Paper deep dive
ChartDensity-Bench: Benchmarking MLLMs for Numerical Data Reconstruction under Visual Density
Xinhe Wu, Yadong Jin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/4/2026, 3:03:23 AM
Summary
The paper introduces ChartDensity-Bench, a benchmark designed to evaluate Multimodal Large Language Models (MLLMs) on their ability to reconstruct structured numerical data from scientific charts under varying visual densities. The benchmark systematically composes source charts into compound figures with 1, 3, 6, or 9 subplots to control visual density. Experiments on five MLLMs (Doubao, Gemini, GLM, Kimi, Qwen) reveal that numerical reconstruction quality generally degrades as visual density increases, with significant variation in sensitivity across models. Kimi demonstrated the highest robustness to density, while GLM showed the most significant degradation. The study highlights visual density as a critical factor in MLLM chart understanding and proposes a multi-dimensional evaluation framework covering structural reliability, completeness, parseability, and numerical fidelity.
Entities (15)
Relation Signals (9)
ChartDensity-Bench → evaluates → Multimodal Large Language Models
confidence 95% · ChartDensity-Bench, a benchmark for evaluating MLLMs on structured numerical data reconstruction from compound chart figures under controlled visual density.
Visual Density → causesdegradationin → Numerical Fidelity
confidence 90% · numerical reconstruction generally degrades as visual density increases
ChartDensity-Bench → constructedfrom → ICPR CHART-Infographics
confidence 90% · ChartDensity-Bench is constructed from the ICPR CHART-Infographics dataset
Kimi-K2.6 → exhibitshighestrobustnessto → Visual Density
confidence 90% · Kimi is the most density-robust model among the evaluated systems, exhibiting only a 19.18% relative increase in MAE from one to nine charts.
GLM-4.6V → exhibitshighestsensitivityto → Visual Density
confidence 90% · GLM shows the strongest sensitivity, with an 84.81% relative increase.
ChartDensity-Bench → usesmetrics → Numerical Fidelity
confidence 90% · We further propose a multi-dimensional evaluation framework covering structural reliability, reconstruction completeness, parseability, and numerical fidelity.
ChartDensity-Bench → usesmetrics → Structural Reliability
confidence 90% · We further propose a multi-dimensional evaluation framework covering structural reliability, reconstruction completeness, parseability, and numerical fidelity.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multimodal large language models (MLLMs) offer a promising approach for recovering numerical data from scientific charts, but their ability to reconstruct chart data from visually dense figures remains poorly understood. Existing chart understanding benchmarks primarily evaluate question answering or chart-level reasoning and provide limited support for evaluating structured numerical reconstruction from scientific figures. We introduce \textbf{ChartDensity-Bench}, a benchmark for evaluating MLLMs on structured numerical data reconstruction from compound chart figures under controlled visual density. Built from charts paired with source-level ground-truth data, ChartDensity-Bench systematically varies the number of simultaneously presented charts ($k\in{1,3,6,9}$), enabling controlled evaluation of density-induced degradation. We further propose a multi-dimensional evaluation framework covering structural reliability, reconstruction completeness, parseability, and numerical fidelity. Experiments on five recent MLLMs show that numerical reconstruction generally degrades as visual density increases, while the magnitude of degradation varies substantially across models. Chart-level paired comparisons further show that the same source chart can incur higher reconstruction error when embedded in denser visual contexts. These findings highlight visual density as an important and previously underexplored factor in MLLM chart data reconstruction and provide a systematic benchmark for evaluating model robustness in this setting.
Tags
Links
- Source: https://arxiv.org/abs/2609.38781v1
- Canonical: https://arxiv.org/abs/2609.38781v1
Trouble viewing inline? Open PDF directly →
Full Text
44,104 characters extracted from source content.
Expand or collapse full text
ChartDensity-Bench: Benchmarking MLLMs for Numerical Data Reconstruction under Visual Density Xinhe Wu Affiliation: iSoftStone AI Research Lab (ISSAIR) Affiliation: iSoftStone Affiliation: Beijing, China Email: xhwude@isoftstone.com Yadong Jin †thanks: Corresponding author. Affiliation: iSoftStone AI Research Lab (ISSAIR) Affiliation: iSoftStone Affiliation: Beijing, China Email: ydjink@isoftstone.com Abstract Multimodal large language models (MLLMs) offer a promising approach for recovering numerical data from scientific charts, but their ability to reconstruct chart data from visually dense figures remains poorly understood. Existing chart understanding benchmarks primarily evaluate question answering or chart-level reasoning and provide limited support for evaluating structured numerical reconstruction from scientific figures. We introduce ChartDensity-Bench, a benchmark for evaluating MLLMs on structured numerical data reconstruction from compound chart figures under controlled visual density. Built from charts paired with source-level ground-truth data, ChartDensity-Bench systematically varies the number of simultaneously presented charts (k∈1,3,6,9k∈1,3,6,9), enabling controlled evaluation of density-induced degradation. We further propose a multi-dimensional evaluation framework covering structural reliability, reconstruction completeness, parseability, and numerical fidelity. Experiments on five recent MLLMs show that numerical reconstruction generally degrades as visual density increases, while the magnitude of degradation varies substantially across models. Chart-level paired comparisons further show that the same source chart can incur higher reconstruction error when embedded in denser visual contexts. These findings highlight visual density as an important and previously underexplored factor in MLLM chart data reconstruction and provide a systematic benchmark for evaluating model robustness in this setting. 