acceptodds
Under review as a conference paper at ICLR 2027

Optimizing Visualization for Time Series Analysis via Vision-embedded Bayesian Optimization

Abstract

Time-series analysis (TSA) is fundamental across many real-world domains, supporting forecasting, anomaly detection, and decision making. Recent vision-based approaches convert time-series data into images and apply models for TSA. However, these methods typically use fixed or manually selected visualization settings, even though different visual presentations may lead to different model performance. Therefore, we introduce Visual Presentation Optimization (VPO), a new formulation that casts time-series visualization design as an optimization problem to improve vision-based TSA performance for a given task and model. To efficiently explore the large combinatorial configuration space, we propose Vision-embedded Bayesian Optimization (VeBO), which incorporates visual embeddings into a deep kernel learning-based Bayesian Optimization surrogate to predict performance from visualization settings and guide the search. Extensive experiments show that optimizing visual presentations without additional training improves performance efficiently across multiple TSA tasks and models, highlighting VPO as an effective, model-agnostic approach for vision-based time-series understanding. Code will be published upon acceptance.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, MLLM, Visualization, Bayesian Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZyVcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VPO shifts from "build better models" to "optimize inputs" for vision-based time-series analysis (TSA). This paper formalizes choosing a better visual presentation as a black-box optimization problem over visualization parameters and proposes using an MLLM's visual embeddings inside a Bayesian Optimization loop (VeBO) to efficiently search the combinatorial space. For closed-source MLLMs, the input-level optimization is a practically meaningful method.

Critical strengths
  • Model-agnostic and applicable to closed models. VPO can work with API-only models where fine-tuning is impossible. This offers a meaningful practical advantage.
  • Insightful sensitivity analysis. The analysis of color and width sensitivity for visualizations provides valuable guidance for understanding model preferences.
Critical weaknesses
  • The method has a high computational cost with 50 BO iterations ×\times full dataset inference per task-model pair. The optimized visualizations show limited evidence of transferring across tasks or datasets.
  • Main results rely on a single dataset, TimeMQA. The cross-dataset results are on a smaller search space without any baselines, and show smaller gains. This raises concerns about generalization and weakens the claim of broad applicability.
  • The compared closed-source and open-source Qwen model baselines are not state-of-the-art , raising the question of whether the latest LVM models have better inherent robustness to visualization variations. If stronger models are more robust to visualization variations, the practical significance of VPO may decrease.
  • The optimization objective is computed on the evaluation dataset, with no reported held-out test set for the main results. This raises concerns about whether the 50-iteration BO search overfits to the evaluation set rather than identifying genuinely better visualizations.
Questions for authors

Please refer to the critical weaknesses.

5 comments

We thank Reviewer ZyVc for the valuable comments and recognizing our work as model-agnostic and applicable to closed models, which is effective method when model fine-tuning is impossible, insightful sensitivity analysis of color and width, which can analyze the performance over visualizations. We have addressed your concerns and questions below.

W1. Computation cost of proposed method

We note that the 50-iteration budget is not specific to VeBO: we applied the same budget to all search-based algorithms in our comparison, from Grid search and Random search to BO-based algorithms such as GoLLuM and standard BO. Thus, the cost of repeated inference on the full validation set is shared by all baselines.

In addition to the runtimes reported in Appendix Table 9 for a search space of 1,536 configurations, we have compared our approach with the efficient fine-tuning (LoRA) below, so that we show the cost-effectiveness of our method compared to the fine-tuning. We have reported the runtimes of the 50-iteration search methods on the full space of 480k candidate configurations. VeBO requires little additional time compared to Random or Grid search, while taking much less time than training-based approaches such as LoRA fine-tuning. This suggests that VeBO offers a favorable trade-off between performance and computational cost. We have clarified these computational considerations and included the corresponding runtime comparisons in the revised manuscript (Appendix C.1).

