedia Analysis for Composed Image Retrieval." pith.science (2026). https://pith.science/paper/4OSNVBKL

@misc{pith2026260213402,
  author       = {Pith},
  title        = {Pith review of: InfoCIR: Multimedia Analysis for Composed Image Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4OSNVBKL}},
  note         = {Machine review of arXiv:2602.13402}
}
read the original abstract

Composed Image Retrieval (CIR) allows users to search for images by combining a reference image with a text prompt that describes desired modifications. While vision-language models like CLIP have popularized this task by embedding multiple modalities into a joint space, developers still lack tools that reveal how these multimodal prompts interact with embedding spaces and why small wording changes can dramatically alter the results. We present InfoCIR, a visual analytics system that closes this gap by coupling retrieval, explainability, and prompt engineering in a single, interactive dashboard. InfoCIR integrates a state-of-the-art CIR back-end (SEARLE arXiv:2303.15247) with a six-panel interface that (i) lets users compose image + text queries, (ii) projects the top-k results into a low-dimensional space using Uniform Manifold Approximation and Projection (UMAP) for spatial reasoning, (iii) overlays similarity-based saliency maps and gradient-derived token-attribution bars for local explanation, and (iv) employs an LLM-powered prompt enhancer that generates counterfactual variants and visualizes how these changes affect the ranking of user-selected target images. A modular architecture built on Plotly-Dash allows new models, datasets, and attribution methods to be plugged in with minimal effort. We argue that InfoCIR helps diagnose retrieval failures, guides prompt enhancement, and accelerates insight generation during model development. All source code allowing for a reproducible demo is available at https://github.com/giannhskp/InfoCIR.

Figures

Figures reproduced from arXiv: 2602.13402 by the authors.

Figure 1
Figure 1. The user interface of our system consists of six main panels. (A) The Composed Image Retrieval Panel allows users to [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The architecture links prompt composition, embedding-based retrieval, and interactive visualization across UI panels (A-F), enabling [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Conceptual overview of the Class Projection Preservation [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Explanation pipeline for visual and textual attribution in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Reference and target image pairs used in the quantitative [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 4 canonical work pages

  1. [1]

    Agnolucci, A

    L. Agnolucci, A. Baldrati, M. Bertini, and A. D. Bimbo. isearle: Improving textual inversion for zero-shot composed image retrieval,

  2. [2]

    A. L. Alfeo, A. G. Zippo, V . Catrambone, M. G. Cimino, N. Toschi, and G. Valenza. From local counterfactuals to global feature im- portance: efficient, robust, and model-agnostic explanations for brain connectivity networks.Computer Methods and Programs in Biomedicine, 236:107550, 2023. doi: 10.1016/j.cmpb.2023.107550 9

    arXiv 2023

  3. [3]

    Amid and M

    E. Amid and M. K. Warmuth. Trimap: Large-scale dimensionality reduction using triplets, 2022. 3, 8

    2022

  4. [4]

    Baldrati, L

    A. Baldrati, L. Agnolucci, M. Bertini, and A. D. Bimbo. Zero-shot composed image retrieval with textual inversion, 2023. 2

    2023

  5. [5]

    Chandna and P

    B. Chandna and P. Sen. A counterfactual explanation framework for retrieval models.CoRR, abs/2409.00860, 2024. 2, 3

    arXiv 2024

  6. [6]

    Chandna and P

    B. Chandna and P. Sen. A counterfactual explanation framework for retrieval models, 2025. 8

    2025

  7. [7]

    Z. Chen, F. Silvestri, J. Wang, Y . Zhang, Z. Huang, H. Ahn, and G. Tolomei. Grease: Generate factual and counterfactual explanations for gnn-based recommendations, 2022. 3

    2022

  8. [8]

    chin Ivan Chang

    Y . chin Ivan Chang. A survey: Potential dimensionality reduction methods, 2025. 5

    2025

Show all 58 references
  1. [9]