1 Introduction Scientific literature contains a vast amount of empirical knowledge encoded in figures and charts. As machine learning is increasingly applied to scientific discovery, researchers in AI for Science increasingly rely on published literature as a source of training and evaluation data Wang et al., (2023); Mobarak et al., (2023). However, much numerical information is available only as visualizations, while the underlying data are not released or are provided in heterogeneous formats Yuan et al., (2026); He et al., (2026). Recovering these data at scale therefore remains labor-intensive, requiring interpretation of axes, legends, data series, and graphical elements. Automated reconstruction of structured numerical data from scientific figures could substantially reduce this bottleneck. The emergence of multimodal large language models (MLLMs) provides a promising approach to this problem. Recent MLLMs have demonstrated strong capabilities in document understanding, optical character recognition, and chart reasoning Achiam et al., (2023); Wang et al., (2024); Liu et al., (2023); Liu et al., (2026). In principle, an MLLM can directly process a scientific figure and reconstruct its underlying data. However, scientific figures often contain multiple subplots, dense data series, small visual elements, and heterogeneous graphical structures. Providing an entire compound figure may therefore introduce substantial visual interference. Decomposing a compound figure into individual subplots is a natural potential solution, but the effect of visual density on numerical data reconstruction has not been systematically studied. Existing multimodal benchmarks provide limited support for studying this problem. ChartQA and ChartQA Pro evaluate chart understanding and numerical reasoning, but provide limited coverage of scientific charts and often contain explicitly displayed numerical values that can be recognized through OCR Masry et al., (2022); Masry et al., (2025). Their numerical evaluation also focuses primarily on point-wise error, without separately characterizing chart separation or reconstruction completeness. ExChart-Bench provides a more direct chart-to-data extraction setting, but its ground truth is manually annotated from chart images and its evaluation primarily emphasizes format success and numerical error He et al., (2026). These limitations motivate a benchmark with source-level ground truth and explicit evaluation of structural reliability, completeness, and numerical fidelity. To address this gap, we introduce ChartDensity-Bench, a benchmark for evaluating MLLMs on structured numerical data reconstruction from compound chart figures under controlled visual density. Rather than answering individual questions about a chart, models must identify each constituent chart and reconstruct its associated tabular data, including data-series identifiers, x-axis entries, and numerical values. The benchmark is constructed from charts paired with ground-truth tables, enabling direct evaluation against the underlying source data. A key feature of ChartDensity-Bench is its controlled visual-density setting. We systematically compose k∈1,3,6,9k∈1,3,6,9 source charts into compound figures, allowing the same source charts to be evaluated under different numbers of simultaneously presented subplots. We evaluate reconstruction from complementary perspectives: structural reliability, reconstruction completeness, and numerical fidelity, while additionally reporting parseability to distinguish formatting or structural failures from numerical errors. Across five recent MLLMs, we find that reconstruction quality generally degrades as visual density increases, with substantial differences in density sensitivity across models. Our contributions are summarized as follows: • We introduce ChartDensity-Bench, a benchmark for structured numerical data reconstruction from compound scientific chart figures with source-level ground truth under controlled visual density. • We propose a multi-dimensional evaluation framework covering structural reliability, reconstruction completeness, parseability, and numerical fidelity. • We systematically evaluate five recent MLLMs under controlled compound-figure settings and characterize their differing sensitivity to increasing visual density. 2 ChartDensity-Bench We introduce ChartDensity-Bench, a benchmark for evaluating the ability of multimodal large language models (MLLMs) to reconstruct structured numerical data from compound chart figures under controlled visual density. Unlike conventional chart understanding benchmarks that primarily evaluate visual question answering or semantic interpretation, ChartDensity-Bench requires models to recover the underlying tabular data of every constituent chart. This formulation exposes three distinct failure modes: structural reliability, reconstruction completeness, and numerical fidelity. 