MethodAD (h)CLS (h)FV mean (h)QA (h)
Grid3.803.552.311.77
Random3.733.502.251.76
BO3.403.852.221.82
GoLLuM3.994.732.572.22
VeBO (ours)3.864.202.252.09
LoRA fine-tuning13.3212.1012.242.96
Exhaustive Search (1.5k space)58.9556.7922.2132.24

We additionally report cross-dataset transfer results for each method using Qwen2.5-VL-7B below. Please refer to our response to W2 for details.

W2. Cross-dataset result and generalizability

As requested, we have extended our cross-dataset evaluation to use the same search space of 480k candidate configurations as in our main experiments. Each method used the same search budget to select a visualization configuration for each task on the TimeMQA validation set using Qwen2.5-VL-7B. We then applied the selected configurations directly to the corresponding tasks on TSAQA, MTBench, and ChatTS without re-optimizing on any of the target datasets.

MethodTSAQA AD Precision ↑AD Recall ↑AD F1 ↑CLS Acc ↑QA T/F ↑QA MCQ ↑MTBench F.C. MAE ↓QA 7 days ↑QA 30 days ↑ChatTS QA Cate. ↑QA Num. ↑
Default0.5870.5470.4840.5380.5430.4724.5580.4850.4010.4090.187
Grid search0.5650.5410.4910.5560.5970.5262.4120.4850.4110.4180.197
Random search0.5530.5360.4920.5610.6050.514103.3530.4620.3900.4150.186
BO0.5250.5190.4830.5680.6010.513102.9150.4830.3990.4230.202
GOLLuM0.5150.5140.5100.5520.5950.5083.0010.4850.3950.4200.194
VeBO0.6160.5750.5370.5690.6030.5172.3200.5050.4220.4290.207

As shown in the table, the configurations selected by VeBO outperformed the default visualizations mostly, those selected by competing methods on most metrics. These results hint that VeBO truly identifies good visualization transferable to unseen datasets with the same tasks. Furthermore, we have examined whether the selected configurations could transfer across models within the same family. We applied the configurations selected using Qwen2.5-VL-7B directly to Qwen3.5-9B without re-optimization.

ModelTransferAD P(↑)AD R(↑)AD F1(↑)Cls P(↑)Cls R(↑)Cls F1(↑)NMSE_v1(↓)NMSE_v2(↓)T/FMCQOverall
Qwen3.5-9B (← Qwen2.5-VL-7B)✗0.5100.5080.4780.1660.2410.1877.5523.640.7350.6410.677
✓0.5330.5330.5310.3570.2740.23010.1519.030.7420.6440.681

As shown, the transferred configurations improved performance on most metrics. This suggests that a visualization configuration selected for one model can also benefit another model in the same family. We have added those results to the revised manuscript (Appendix C.1).

W3. Results on more recent LVLMs

To examine whether more recent VLMs still benefit from visualization optimization, we have newly evaluated two recent open-weight models: Qwen3.5-9B and InterVL3.5-8B. For each model, we have compared the default visualization with the configuration selected by VeBO. Interestingly, we observed that both models benefited from the optimized visualizations across several tasks. In some cases, the gains were even larger for Qwen3.5-9B than for Qwen2.5-VL-7B. Our results suggest that moving to a more recent VLM make input-level visualization optimization more important. We have added this result to the revised manuscript (Appendix C.1).

ModelVeBOAnomaly P↑R↑F1↑Classification P↑R↑F1↑Forecast NMSE v1↓v2↓OpenQA T/F↑MCQ↑Overall↑
Qwen3.5-9B✗0.51020.50790.47760.16540.24070.186613.4422.650.73850.62000.6652
Qwen3.5-9B✓0.56640.56640.56640.31500.32640.28937.6918.090.76150.64610.6902
InternVL3.5-8B✗0.31550.47210.32990.29520.31170.149013.1528.860.74620.63900.6799
InternVL3.5-8B✓0.54660.54500.54110.28080.33750.281620.0924.870.75380.64850.6887

We thank the reviewer for prompting this comparison. We find it particularly interesting that a newer model can sometimes gain more from visualization optimization, and we believe this is worth exploring further.