    J. Choi. Efficient prompt optimization for relevance evaluation via llm-based confusion matrix feedback.Applied Sciences, 15(9), 2025. doi: 10.3390/app15095198 9

    2025 doi

  2. [10]

    B. Dong, R. Collins, and A. Hoogs. Explainability for content-based image retrieval. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition Workshops, pp. 95–98, 2019. 3

    2019

  3. [11]

    X. Dong, S. Wang, D. Lin, G. K. Rajbahadur, B. Zhou, S. Liu, and A. E. Hassan. Promptexp: Multi granularity prompt explanation of large language models.CoRR, abs/2410.13073, 2024. 3, 6

    2024 arXiv

  4. [12]

    Efthymiadis, B

    N. Efthymiadis, B. Psomas, Z. Laskar, K. Karantzalos, Y . Avrithis, O. Chum, and G. Tolias. Composed image retrieval for training-free domain conversion, 2024. 2, 3, 8

    2024

  5. [13]

    Y . Feng, X. Wang, K. K. Wong, S. Wang, Y . Lu, M. Zhu, B. Wang, and W. Chen. Promptmagician: Interactive prompt engineering for text to image creation.IEEE Transactions on Visualization and Computer Graphics, 30(1):295–305, 2024. doi: 10.1109/TVCG.2023.3327168 2

    2024

  6. [14]

    Fu and D

    J. Fu and D. Zhao. Mpad: A new dimension-reduction method for preserving nearest neighbors in high-dimensional vector search, 2025. 8

    2025

  7. [15]

    Ghosh and M

    T. Ghosh and M. Kirby. Supervised dimensionality reduction and vi- sualization using centroid-encoder, 2020. 5

    2020

  8. [16]

    S. M. V . Hatefi, M. Dreyer, R. Achtibat, T. Wiegand, W. Samek, and S. Lapuschkin. Pruning by explaining revisited: Optimizing attribu- tion methods to prune cnns and transformers, 2024. 9

    2024

  9. [17]

    Huang, D

    Z. Huang, D. Witschard, K. Kucher, and A. Kerren. Va + embeddings star: A state-of-the-art report on the use of embeddings in visual an- alytics.Computer Graphics Forum, 42:539–571, 06 2023. doi: 10. 1111/cgf.14859 3

    2023

  10. [18]

    I. T. Jolliffe.Principal Component Analysis. Springer Series in Statis- tics. Springer-Verlag, New York, 2002. doi: 10.1007/b98835 3

    2002 doi

  11. [19]

    Kobak and G

    D. Kobak and G. Linderman. Initialization is critical for preserving global data structure in both t-sne and umap.Nature Biotechnology, 39:1–2, 02 2021. doi: 10.1038/s41587-020-00809-z 5

    2021 doi

  12. [20]

    Q. R. Lauro, S. Shankar, S. Zeighami, and A. Parameswaran. Rag without the lag: Interactive debugging for retrieval-augmented gener- ation pipelines, 2025. 8

    2025

  13. [21]

    Q. Li, K. S. Njotoprawiro, H. Haleem, Q. Chen, C. Yi, and X. Ma. Embeddingvis: A visual analytics approach to comparative network embedding inspection, 2018. 3

    2018

  14. [22]

    Z. Liu, H. Zhang, C. Xiong, Z. Liu, Y . Gu, and X. Li. Dimension re- duction for efficient dense retrieval via conditional autoencoder, 2022. 3

    2022

  15. [23]

    F. Luus, N. Khan, and I. Akhalwaya. Interactive supervision with t-sne. InProceedings of the 10th International Conference on Knowl- edge Capture, K-CAP ’19, p. 85–92. Association for Computing Ma- chinery, New York, NY , USA, 2019. doi: 10.1145/3360901.3364414 3

    2019

  16. [24]

    McInnes, J

    L. McInnes, J. Healy, and J. Melville. Umap: Uniform manifold ap- proximation and projection for dimension reduction, 2020. 2, 3, 5

    2020

  17. [25]