2.1 Benchmark Overview Given a compound figure containing k charts, where k∈1,3,6,9k∈\1,3,6,9\, an MLLM is required to identify the constituent charts and reconstruct the data table associated with each chart. Each table contains data-series identifiers, x-axis entries, and their corresponding numerical values. Importantly, the model is not provided with chart-level annotations or source-chart identifiers. The task therefore requires three levels of correspondence: (i) structural correspondence, i.e., correctly separating the constituent charts; (i) point correspondence, i.e., matching extracted series and x-axis entries to the ground truth; and (i) numerical correspondence, i.e., accurately recovering the underlying y-values. We evaluate these capabilities under four controlled visual-density levels, with k=1k=1 serving as the single-chart baseline and larger k corresponding to increasingly dense compound figures. Figure 1: Overview of ChartDensity-Bench. Source charts are systematically composed into increasingly dense compound figures with k∈1,3,6,9k∈\1,3,6,9\ charts. MLLMs must identify each constituent chart and reconstruct its underlying numerical data. The benchmark separately measures structural reliability, reconstruction completeness, and numerical fidelity, enabling controlled analysis of performance degradation under increasing visual density. 2.2 Dataset Construction ChartDensity-Bench is constructed from the ICPR CHART-Infographics dataset Davila et al., (2022), which provides chart images paired with ground-truth tables. We focus on line and bar charts and systematically compose individual charts into compound figures. Controlled visual density. For each density level k∈1,3,6,9k∈\1,3,6,9\, we compose k source charts into a single image. The k=1k=1 condition uses the original chart, while k=3k=3, 66, and 99 arrange charts in 1×31× 3, 2×32× 3, and 3×33× 3 grids, respectively. Adjacent charts are separated by a fixed 10-pixel margin. Each source image is resized to a common width of 800 pixels while preserving its aspect ratio. Thus, the number of simultaneously presented charts provides a controlled density variable while preserving the underlying chart content. Ground-truth manifest. For every compound figure, we retain the ordered list of source-chart identifiers and their corresponding ground-truth. This manifest provides the reference for chart-level and point-level alignment during evaluation. Data partitioning. We use a fixed random seed of 42 for reproducible composition. Source charts are sampled without replacement within each compound, and incomplete groups are discarded. The resulting benchmark contains four controlled density conditions corresponding to 11, 33, 66, and 99 simultaneously presented charts. 2.3 Data Reconstruction and Alignment Each compound figure is independently provided to the target MLLM with a unified extraction prompt. The model is instructed to identify individual charts, distinguish data series, interpret x-axis values, and recover chart-specific data points. For line charts, values are associated with visible line markers or nodes; for bar charts, values are associated with the centers of bar tops. The model is required to produce a standardized table representation: | Series Name | X Value | Y Value |. The raw responses are subsequently parsed into chart-level tables. Explicit chart delimiters are used when available, while structural cues such as x-axis sequences and series structure are used for continuous outputs. Formatting and parsing failures are recorded separately from numerical reconstruction errors. Predicted tables are then matched to the source charts using the ground-truth manifest. Within each chart, series are matched using normalized string similarity with one-to-one assignment. Numerical x-values are matched by nearest-neighbor assignment subject to a predefined tolerance, while categorical x-values are matched by their labels or source ordering. Only successfully matched points are used for numerical fidelity evaluation. 2.4 Evaluation Metrics We evaluate reconstruction quality along four complementary dimensions: structural reliability, parseability, completeness, and numerical fidelity. We report all metrics separately for visual densities k∈1,3,6,9k∈1,3,6,9. Detailed definitions and implementation details are provided in Appendix A.1. Structural reliability. We report Response Format Validity (RFV), which measures whether the predicted number of charts matches the expected number at the compound level, and Chart Alignment Rate (CAR), which measures the fraction of expected charts that are recovered from compounds with structurally valid chart counts. Thus, RFV captures compound-level structural correctness, while CAR measures chart-level recovery. Parseability. We report Parseability Rate (PR), the fraction of aligned charts for which valid numerical evaluation can be performed. A chart is considered parseable when valid point correspondences can be established and numerical errors can be computed. Charts for which no valid matches remain, or whose resulting MAE or RMSE exceeds 100, are treated as unparseable. These cases predominantly arise from unsuccessful structural matching, such as incorrect row ordering or series assignment. Completeness. For each aligned chart i, we define the Completeness Ratio as Ci=N^iNi,C_i= N_iN_i, (1) where NiN_i and N^i N_i denote the numbers of valid ground-truth and extracted x-axis entries, respectively. We report the mean across aligned charts as Average Completeness Ratio (ACR). We additionally report Point Coverage, which measures the fraction of valid ground-truth points that are successfully matched by the extracted reconstruction. To jointly account for chart-level recovery and within-chart completeness, we define Effective Completeness (EC) as EC=CAR×ACR.EC=CAR×ACR. (2) Numerical fidelity. For successfully matched point pairs, we normalize the extracted and ground-truth y-values to [0,100][0,100] using the minimum and maximum ground-truth values among the matched points within the corresponding chart. We then report Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). When normalization is not possible, relative errors are used as a fallback, followed by bounded absolute errors. For multi-series charts, we first attempt to associate extracted and ground-truth series using normalized string matching based on edit distance Levenshtein, (1966). If a reliable series mapping cannot be established, we fall back to sequential row-order matching. Within successfully mapped series, numeric x-values are matched using one-to-one nearest-neighbor assignment Kuhn, (1955), while categorical values are matched by row order. Further details are provided in Appendix A.1. 