W4. Clarification on held-out optimization & test set report

We would like to clarify that the term "evaluation dataset" in the original manuscript refers to the validation set, not the test set. Throughout the paper, VPO is optimized on the validation set, while all reported results are evaluated on an unseen test set. We have clarified this distinction in the revised manuscript (Appendix A.3). We have additionally reported the number of validation and test sets below. Following the sample size used in the original TimeMQA benchmark, we sampled 1,400 instances per task for each of the validation and test sets. For FV1 and FV2, we retained only forecasting instances and excluded imputation instances, while QA was restricted to true/false and multiple-choice questions. The table below gives the number of instances used for validation and test in each task. We have clarified this protocol in the revised manuscript (Section 4).

TaskHeld-outTest
Anomaly Detection14001400
Classification14001400
Forecasting (V1)671706
Forecasting (V2)739707
Open QA713681

Therefore, all reported improvements reflect generalization to unseen test data rather than overfitting to the validation set used for optimization.

Dear Reviewer ZyVc,

We sincerely appreciate the time for the review. If our rebuttal has addressed your concerns, we would be grateful if you took our clarifications and additional results into account in your assessment. We would also welcome any further feedback before the discussion closes.

Best regards,

The Authors

Reviewer_EXSLOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a valuable technical contribution under a background that many recent time series forecasting model studies started to look into using the image sequence representation rather than numerical value sequence itself in the time series forecasting tasks. However, there exists a significant gap which desires a theoretical contribution that how can we take the optimal visualization setting for transforming numerical time series into an image representation to better help those visual time series models. This work hence develops the vision embedded Bayesian optimization method for addressing such a problem.

Critical strengths

Strengths of this work include:

  1. The problem targeted is novel and specific - under the context, the problem shifts from the model-centric to data input-centric optimization.

  2. The methodological development is complete and well motivated.

  3. The validation is broad and comprehensive.

Critical weaknesses

There are some minor issues:

  1. The VeBO uses Qwen2,5-VL-3B for visual features while others might use Qwen2.5-3B, which might cause some fairness concerns;

  2. This work only reports single run experimental results, which might make readers difficult in assessing the robustness of the result;

Questions for authors

Q1. I am curious about the impact of VLMs in affecting the results. Here, we just compare QWen and Llama serious, can authors show more to strengthen the confidence of the results?

Q2. If the input data have quality issues, such as data missing or noisy values, will these affect the proposed method?

4 comments

We thank Reviewer EXSL for the valuable comments, and recognizing our work as the problem target is novel and practical, well-motivated and completed methodological development broad and comprehensive validation. We have addressed concerns and comments below.

W1. Impact of feature extractor selection

We originally selected Qwen2.5-3B as an LLM for GoLLuM because it guides BO using textual descriptions of candidate configurations, whereas VeBO uses a VLM (Qwen2.5-VL-3B) to encode the visually rendered plots. We used 3B models for the fairness in the number of parameters. To further examine whether the performance differences between VeBO and GoLLuM are influenced by their respective feature extractors, we have conducted an additional controlled experiment to assess the impact of backbone selection on surrogate modeling and optimization performance. We report the result of GoLLuM using the feature extractor of Qwen2.5-VL-3B below:

Feature extraction modelAD P ↑AD R ↑AD F1 ↑CLS P ↑CLS R ↑CLS F1 ↑FV1 NMSE ↓FV2 NMSE ↓QA T/F ↑QA MCQ ↑QA Overall ↑
Qwen2.5-VL-3B0.5520.5520.5520.2080.2680.21718.8226.940.7040.5910.634
Qwen2.5-3B0.5590.5590.5590.2180.2720.22317.7428.980.7120.5890.636

As shown, GoLLuM with Qwen2.5-3B mostly performs better on anomaly detection, classification, and Forecasting V1, and achieves slightly higher overall QA accuracy. These results suggest that adopting the same vision-language backbone as VeBO does not consistently improve GoLLuM’s optimization performance, suggesting that the original backbone choice does not unfairly disadvantage GoLLuM in our comparison.