    R. Nara, Y . C. Lin, Y . Nozawa, Y . Ng, G. Itoh, O. Torii, and Y . Mat- sui. Revisiting relevance feedback for clip based interactive image retrieval.CoRR, abs/2404.16398, 2024. 2

    2024 arXiv

  18. [26]

    W. A. Pike, J. Stasko, R. Chang, and T. A. O’Connell. The science of interaction.Information Visualization, 8(4):263–274, Dec. 2009. doi: 10.1057/ivs.2009.22 4

    2009 doi

  19. [27]

    B. A. Plummer, M. I. Vasileva, V . Petsiuk, K. Saenko, and D. Forsyth. Why do these match? explaining the behavior of image similarity models.arXiv, 2019. 3

    2019

  20. [28]

    Psomas, I

    B. Psomas, I. Kakogeorgiou, N. Efthymiadis, G. Tolias, O. Chum, Y . Avrithis, and K. Karantzalos. Composed image retrieval for remote sensing, 2024. 2

    2024

  21. [29]

    P ´oczos

    B. P ´oczos. Independent component analysis.https: //www.cs.cmu.edu/˜bapoczos/Classes/ML10715_2015Fall/ slides/ICA.pdf, 2015. Lecture slides, CMU 10-715: Advanced Introduction to Machine Learning. 5

    2015

  22. [30]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervi- sion, 2021. 2

    2021

  23. [31]

    Rorseth, P

    J. Rorseth, P. Godfrey, L. Golab, M. Kargar, D. Srivastava, and J. Szlichta. Credence: Counterfactual explanations for document rank- ing. InProceedings of the 39th International Conference on Data Engineering, 2023. 2, 6

    2023

  24. [32]

    Sainburg, L

    T. Sainburg, L. McInnes, and T. Gentner. Parametric umap embed- dings for representation and semisupervised learning.Neural Compu- tation, 33:2881–2907, 10 2021. doi: 10.1162/neco a 01434 5

    2021 doi

  25. [33]

    Sedlmair, M

    M. Sedlmair, M. Meyer, and T. Munzner. Design study methodology: Reflections from the trenches and the stacks.IEEE Transactions on Visualization and Computer Graphics, 18(12):2431–2440, 2012. doi: 10.1109/TVCG.2012.213 2

    2012 doi

  26. [34]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization.International Journal of Computer Vi- sion, 128(2):336–359, Oct. 2019. doi: 10.1007/s11263-019-01228-7 2, 3

    2019 doi

  27. [35]

    C. Shi, W. Cui, S. Liu, P. Xu, W. Chen, and H. Qu. Rankexplorer: Vi- sualization of ranking changes in large time series data.IEEE Trans- actions on Visualization and Computer Graphics, 18(12):2669–2678,

  28. [36]

    Smilkov, N

    D. Smilkov, N. Thorat, C. Nicholson, E. Reif, F. B. Vi ´egas, and M. Wattenberg. Embedding projector: Interactive visualization and interpretation of embeddings, 2016. 2, 3

    2016

  29. [37]

    X. Song, H. Lin, H. Wen, B. Hou, M. Xu, and L. Nie. A comprehen- sive survey on composed image retrieval, 2025. 2

    2025

  30. [38]

    Y . Tang, J. Yu, K. Gai, J. Zhuang, G. Xiong, G. Gou, and Q. Wu. Missing target-relevant information prediction with world model for accurate zero-shot composed image retrieval, 2025. 2

    2025

  31. [39]

    J. B. Tenenbaum, V . De Silva, and J. C. Langford. A global geo- metric framework for nonlinear dimensionality reduction.Science, 290(5500):2319–2323, 2000. doi: 10.1126/science.290.5500.2319 3

  32. [40]

    Tovstogan, X

    P. Tovstogan, X. Serra, and D. Bogdanov. Visualization of deep audio embeddings for music exploration and rediscovery. InProceedings of the 19th Sound and Music Computing Conference (SMC), pp. 493–