3 Benchmark Results We evaluate five API-accessed MLLMs—Doubao-Seed-2.0-Pro ByteDance, (2026), Gemini-3.1-Pro Google DeepMind, (2026), GLM-4.6V GLM-V Team et al., (2025), Kimi-K2.6 Moonshot AI, (2025), and Qwen-3.6-Plus Bai et al., (2025)—under four controlled visual-density conditions with 1, 3, 6, and 9 simultaneously presented charts. We ask three questions: (i) how accurately do current MLLMs reconstruct chart data, (i) how does reconstruction quality change as visual density increases, and (i) do different models exhibit distinct sensitivity to dense visual contexts? 3.1 Overall Reconstruction Performance Table S1 summarizes reconstruction performance across density levels. Even in the single-chart setting, where no competing charts are present, substantial differences exist across models. Gemini achieves the strongest overall reconstruction quality, obtaining both the highest point coverage (60.5%) and the lowest normalized MAE (5.51), followed by Qwen, Kimi, Doubao, and GLM in terms of numerical error. This result indicates that chart data reconstruction remains non-trivial even without additional visual interference. More importantly, single-chart performance does not fully predict robustness to dense compound figures. While Gemini provides the best single-chart numerical accuracy, Kimi exhibits a substantially smaller increase in error as the number of simultaneously presented charts grows. This distinction between absolute reconstruction quality and density robustness motivates a direct analysis of performance degradation under increasing visual density. 3.2 Effect of Increasing Visual Density (a) (b) Figure 2: Numerical reconstruction error under increasing chart density. The x-axis denotes the number of simultaneously presented subplots. Lower values indicate better reconstruction fidelity. Table 1: Density-induced degradation in numerical reconstruction from 1 to 9 subplots. Δ denotes the absolute increase in error, while Relative Δ measures the increase relative to the single-chart baseline. Model Δ ↑ Rel. Δ ↑ Δ ↑ Rel. Δ ↑ Doubao 4.4009 51.82% 5.6338 47.30% Gemini 3.0017 54.46% 3.8754 47.57% GLM 7.3937 84.81% 8.5409 67.31% Kimi 1.5191 19.18% 2.2397 19.89% Qwen 5.7183 81.01% 7.4899 73.30% Figure 2 shows normalized MAE and RMSE as the number of simultaneously presented charts increases from 1 to 9. Reconstruction error generally increases with visual density across all five models, although the trajectories are not strictly monotonic. The effect is strongest for GLM and weakest for Kimi, while Gemini and Qwen show intermediate sensitivity. The 1-to-9 degradation in Table 1 further quantifies this model-dependent effect. Relative MAE degradation ranges from 19.18% for Kimi to 84.81% for GLM, with RMSE showing the same overall ordering. These consistent trends across two numerical error measures indicate that density-induced degradation is not specific to a single fidelity metric. 3.3 Completeness–Fidelity Trade-off (a) (b) Figure 3: Relationship between point coverage and numerical reconstruction error under increasing chart density. Better reconstruction lies toward the upper-right region, corresponding to higher point coverage and lower numerical error. Numerical accuracy alone does not capture whether a model reconstructs the complete chart data. Figure 3 therefore jointly visualizes point coverage and numerical error under increasing density. Each trajectory connects the four density conditions for the same model. Overall, denser compound figures tend to move models toward lower point coverage and higher numerical error, indicating that visual density induces both omission and value-reconstruction errors. However, the two failure modes are not perfectly coupled. Kimi provides a robustness-oriented example: its point coverage decreases only moderately from 55.9% to 52.9% between one and nine charts, while its MAE increases from 7.92 to 9.44. In contrast, GLM experiences simultaneous degradation in both dimensions, with point coverage decreasing from 56.2% to 40.9% and MAE increasing from 8.72 to 16.11. Qwen further illustrates that completeness and numerical fidelity can change independently. Its point coverage continues to decrease from three to six charts, while its MAE temporarily improves from 9.93 to 9.00. Thus, a model may recover fewer points while estimating the remaining points more accurately. This observation motivates treating completeness and numerical fidelity as separate evaluation dimensions. 3.4 Model-specific Density Sensitivity The results reveal substantial differences in how models respond to increasing visual density. Kimi is the most density-robust model among the evaluated systems, exhibiting only a 19.18% relative increase in MAE from one to nine charts. GLM shows the strongest sensitivity, with an 84.81% relative increase. Qwen also exhibits pronounced sensitivity, with an 81.01% relative increase in MAE. Structural reliability shows a similarly model-dependent pattern. GLM’s response format validity decreases from 100% in the single-chart condition to 55.3% at nine charts, whereas Kimi remains above 98% throughout. These results suggest that dense visual contexts can affect reconstruction at multiple levels, ranging from structural parsing to point recovery and numerical estimation. Overall, the experiments establish that visual density is a meaningful and model-dependent source of reconstruction degradation. In particular, the substantially different responses of models such as Kimi and GLM raise the question of why some models remain robust while others degrade sharply. We investigate this question through attention-level analysis in Section 4. 