W2. Multi-seed evaluation of default & optimized visualization

As requested, we have evaluated VeBO, Grid search, Random search, BO, and GoLLuM using Qwen2.5-VL-7B and report means and standard deviations over three seeds. Within each seed, all methods share ten initialization points, followed by 50 iterations. Grid search uses a fixed uniform sampling scheme and is therefore evaluated once. Evaluation temperature is fixed at 0 on both held-out and test sets.

MethodAD F1 ↑CLS F1 ↑Forecasting V1 NMSE ↓Forecasting V2 NMSE ↓OpenQA Overall ↑
Grid46.0124.3417.5029.6463.73
Random53.41 ± 1.2222.62 ± 0.8416.84 ± 1.9030.69 ± 1.4764.02 ± 0.21
BO54.62 ± 0.8324.40 ± 0.6516.57 ± 0.3432.27 ± 2.0764.32 ± 2.03
GoLLuM55.16 ± 0.7224.45 ± 1.8017.48 ± 0.3128.45 ± 4.1965.00 ± 1.33
VeBO (Ours)55.20 ± 1.0825.31 ± 2.1715.63 ± 1.6630.35 ± 1.6165.15 ± 1.22

VeBO achieves the best mean test performance overall. It outperforms Random search and Grid search, supporting VeBO’s benefits over simpler search strategies. VeBO’s leading results across three task categories demonstrate its effectiveness across diverse tasks. Together with our ablation results, which show an advantage over GoLLuM, these findings provide complementary support for the vision-embedded surrogate design.

Q1. Impact of different VLMs

As requested, we have extended our evaluation to the more diverse LVLMs including Qwen-3.5-9B, InternVL3.5-8B. As shown, the same tendency is maintained along the recent and diverse LVLMs, which implies that the proposed method exhibits strong generalizability and robustness across different model architectures. These results provide further empirical evidence that VeBO's effectiveness is not limited to the Qwen and Llama model families, supporting its generalizability across diverse VLM architectures. We have added this result on revised manuscript (Appendix C.1).

ModelVeBOAD P ↑R↑F1↑CLS P↑R↑F1↑FC NMSE v1↓v2↓OpenQA T/F↑MCQ↑Overall↑
Qwen3.5-9B✗0.51020.50790.47760.16540.24070.186613.4422.650.73850.62000.6652
Qwen3.5-9B✓0.56640.56640.56640.31500.32640.28937.6918.090.76150.64610.6902
InternVL3.5-8B✗0.31550.47210.32990.29520.31170.149013.1528.860.74620.63900.6799
InternVL3.5-8B✓0.54660.54500.54110.28080.33750.281620.0924.870.75380.64850.6887

Q2. Robustness to input data quality

To examine VeBO’s robustness to input quality issues, we have conducted data perturbation experiments by missing 10% data and adding 10% of noise into the original time-series data. We then compared the performance of the default and optimized plots under these noisy inputs.

TaskPlotCleanMissing 10%CleanNoise 10%
AD (F1)Default33.80 ± 0.8333.51 ± 0.9433.80 ± 0.8336.72 ± 2.44
AD (F1)Optimized54.70 ± 1.2554.11 ± 0.5854.70 ± 1.2555.91 ± 0.54
CLS (F1)Default18.84 ± 0.0018.84 ± 0.0118.84 ± 0.0018.84 ± 0.004
CLS (F1)Optimized21.63 ± 0.0920.45 ± 0.8721.63 ± 0.0919.14 ± 0.30
OpenQA (Overall)Default64.32 ± 0.9663.63 ± 0.1764.32 ± 0.9658.84 ± 0.61
OpenQA (Overall)Optimized64.66 ± 1.0463.44 ± 1.4264.66 ± 1.0460.01 ± 1.96

As shown in the table, perturbing the input time series affected performance of visualization in a task-dependent way. The optimized anomaly-detection plot maintained its performance under Gaussian noise, whereas the optimized classification plot and both OpenQA plots degraded. Missing values also reduced the optimized anomaly-detection and classification scores. These results suggest that configurations found by VeBO can remain useful under noisy inputs and generally do not exhibit greater sensitivity compared to the default visualizations. We have added the result in the revised manuscript (Appendix D).