  33. [41]

    van der Maaten and G

    L. van der Maaten and G. Hinton. Visualizing data using t-sne.Jour- nal of Machine Learning Research, 9(86):2579–2605, 2008. 3

    2008

  34. [42]

    B. Vasu, B. Hu, B. Dong, R. Collins, and A. Hoogs. Explainable interactive content-based image retrieval.Applied AI Letters, 2(4):e41,

  35. [43]

    N. V o, L. Jiang, C. Sun, K. Murphy, L.-J. Li, L. Fei-Fei, and J. Hays. Composing text and image for image retrieval - an empirical odyssey,

  36. [44]

    H. Wang, X. Jiang, A. Nagarajan, X. Guo, L. Ding, D. Wan, J. Zhao, and Y . Chen. Colorslope: A balanced visualization of overview and 10 To appear in the Proceedings of the 19th IEEE Pacific Visualization Conference (PacificVis 2026), Sydney, Australia (Conference Track). deta...

    2026 doi

  37. [45]

    J. Wang, Z. Deng, D. Deng, X. Wang, R. Sheng, Y . Cai, and H. Qu. Empowering multimodal analysis with visualization: A survey.Com- puter Science Review, 57:100748, 2025. doi: 10.1016/j.cosrev.2025. 100748 5

    2025 doi

  38. [46]

    Y . Wang, H. Huang, C. Rudin, and Y . Shaposhnik. Understanding how dimension reduction tools work: An empirical approach to decipher- ing t-sne, umap, trimap, and pacmap for data visualization, 2021. 3, 8

    2021

  39. [47]

    Z. J. Wang, F. Hohman, and D. H. Chau. Wizmap: Scalable interactive visualization for exploring large machine learning embeddings, 2023. 3

    2023

  40. [48]

    Worring, J

    M. Worring, J. Zah ´alka, S. van den Elzen, M. T. Fischer, and D. A. Keim. A multimedia analytics model for the foundation model era,

  41. [49]

    Z. Xu, H. Lamba, Q. Ai, J. Tetreault, and A. Jaimes. Counterfactual editing for search result explanation.CoRR, abs/2301.10389, 2023. 2

    2023 arXiv

  42. [50]

    Z. Yang, D. Xue, S. Qian, W. Dong, and C. Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image re- trieval. InProceedings of the 47th International ACM SIGIR Confer- ence on Research and Development in Information Retrieval, SIGIR ’24, p. 80–90. As...

    2024

  43. [51]

    S. Ye, Z. Sun, G. Wang, L. Guo, Q. Liang, Z. Li, and Y . Liu. Prompt alchemy: Automatic prompt refinement for enhancing code genera- tion, 2025. 9

    2025

  44. [52]

    H. Yin, X. Xu, Z. Wu, J. Zhou, and J. Lu. Sg-nav: Online 3d scene graph prompting for llm-based zero-shot object navigation, 2024. 9

    2024

  45. [53]

    Zah ´alka and M

    J. Zah ´alka and M. Worring. Towards interactive, intelligent, and in- tegrated multimedia analytics. In2014 IEEE Conference on Visual Analytics Science and Technology (VAST), pp. 3–12, 2014. doi: 10. 1109/V AST.2014.7042476 3

    2014

  46. [54]

    C. Zhao, K. Wang, J. H. Hsiao, and A. B. Chan. Grad eclip: Gradient based visual and textual explanations for clip.CoRR, abs/2502.18816,

    arXiv

  47. [55]

    K. Zhu, Q. Zhao, H. Chen, J. Wang, and X. Xie. Promptbench: A unified library for evaluation of large language models, 2024. 3 11

    2024

  48. [500]

    Available from UPF Digital Repository

    Saint- ´Etienne, France, June 2022. Available from UPF Digital Repository. 3

    2022

  49. [2012]

    doi: 10.1109/TVCG.2012.253 2, 3

    2012 doi

  50. [2021]

    doi: 10.1002/ail2.41 3

    doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.