4 Chart-level Analysis of Density-induced Error The aggregate results in Section 3 show that numerical reconstruction generally deteriorates as more charts are presented simultaneously. However, aggregate comparisons may confound density with differences in chart composition. We therefore exploit the controlled construction of ChartDensity-Bench and compare the reconstruction of the same source chart under different density conditions. 4.1 Paired Chart-level Density Effects We first compare reconstruction errors for the same source chart under the single-chart and nine-chart conditions. Figure 4 presents paired comparisons of normalized MAE. (a) (b) Figure 4: Paired chart-level comparison between the 1- and 9-chart conditions. Each point corresponds to the same source chart; values above the diagonal indicate higher error under density 9. The right panel shows the distribution of per-chart error changes. A substantial fraction of charts lie above the diagonal in Figure 4a, indicating higher reconstruction error when the same underlying chart is embedded in a nine-chart compound figure. This comparison controls for intrinsic chart difficulty because the source chart is held fixed while its visual context changes. For source chart c, we define the change in reconstruction error as ΔMAEc=MAEc,9−MAEc,1. _c=MAE_c,9-MAE_c,1. (3) As shown in Figure 4b, the distribution of ΔMAE is generally shifted toward positive values across models, although substantial chart-to-chart variation remains. Thus, density does not affect all charts uniformly, but the aggregate effect is predominantly toward higher reconstruction error. Corresponding RMSE results are provided in Figure S1. 4.2 Model-specific Density Sensitivity The magnitude of the chart-level density effect varies substantially across models. To summarize the overall trend, we fit a descriptive linear trend for each model: MAEc,d=αm+βmd+ϵc,d,MAE_c,d= _m+ _md+ _c,d, where d∈1,3,6,9d∈\1,3,6,9\ denotes the number of simultaneously presented subplots. Here, βm _m summarizes the average direction and magnitude of the density–error relationship across the four controlled density levels. Because the same source charts are observed repeatedly, these slopes are treated as descriptive summaries rather than independent causal estimates. Table 2: Per-model sensitivity of chart reconstruction error to visual density. β is the slope from MAE=α+β⋅Density+ϵMAE=α+β· Density+ε. Positive slopes indicate increasing numerical error as more charts are presented simultaneously. Model β Std. Error 95% CI p R2R^2 Doubao 0.582 0.117 [0.354, 0.811] <.001<.001 .013 Gemini 0.296 0.099 [0.103, 0.490] .003 .004 GLM 0.940 0.138 [0.671, 1.210] <.001<.001 .027 Kimi 0.181 0.096 [-0.008, 0.369] .060 .002 Qwen 0.627 0.109 [0.413, 0.841] <.001<.001 .018 As shown in Table 2, all five models have positive estimated density slopes, although the strength of statistical evidence varies. GLM exhibits the largest slope (β=0.940β=0.940), followed by Qwen (0.6270.627), Doubao (0.5820.582), Gemini (0.2960.296), and Kimi (0.1810.181). The estimated slope for Kimi is the smallest and does not reach the conventional p<.05p<.05 threshold (p=.060p=.060), indicating weaker evidence for a systematic linear trend across the four density levels. The low R2R^2 values reflect substantial heterogeneity in the intrinsic difficulty of individual charts. They should therefore not be interpreted as evidence against a density effect. To account explicitly for repeated observations of the same source charts and to test model-specific density effects, we next perform a mixed-effects analysis. 4.3 Mixed-effects Analysis of Density Sensitivity Because each source chart is evaluated under multiple density conditions and by multiple models, we use a linear mixed-effects model Bates et al., (2015) with chart identity as a random intercept: Yc,m,d=β0+βmodel,m+βdensity,d+βmodel×density,m,d+γc+ϵc,m,d,Y_c,m,d= _0+ _model,m+ _density,d+ _model×density,m,d+ _c+ _c,m,d, where Yc,m,dY_c,m,d denotes the normalized reconstruction error for source chart c, model m, and density condition d. Kimi and the single-chart condition are used as reference categories. The results in Table S2 provide evidence for substantial model-dependent density sensitivity. Relative to Kimi, Doubao, GLM, and Qwen exhibit significantly larger density effects at nine subplots, with interaction coefficients of 3.757 (p<.001p<.001), 6.077 (p<.001p<.001), and 4.803 (p<.001p<.001), respectively. The corresponding interaction for Gemini is smaller and does not reach conventional significance (1.616, p=.080p=.080). The main density effect for Kimi at nine subplots is 1.552 MAE points (p=.014p=.014). Combining this reference effect with the corresponding model–density interactions gives estimated density-related increases of approximately 5.31 MAE points for Doubao, 3.17 for Gemini, 7.63 for GLM, 1.55 for Kimi, and 6.36 for Qwen at nine subplots relative to the single-chart reference, with the latter quantities interpreted through the fitted model coefficients. 