Reviewer_EXSLOpenReview ↗

A Response to Authors

Thank you very much for your great efforts. I have read your responses and I keep my supportive evaluation on this work.

Dear Reviewer EXSL,

Thank you for taking the time to read our rebuttal and for taking the time to consider our response. We’re glad our clarifications addressed your concerns.

Best regards,

The Authors

Reviewer_a6eoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Visual Presentation Optimization (VPO), which treats how a time series is plotted as parameters to optimize for a frozen vision-language model on a given task. It introduces VeBO, a Bayesian optimization method whose GP surrogate operates on MLLM image embeddings of the rendered plots. Experiments on TimeMQA with four MLLMs show that optimized plots outperform the default line plot, with comparisons some baseline methods.

Critical strengths

S1. The problem framing is novel and practically relevant. Vision-based time-series analysis with MLLMs has mostly fixed the rendering by hand, and the paper shows that rendering choices matter a great deal. Considering rendering as a black-box optimization over a structured configuration space is a sensible and reusable formulation.

S2. The paper is well written and the method is clearly described, with a reasonable set of models, tasks, and ablations, making it easy to follow.

Critical weaknesses

W1. Optimization and reporting appear to use the same data, so the reported gains are inflated by selection. Eq. (1) and Algorithm 1 define the objective on "the evaluation dataset" and return the best observed configuration, and the train/validation/test split of TimeMQA is unclear. The headline comparison in Table 1 therefore contrasts a single evaluation of the default plot with the maximum over 50 evaluated configurations on the same test instances. With stochastic API models and small test sets, the max of 50 noisy evaluations will exceed a single evaluation even if the configurations were equivalent. The small differences between Random/Grid search and VeBO in Table 2 may be consistent with selection on test noise rather than real improvements. The paper needs to optimize on a held-out optimization split and report on a disjoint test split, with repeated runs and variance.

W2. The specific method contribution (VeBO) is not shown to be better than simple baselines. In Table 2, VeBO does not demonstrate a clear advantage. GoLLuM outperforms it on anomaly detection for Qwen, Random search matches it on several columns, and the appendix convergence analysis states VeBO "performs moderately among the baselines". Given that the paper's claimed novelty beyond VPO is the vision-embedded surrogate, the evidence that visual embeddings improve search appears to be weak. Reporting means and standard deviations over repeated runs, and best-so-far curves for all methods in the main text, can help support the claim.

Questions for authors

Q1. Could the authors please describe which TimeMQA instances are used during optimization and which are used for the results in Tables 1 and 2? If they overlap, could the authors please rerun with disjoint optimization/test splits and report the test performance of the configuration selected on the optimization split?

Q2. How many instances N are used per task, and how was the stochasticity of GPT-5-mini/Gemini handled (temperature, repeated queries)?

Q3. What is the standard deviation of the default configuration's score under repeated evaluation?

Q4. For the default-visualization baseline in Table 1, is the model prompted identically (same prompt, same resolution) as for the optimized plots? Several discovered configurations change canvas size, which also changes the MLLM's input token count.

5 comments

We thank Reviewer a6eo for the valuable comments and recognizing our work as novel problem framing and practically relevant black-box formulation, well written with the clearly described method. We have addressed concerns and comments below.

W1-1/Q1/Q2. Clarification on held-out evaluation & test set report

We would like to clarify that the “evaluation dataset” used for Eq. (1) and Algorithm 1 refers to the validation set, not the test set. Throughout the paper, VPO is optimized on each validation set, while all reported results are evaluated on their corresponding unseen test sets, respectively. We selected the best of 50 candidate visualization configurations using validation scores and evaluated it on a disjoint test set for final reporting. Thus, Tables 1 and 2 report the test performance of the validation-selected configuration, not the highest test score of validation among 50 candidates. We have clarified the dataset split terms in the revised manuscript. (Appendix A.3)

We have additionally reported the number of validation and test sets below. Following the sample size used in the original TimeMQA benchmark, we sampled 1,400 instances per task for each of the validation and test sets. For FV1 and FV2, we retained only forecasting instances and excluded imputation instances, while QA was restricted to true/false and multiple-choice questions. The table below gives the number of instances used for validation and test in each task. We have clarified this protocol in the revised manuscript (Section 4).

TaskValidationTest
Anomaly Detection14001400
Classification14001400
Forecasting (V1)671706
Forecasting (V2)739707
Open QA713681

Small difference of VeBO over Random/Grid search. We address this concern in W2. Across multiple seeds, the small performance gap no longer persists, optimization-based methods outperform Random/Grid search while VeBO achieving the best overall performance.

W1-2/Q3. Multi-run results for model inference stochasticity

As requested, we have conducted the 3-run evaluation of models at Table 1 (default vs discovered best config) in the main paper (Refer to the below table). For Gemini, we set the temperature 0.5, for GPT, we used the default version. For open-source models, we added the stochasticity with temperature 0.1.

As shown below, with the 3-run evaluation, optimized visualization mostly outperforms the default visualization across diverse models and tasks, which implies that solely optimizing the visual presentation itself can affect the performance indeed, even at the multi-run evaluation with the stochasticity. We have revised the result in the revision using 3-repeated experiments at Table 1.

ModelVeBOA.D. F1 ↑CLS. F1 ↑F.C. V1 NMSE ↓F.C. V2 NMSE ↓QA Overall ↑
GPT-5-mini✗54.60 ± 2.6026.50 ± 0.506.79 ± 2.889.78 ± 6.0868.90 ± 0.60
GPT-5-mini✓59.20 ± 0.8029.00 ± 3.904.41 ± 0.666.20 ± 4.3269.50 ± 2.40
Gemini-2.5-Flash-Lite✗49.20 ± 0.4024.00 ± 0.405.29 ± 0.636.51 ± 2.0164.30 ± 0.20
Gemini-2.5-Flash-Lite✓53.80 ± 0.4026.30 ± 0.405.06 ± 0.415.40 ± 0.8464.80 ± 0.40
LLaMA3.2-11B-Vision✗39.68 ± 0.308.73 ± 0.2819.07 ± 1.4229.83 ± 6.0153.70 ± 0.22
LLaMA3.2-11B-Vision✓53.47 ± 1.0317.29 ± 0.7816.79 ± 0.9629.00 ± 4.5854.58 ± 0.52
Qwen2.5-VL-7B✗33.90 ± 0.7418.84 ± 0.0019.53 ± 4.6335.09 ± 2.2864.32 ± 0.96
Qwen2.5-VL-7B✓55.37 ± 1.0421.63 ± 0.0919.73 ± 2.8428.91 ± 2.2564.81 ± 0.98

We have additionally reported the precise 3-seed repeated optimization result with Qwen2.5-VL of each method below. Please refer to our response to W2 for details.

W2. Multi-seed evaluation of baselines

Following the reviewer’s suggestion, we have repeated the search with Qwen2.5-VL-7B across three seeds to assess whether VeBO’s advantage over simpler baselines hold beyond a single run (See the below table). We have added the multi-seed results and the experiment details to Appendix C.1 and clarified the comparison.

MethodAD F1 ↑CLS F1 ↑Forecasting V1 NMSE ↓Forecasting V2 NMSE ↓OpenQA Overall ↑
Grid46.0124.3417.5029.6463.73
Random53.41 ± 1.2222.62 ± 0.8416.84 ± 1.9030.69 ± 1.4764.02 ± 0.21
BO54.62 ± 0.8324.40 ± 0.6516.57 ± 0.3432.27 ± 2.0764.32 ± 2.03
GoLLuM55.16 ± 0.7224.45 ± 1.8017.48 ± 0.3128.45 ± 4.1965.00 ± 1.33
VeBO (Ours)55.20 ± 1.0825.31 ± 2.1715.63 ± 1.6630.35 ± 1.6165.15 ± 1.22

Experiment detail (can be found in the revised version): Within each seed, all stochastic methods started from the same 10 randomly sampled configurations and then evaluated 50 additional configurations. For each run, we selected the best configuration on the validation set and evaluated it on the disjoint test set. We report the mean and standard deviation of the resulting test scores. We report original Grid search because it is deterministic.