4.4 Evidence Consistent with Attention Dilution The paired chart-level analysis and mixed-effects results show that increasing visual density is associated with higher reconstruction error even when the underlying source chart is held fixed, with substantially stronger effects for models such as GLM and Qwen than for Kimi. This pattern is consistent with an attention dilution hypothesis Chen et al., (2024); He et al., (2024); Leng et al., (2025), in which additional charts introduce competing visual elements and reduce the effective focus available to each target chart. However, ChartDensity-Bench measures behavioral reconstruction rather than internal attention allocation and therefore cannot establish this mechanism directly. We interpret the results as behavioral evidence compatible with attention dilution, while direct analysis of multimodal attention or internal representations remains an important direction for future work. 5 Discussion Our results have several implications for evaluating MLLMs for numerical chart data reconstruction. Visual density as an evaluation factor. A central finding of ChartDensity-Bench is that visual density itself can substantially affect reconstruction quality. Numerical error generally increases as more charts are presented simultaneously, and the paired chart-level analysis shows that the same source chart can become harder to reconstruct when embedded in a denser compound figure. This suggests that chart reconstruction difficulty depends not only on the intrinsic complexity of the target chart but also on its surrounding visual context. Since scientific figures frequently contain multiple panels, robustness to visual density is therefore an important consideration for chart-to-data extraction systems. Multi-dimensional evaluation and model robustness. The results also show that completeness and numerical fidelity capture distinct failure modes. A model may omit data points while estimating the remaining values relatively accurately, or experience simultaneous degradation in both dimensions. Structural reliability, completeness, parseability, and numerical fidelity therefore provide complementary views of reconstruction behavior. Moreover, strong single-chart performance does not necessarily imply robustness to dense visual contexts: the evaluated models exhibit substantially different degradation patterns, with Kimi showing relatively small density-induced degradation and GLM substantially larger sensitivity. Limitations and future directions. Our benchmark currently focuses on line and bar charts and therefore does not cover the broader diversity of scientific visualizations. In addition, visual density is operationalized by the number of simultaneously presented charts, which provides a controlled variable but does not capture all forms of visual complexity, such as the number of data series, text density, or visual overlap. Finally, the benchmark measures behavioral outcomes and cannot by itself establish the internal mechanism underlying density-induced degradation. Future work could extend the benchmark to additional visualization types and finer-grained sources of visual interference, while directly investigating model representations or inference strategies that improve robustness to dense compound figures. Reproducibility statement The benchmark construction procedure, controlled density settings, data reconstruction protocol, alignment procedure, and evaluation metrics are described in Sections 2 and 3 and further detailed in the Appendix. We plan to release the benchmark construction and evaluation code upon publication. The benchmark construction and evaluation code will be publicly available at https://github.com/ISSAIR/ChartDensity-Bench. References Achiam et al., (2023) Achiam, J., Adler, S., Agarwal, S., et al. (2023). GPT-4 technical report. arXiv preprint arXiv:2303.08774. Bai et al., (2025) Bai, S., Cai, Y., Chen, R., et al. (2025). Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Bates et al., (2015) Bates, D., Mächler, M., Bolker, B., and Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1):1–48. ByteDance, (2026) ByteDance (2026). Doubao-seed-2.0. https://team.doubao.com/seed2. Chen et al., (2024) Chen, L., Zhao, H., Liu, T., Bai, S., Lin, J., Zhou, C., and Chang, B. (2024). An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In Proceedings of the European Conference on Computer Vision (ECCV). Davila et al., (2022) Davila, K., Xu, F., Ahmed, S., Mendoza, D. A., Setlur, S., and Govindaraju, V. (2022). ICPR 2022: Challenge on harvesting raw tables from infographics (CHART-Infographics). In 2022 26th International Conference on Pattern Recognition (ICPR), pages 4995–5001. IEEE. GLM-V Team et al., (2025) GLM-V Team, Hong, W., et al. (2025). GLM-4.5V and GLM-4.1V-Thinking: Towards versatile multimodal reasoning with scalable reinforcement learning. arXiv preprint arXiv:2507.01006. Google DeepMind, (2026) Google DeepMind (2026). Gemini 3.1 pro model card. https://deepmind.google/models/model-cards/gemini-3-1-pro/. He et al., (2026) He, Y., Ying, P., Cheng, L., Peng, K., Tian, Y., Deng, D., and Wu, Y. (2026). Making multimodal llms reliable chart data extractors: A benchmark and training framework. arXiv preprint arXiv:2606.29808. He et al., (2024) He, Z. et al. (2024). LOOK-M: Look-once optimization in KV cache for efficient multimodal long-context inference. arXiv preprint arXiv:2406.18139. Kuhn, (1955) Kuhn, H. W. (1955). The Hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2(1–2):83–97. Leng et al., (2025) Leng, X. et al. (2025). RAVE: Re-allocating visual attention in large multimodal models. arXiv preprint arXiv:2605.18359. Levenshtein, (1966) Levenshtein, V. I. (1966). Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady, 10(8):707–710. Liu et al., (2023) Liu, Y., Li, Z., Yang, B., et al. (2023). OCRBench: On the hidden mystery of OCR in large multimodal models. arXiv preprint arXiv:2305.07895. Liu et al., (2026) Liu, Z., Gao, X., Niu, F., Gao, Q., Liu, L., and Piramuthu, R. (2026). Start: Spatial and textual learning for chart understanding. In 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8146–8156. IEEE. Masry et al., (2025) Masry, A., Islam, M. S., Ahmed, M., Bajaj, A., Kabir, F., Kartha, A., Laskar, M. T. R., Rahman, M., Rahman, S., Shahmohammadi, M., Thakkar, M., Parvez, M. R., Hoque, E., and Joty, S. (2025). ChartQAPro: A more diverse and challenging benchmark for chart question answering. In Findings of the Association for Computational Linguistics: ACL 2025, pages 19123–19151. Association for Computational Linguistics. Masry et al., (2022) Masry, A., Long, D. X., Tan, J. Q., Joty, S., and Hoque, E. (2022). ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263–2279. Association for Computational Linguistics. Mobarak et al., (2023) Mobarak, M. H., Mimona, M. A., Islam, M. A., Hossain, N., Zohura, F. T., Imtiaz, I., and Rimon, M. I. H. (2023). Scope of machine learning in materials research—a review. Applied Surface Science Advances, 18:100523. Moonshot AI, (2025) Moonshot AI (2025). Kimi k2: A trillion-parameter MoE language model. https://github.com/MoonshotAI/Kimi-K2. Wang et al., (2023) Wang, H., Fu, T., Du, Y., Gao, W., Huang, K., Liu, Z., Chandak, P., Liu, S., Van Katwyk, P., Deac, A., Anandkumar, A., Bergen, K. J., Gomes, C. P., Ho, S., Kohli, P., Lasenby, J., Leskovec, J., Liu, T.-Y., Manrai, A. K., Marks, D. S., Ramsundar, B., Song, L., Sun, J., Tang, J., Veličković, P., Welling, M., Zhang, L., Coley, C. W., Bengio, Y., and Zitnik, M. (2023). Scientific discovery in the age of artificial intelligence. Nature, 620:47–60. Wang et al., (2024) Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al. (2024). Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191. Yuan et al., (2026) Yuan, Y. et al. (2026). A deep learning framework for scientific chart data extraction and reconstruction. Communications Engineering. Appendix A Appendix A.1 Detailed Evaluation Metrics This section provides the formal definitions and implementation details of the evaluation metrics used in ChartDensity-Bench. Reconstruction quality is evaluated at the compound, chart, and point levels, covering structural reliability, parseability, completeness, and numerical fidelity. Response Format Validity (RFV). Let NprocessedN_processed denote the number of processed compound figures, and let NmismatchN_mismatch denote the number of compounds for which the number of extracted charts differs from the expected number. We define Response Format Validity as RFV=1−NmismatchNprocessed.RFV=1- N_mismatchN_processed. (4) A compound is considered structurally valid only when the number of extracted charts exactly matches the number of source charts. If the chart counts differ, the compound is recorded as a structural mismatch and its charts are excluded from subsequent chart-level evaluation. Chart Alignment Rate (CAR). Let NalignedN_aligned denote the number of charts evaluated from structurally valid compounds, and let NexpectedN_expected denote the total number of expected charts across all processed compounds. The Chart Alignment Rate is CAR=NalignedNexpected.CAR= N_alignedN_expected. (5) For a visual density k, where each compound contains k source charts, Nexpected=Nprocessed×k.N_expected=N_processed× k. (6) Because chart-level evaluation is performed only for compounds with the correct chart count, CAR jointly reflects successful compound-level structural recovery and subsequent chart alignment. Parseability Rate (PR). Let NparseableN_parseable denote the number of aligned charts for which valid numerical evaluation can be performed. We define PR=NparseableNaligned.PR= N_parseableN_aligned. (7) A chart is considered parseable when at least one valid point correspondence can be established and a numerical error value can be computed. If no valid matches or error values remain, the chart is considered unparseable. In addition, charts with aggregate MAE or RMSE greater than 100 are treated as unparseable. The latter criterion serves as a numerical sanity check for cases in which structural matching produces degenerate errors. Completeness Ratio. For aligned chart i, let NiN_i denote the number of valid ground-truth x-axis entries and let N^i N_i denote the number of extracted entries. The chart-level Completeness Ratio is Ci=N^iNi.C_i= N_iN_i. (8) For numeric x-values, only valid numeric entries are counted. For non-numeric x-values, non-empty entries are counted. Average Completeness Ratio (ACR). Let A denote the set of aligned charts. The Average Completeness Ratio is computed as a macro-average over aligned charts: ACR=1||∑i∈Ci.ACR= 1|A| _i C_i. (9) Thus, each aligned chart contributes equally to ACR regardless of the number of points or series it contains. Point Coverage. For aligned chart i, let Li,jL_i,j denote the number of successfully matched points in series j, and let Ni,jpointsN^points_i,j denote the number of valid ground-truth points in that series. We define chart-level Point Coverage as Pi=∑jLi,j∑jNi,jpoints.P_i= _jL_i,j _jN^points_i,j. (10) We report Point Coverage as the macro-average across aligned charts: PointCoverage=1||∑i∈Pi.PointCoverage= 1|A| _i P_i. (11) This definition distinguishes Point Coverage from ACR: ACR measures the quantity of extracted x-axis entries, whereas Point Coverage measures the fraction of ground-truth points that can be successfully matched. Effective Completeness (EC). To jointly account for chart-level recovery and within-chart completeness, we define EC=CAR×ACR.EC=CAR×ACR. (12) EC therefore penalizes both failures to recover the expected charts and incomplete reconstruction within successfully aligned charts. Series Matching. For multi-series charts, extracted and ground-truth series are first grouped according to