As shown, VeBO’s mean test performance was better than both the Random search mean and the Grid search result on all five reported metrics. Among all methods, VeBO achieved the highest mean F1 and overall QA accuracy, as well as the lowest mean Forecasting V1 NMSE. Although GoLLuM remains strong, VeBO’s mean test performance across three seeds shows mostly outperforming result. These results support the effectiveness of VeBO’s vision-guided search in identifying better visualization configurations than simple search baselines under the same evaluation budget.

Q4. Canvas size and token budget

We clarify that we used the same text prompts for the default and optimized visualizations in Table 1. The default plot was rendered at 1200 × 800 pixels, and canvas size for optimized visualization was not fixed, since it is one of the variables of VPO: 800 × 600, 1200 × 800 (default), 1280 × 720, and 1600 × 900. As Reviewer a6eo notes, varying the canvas size can change the number of image tokens processed by the model. For Qwen, however, the observed difference was small: the image token count ranged from 308 to 315 (7 token difference), with the highest count approximately 2.3% above the lowest. Taken together, the canvas size is intentionally varied as part of VPO, and the associated change in image token count is marginal under our evaluated setting.

Dear Reviewer a6eo,

We sincerely appreciate the time for the review. If our rebuttal has addressed your concerns, we would be grateful if you took our clarifications and additional results into account in your assessment. We would also welcome any further feedback before the discussion closes.

Best regards,

The Authors

Comments on the paper

3 comments

Dear Area Chair and All Reviewers

We thank for the Area Chair for coordinating the review process of our work, and thank all Reviewers for their meaningful and constructive comments. We appreciate Reviewers for recognizing our work as:

  • Novel problem framing, formulating visualization as black-box optimization, shifting the focus from model-centric to input-centric optimization (Reviewers a6eo, EXSL, ZyVc1)
  • Well motivated and clear methodological development (Reviewers a6eo, EXSL)
  • Meaningful, practical and relevant advantage (Reviewers a6eo, ZyVc)
  • Broad and comprehensive validation with reasonable set of models, tasks, and ablations, and insightful sensitivity analysis (Reviewer a6eo, EXS, ZyVc)

In the rebuttal, we have addressed the reviewer's comments including:

  • We have added multi-run & multi-seed evaluations, involving substantial experimental effort during the rebuttal period, to assess the robustness of VeBO, including 3-run evaluations under model stochasticity and 3-seed optimization experiments with competing baselines (Grid, Random, BO, and GoLLuM), and added at revision (Table 13). (Reviewer a6eO, EXSL)
  • We have extended our cross-dataset transferability experiments at larger search space (480K), incorporating comparisons with competing baselines. VeBO has outperformed other baselines at transferability, and Qwen2.5-VL-7B has shown the transferability within the same model family, Qwen3.5-9B. We have added those results at revision (Table 15). (Reviewer ZyVc)
  • We have clarified the held-out validation and test set evaluation, confirming that visualization optimization is performed exclusively on the validation set. We have also added detailed dataset split statistics at revision (Table 6). (Reviewer a6eO, ZyVc)
  • We added additional experiments on more recent LVLMs including Qwen3.5-9B and InternVL3.5-8B, showing that VPO remains effective across diverse models. Those results are all added to the revision (Table 18). (Reviewer EXSL, ZyVc)

We sincerely thank the Area Chair and all reviewers for their insightful feedback and constructive suggestions, which have helped us strengthen our work.

Dear Area Chair,

We appreciate your time and effort in coordinating the review process. We are writing this as few days remain in the discussion period, and we have not received feedbacks from the reviewers except reviewer EXSL since submitting our rebuttal. We took all our effort to address the concerns in a short time, and to discuss with reviewers during the discussion period. We believe that our responses and experiments have helped address the reviewers’ concerns.