their series labels. Labels are normalized using case folding and whitespace stripping. We then compute the Levenshtein edit distance Levenshtein, (1966) between each extracted series name and candidate ground-truth series names. For an extracted series ses_e and its closest ground-truth candidate sgs_g, the mapping is accepted only when d(se,sg)≤12max(|se|,|sg|),d(s_e,s_g)≤ 12 (|s_e|,|s_g| ), (13) where d(⋅,⋅)d(·,·) denotes Levenshtein distance. If a reliable series mapping cannot be established, the evaluator falls back to sequential row-order matching. x-Value Matching. Within successfully mapped series, numeric x-values are matched using greedy one-to-one nearest-neighbor assignment. Each extracted point is associated with its closest valid ground-truth x-value, and candidate matches are processed in ascending order of distance. Each ground-truth point can be used at most once. For categorical or otherwise non-numeric x-values, points are matched according to row order. Only pairs for which both x and y values can be converted to valid numerical values are included in numerical error computation. Normalization. For each aligned chart, let ℳiM_i denote the set of successfully matched point pairs. When at least two matched points are available, we compute ymin(i)=min(y^,y)∈ℳiy,ymax(i)=max(y^,y)∈ℳiy.y_ ^(i)= _( y,y) _iy, y_ ^(i)= _( y,y) _iy. (14) If ymax(i)−ymin(i)>0y_ ^(i)-y_ ^(i)>0, the extracted and ground-truth values are normalized to [0,100][0,100] as y^ℓnorm y_ ^norm =100y^ℓ−ymin(i)ymax(i)−ymin(i), =100 y_ -y_ ^(i)y_ ^(i)-y_ ^(i), (15) yℓnorm y_ ^norm =100yℓ−ymin(i)ymax(i)−ymin(i). =100 y_ -y_ ^(i)y_ ^(i)-y_ ^(i). (16) Thus, normalization is performed using the range of the matched ground-truth values within each chart, rather than independently for each series. Mean Absolute Error (MAE). For L successfully matched point pairs, the normalized Mean Absolute Error is MAE=1L∑ℓ=1L|y^ℓnorm−yℓnorm|.MAE= 1L _ =1^L | y_ ^norm-y_ ^norm |. (17) Root Mean Squared Error (RMSE). The normalized Root Mean Squared Error is RMSE=1L∑ℓ=1L(y^ℓnorm−yℓnorm)2.RMSE= 1L _ =1^L ( y_ ^norm-y_ ^norm )^2. (18) If normalization cannot be performed, the evaluator first uses relative errors, eℓrel=100|y^ℓ−yℓ||yℓ|,e_ ^rel=100 | y_ -y_ ||y_ |, (19) when |yℓ|>10−6|y_ |>10^-6. Relative errors are clipped at 100. If neither normalized nor relative errors are available, bounded absolute errors are used: eℓabs=min(|y^ℓ−yℓ|,100).e_ ^abs= (| y_ -y_ |,100 ). (20) MAE and RMSE are then computed as the arithmetic mean and root mean square of the selected pointwise errors, respectively. If either resulting value exceeds 100, both values are discarded and the chart is treated as unparseable. Aggregation. At the chart level, completeness, Point Coverage, MAE, and RMSE are averaged equally across aligned charts. For model-level reporting, these chart-level statistics are therefore aggregated using the number of aligned charts as the weight. Structural metrics such as RFV and CAR are computed directly from the corresponding compound- and chart-level counts. A.2 Overall reconstruction performance Table S1: Overall reconstruction performance comparison across different models and subplot counts. Model Subplots RFV EC PR Point Coverage MAE RMSE doubao 1 1.0000 0.8366 0.7629 0.5331 8.4932 11.9111 3 0.9914 0.7817 0.7652 0.4844 9.1226 13.1354 6 0.9310 0.6887 0.7500 0.4411 11.4940 16.2040 9 0.8182 0.5754 0.7213 0.3896 12.8941 17.5449 gemini 1 1.0000 0.8731 0.8428 0.6052 5.5119 8.1466 3 0.9310 0.7771 0.8889 0.5559 8.2033 11.6073 6 0.8879 0.7373 0.8819 0.5448 7.5849 11.0958 9 0.7273 0.6860 0.8770 0.5503 8.5136 12.0220 glm 1 1.0000 0.8343 0.8779 0.5622 8.7179 12.6893 3 0.8485 0.6727 0.8095 0.4527 11.5136 15.9457 6 0.6814 0.5442 0.7381 0.3878 14.2203 19.7835 9 0.5526 0.4872 0.7593 0.4085 16.1116 21.2302 kimi 1 1.0000 0.8793 0.7830 0.5586 7.9208 11.2605 3 0.9957 0.8436 0.8188 0.5426 8.4153 12.2063 6 1.0000 0.8410 0.8319 0.5364 8.7669 12.7310 9 0.9870 0.8250 0.8333 0.5285 9.4399 13.5002 qwen 1 1.0000 0.8442 0.8403 0.5801 7.0583 10.2186 3 0.7241 0.5718 0.8214 0.4854 9.9257 14.0155 6 0.6207 0.4583 0.7593 0.3935 8.9965 12.5950 9 0.9740 0.7161 0.7793 0.3987 12.7767 17.7086 A.3 RMSE density figures (a) (b) Figure S1: Paired chart-level comparison between the 1- and 9-chart conditions. Each point corresponds to the same source chart; values above the diagonal indicate higher error under density 9. The right panel shows the distribution of per-chart error changes. A.4 Mixed-effects model results Table S2: Regression results for the linear mixed-effects model. The reference categories are kimi for model and 1 for density. Term Coef. Std. Error z p-value CI Low CI High Intercept 8.7545 0.6280 13.9398 <<0.0001 7.5235 9.9854 Model (ref. = kimi) doubao 0.4171 0.6438 0.6479 0.5171 -0.8448 1.6789 gemini -2.4025 0.6306 -3.8097 0.0001 -3.6385 -1.1664 glm 0.3972 0.6235 0.6370 0.5241 -0.8249 1.6193 qwen -1.0514 0.6309 -1.6665 0.0956 -2.2879 0.1852 Density (ref. = 1) density = 3 0.3549 0.6339 0.5599 0.5755 -0.8876 1.5975 density = 6 0.6518 0.6305 1.0338 0.3012 -0.5839 1.8875 density = 9 1.5519 0.6331 2.4511 0.0142 0.3109 2.7929 Model × Density interactions doubao × 3 0.7319 0.9068 0.8071 0.4196 -1.0454 2.5093 gemini × 3 1.9301 0.8878 2.1741 0.0297 0.1901 3.6701 glm × 3 2.7341 0.9067 3.0156 0.0026 0.9571 4.5111 qwen × 3 3.0651 0.9322 3.2881 0.0010 1.2381 4.8922 doubao × 6 3.1270 0.9151 3.4169 0.0006 1.3333 4.9207 gemini × 6 1.7784 0.8924 1.9929 0.0463 0.0294 3.5275 glm × 6 4.0658 0.9569 4.2491 <<0.0001 2.1903 5.9412 qwen × 6 1.5839 0.9695 1.6338 0.1023 -0.3163 3.4840 doubao × 9 3.7571 0.9413 3.9912 0.0001 1.9121 5.6021 gemini × 9 1.6163 0.9217 1.7536 0.0795 -0.1902 3.4227 glm × 9 6.0766 0.9922 6.1243 <<0.0001 4.1319 8.0213 qwen × 9 4.8029 0.8987 5.3440 <<0.0001 3.0414 6.5644