In our rebuttal, we have addressed the reviewers’ concerns through additional experiments and clarifications:

  • Reviewer a6eo: Clarification of validation and test set, multi-run evaluations for discovered visualization, multi-seed optimization baseline comparison results, and effect of canvas size has been conducted, addressing the concerns raised by Reviewer a6eo.
  • Reviewer EXSL: Impact of feature extractor selection at baseline (GoLLuM), multi-seed baseline comparsion result, and result on recent LVLMs are provided, addressing all the concerns raised by Reviewer EXSL.
  • Reviewer ZyVc: Computational cost comparison, cross dataset generalizability result, results on recent LVLMs, and clarification of validation and test set have been provided to address the ZyVc's review.

We understand the demands on the reviewers’ time, and would be grateful if you could encourage them to revisit our responses before the discussion closes. We had hoped for an opportunity to discuss with all Reviewers, and would have appreciated a chance to engage further with the reviewers.

Thank you for your time and consideration.

Best regards,
The Authors

Discussion Summary

We thank all reviewers and the Area Chair for their time and feedback, which have helped us strengthen our work. We thank Reviewer EXSL’s follow-up and continued support, as well as Reviewers a6eo and ZyVc’s time and effort in reviewing our work.

We appreciate the reviewers’ recognition of the following strengths of our work:

  • Novel problem formulation: Framing visualization design as black-box optimization (VPO), and shifting from model-centric to input-centric optimization. (Reviewers a6eo, EXSL)
  • Practical relevance: A model-agnostic approach applicable to closed-source models where fine-tuning is unavailable. (Reviewers a6eo, ZyVc)
  • Clear presentation and methodology: A well-written paper with a clearly described, well-motivated methodology (VeBO). (Reviewers a6eo, EXSL)
  • Broad empirical validation: Comprehensive experiments across models and tasks, including ablations. (Reviewers a6eo, EXSL)
  • Insightful sensitivity analysis: Analysis of color and line-width sensitivity that provides guidance on model preferences. (Reviewer ZyVc)

During the discussion period, we added the following experiments:

  • Multi-run evaluations: Conducted three-run evaluations of default/optimized visualizations to check inference stochasticity, and showed VPO improves performance across most models and tasks. (Reviewer a6eo)
  • Multi-seed search baseline comparisons: Conducted three-seed optimization experiments comparing VeBO with Grid Search, Random Search, BO, and GoLLuM, showing best mean performance at VeBO. (Reviewers a6eo, EXSL)
  • Additional LVLM evaluations: Conducted experiments with recent LVLMs: Qwen3.5-9B and InternVL3.5-8B, and showed VPO also improves performance at recent LVLMs. (Reviewers EXSL, ZyVc)
  • Cross-dataset & model transfer: Conducted evaluations on TSAQA, MTBench, and ChatTS with no re-optimization, and showed VeBO outperforms competing baselines on most metrics. We also showed direct transfer of visualizations works along same model familiy: Qwen2.5-VL-7B to Qwen3.5-9B without re-optimization. (Reviewer ZyVc)
  • Input robustness tests of missing and noise data: Conducted evaluations with 10% missing data and gaussian noise, showing that optimized visualizations generally retained their benefits at input perturbations. (Reviewer EXSL)
  • Runtime comparisons: Reported computational cost comparisons with search baselines and LoRA fine-tuning, showing VeBO’s small search overhead and lower cost than LoRA fine-tuning. (Reviewer ZyVc)
  • Controlled backbone comparison of GoLLuM: Conducted evaluation of GoLLuM using Qwen2.5-VL-3B to check backbone selection affects comparison fairness. (Reviewer EXSL)

We also provided the following clarifications:

  • Held-out evaluation protocol: Clarified that optimization and final evaluation use disjoint validation and test sets, and reported dataset split statistics. (Reviewers a6eo, ZyVc)
  • Canvas size and visual token counts: Clarified the use of identical text prompts and image token differences of different canvas sizes. (Reviewer a6eo)

We believe that these additional experiments and clarifications have helped address the reviewers’ concerns, and we appreciate all Reviewer and Area Chair's time of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.