Paper deep dive
LongChart VQA: A Comprehensive Benchmark for MLLMs with Complex Multi-Chart Reasoning
Ziyan Xiao, Yinghao Zhu, Wenting Zhang, Heaju Kim, Lequan Yu
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multimodal large language models (MLLMs) are rapidly evolving with expanded context windows and stronger reasoning capabilities, enabling multi-chart understanding and multi-step inference. These abilities are increasingly important as MLLMs are adopted in complex agentic tasks. However, existing benchmarks largely emphasize single-chart perception, while simple chart-to-chart connections are insufficient to evaluate these capabilities. To capture multi-chart complexity while ensuring consistency and validity, we design a synthesis pipeline supported by latent graphs. Building on this pipeline, we introduce LongChart, a benchmark whose VQA sets contain an average of 6.5 images and 31.2 questions. We evaluate 10 state-of-the-art MLLMs and examine three factors that influence performance: reasoning patterns, auxiliary tools, and robustness to image perturbations. Our results show that MLLM accuracy decreases and varies substantially as computational complexity increases, highlighting directions for future research in multi-chart reasoning.
Tags
Links
- Source: https://arxiv.org/abs/2608.01328v1
- Canonical: https://arxiv.org/abs/2608.01328v1
Trouble viewing inline? Open PDF directly â
Full Text
73,049 characters extracted from source content.
Expand or collapse full text
LongChart VQA: A Comprehensive Benchmark for MLLMs with Complex Multi-Chart Reasoning Ziyan Xiao, Yinghao Zhu, Wenting Zhang, Heaju Kim, Lequan Yu * The University of Hong Kong Abstract Multimodal large language models (MLLMs) rapidly evolve with expanding context and strengthened reasoning to support multi-chart understanding and multi-step inference. These abilities become increasingly important as MLLMs are adopted in complex agentic tasks. However, existing benchmarks largely empha- size single-chart perception, and simple chart- to-chart connections remain insufficient for evaluating MLLMs might be insufficient in measuring MLLMâs capability in these aspects. To capture multi-chart complexity while ensur- ing consistency and validity, we design a latent- graphâsupported synthesis pipeline. Building on this pipeline, we introduce LongChart, a benchmark that extends VQA sets to an average of 6.5 images and 31.2 questions. We evaluate 10 SOTA MLLMs on this benchmark, and fur- ther examine three aspects influencing perfor- mance, including reasoning patterns, auxiliary tools, and robustness against image perturba- tions. Our results show that MLLM accuracy decreases and varies significantly as compu- tational complexity increases, which offer a direction for future research in multi-chart rea- soning. 1 Introduction Multimodal large language models (MLLMs) have demonstrated strong reasoning ability on a wide range of tasks, including commonsense understand- ing (Yue et al., 2024; Lu et al., 2022; Marino et al., 2019; Yue et al., 2025), mathematical reason- ing (Lu et al., 2023; Zhang et al., 2024; Wang et al., 2024a), and visual understanding across diverse image sources. Driven by advances such as mul- timodal chain-of-thought prompting (Zhang et al., 2023) and reinforcement learning (Guo et al., 2025; Team et al., 2025), MLLMs are moving beyond single-image interpretation toward more complex cross-image and long-context settings (Fu et al., * Correspondence to Lequan Yu (lqyu@hku.hk) Robustness ĺ°˝ĺŻ č˝ĺžč˝ĺä¸é ďźć˘ ä¸ä¸ŞčŻ´ćł How MLLMsâ performance degrade in complex multi-chart scenarios? What might impact the reasoning performance? Auxiliary Tools Question(s) Reasoning Pattern Robustness - Bounding Box - Points - Extended Context - Multi-hop in-Depth - Perturbation - Transformation Input Image Set All intermediate data node separately as image input Input Image Set Minimal data node as image input Input Image Set Image-2 Style Transfer via GPT-Image-2 Resolution Noise Input Image Set Questions on Input Image Set cross-image interaction multi-hop reasoning multi-point inference DataComputation Node Image Ground Truth Prediction Latent Graph Depth â Context â Latent Graph support multi-chart VQA curation: Ground Truth Prediction Figure 1: Motivation of this work 2023; Kil et al., 2024; Huang et al., 2026). In these scenarios, maintaining precision in understanding across multiple images with intricate relationships among them become central challenges. Multi-chart, as a common medium for rich information, remains an underexplored area for MLLMâs performance evaluation.Pioneering works of chart-related benchmarks (Masry et al., 2022; Methani et al., 2020), primarily evaluate ba- sic chart understanding, such as element identi- fication and data retrieval. The follow-up works focuses on enhancing task complexity by compli- cated scientific plots (Wang et al., 2024b; Foroutan et al., 2025; Shen et al., 2024), injecting mislead- ing information(Chen et al., 2025), or entend to multilingual settings(Xu et al., 2025a). Some recent works have begun to explore multi- chart relationships, such as MultiChartQA (Zhu et al., 2025), which focuses on extraction and com- parison across charts, and InterChart (Iyengar et al., 2025), which increases task difficulty by synthe- sizing additional charts from single-chart datasets. However, these approaches often building up com- 1 arXiv:2608.01328v1 [cs.CL] 2 Aug 2026 Table 1: Comparison of chart-based reasoning benchmarks. Abbreviations: Extract = Data perception and extraction, Compare = Compare and analysis data pattern, Compute = Complex Data Computation, Ground. = Grounding, Point = Pointing, Eval. = Evaluation Format; = Full support, G#= Partial support, #= No support. Work CAPABILITIESAUXILIARY TOOLSVQA ExtractCompareComputeGround.PointImageEval. ChartQA(Masry et al., 2022) #G###1Finite Answer ChartBench(Xu et al., 2023) # ##1Finite Answer ChartXiv(Wang et al., 2024b) ##1â6+Open Answer MultiChartQA(Zhu et al., 2025) ###1-3Open Answer Ours 3-18Finite Answer plexity by LLM-based question generation and open-ended answers, which can introduce ambi- guity in evaluation. More importantly, the relation- ships between multiple charts that distinguishes multi-chart QA from single-chart tasks, especially considering professional multiple chart analysis involves intensive data computation and analysis. The automated synthesizing or grouping in previ- ous work may leave the benchmark lack of evalua- tion on these skills. To address these challenges, we design a latent graph-supported curation pipeline that provides a structured foundation for construct- ing complex multi-chart benchmarks. One of the central challenges is ensuring that charts are grounded in a shared and internally con- sistent data space, so that cross-image interactions yield meaningful results. To ensure the internal consistency, the synthetic pipeline is designed with the following properties: (1) latent graph as back- bone: data and images are defined and connected with logical relationship; (2) data consistency: data is synthesized only with a minimal set of nodes, with propagation to the rest to ensure unique- ness; and (3) difficulty through a growing hops not answer length: through close-end question, but the difficulty of questions is distributed with the number of hops, avoiding reliance on external knowledge or open-ended judgments. Based on the above synthesizing framework, we propose LONGCHART benchmark, a multiple chart VQA benchmark for evaluating the reasoning capa- bilities of MLLMs. The benchmark includes 557 images and 2,876 questions. We further largely ex- tending the number of question for each set of im- ages to study change of performance with question complexity. In the main set, each VQA instance contains on average 6.5 images, 31.2 questions, and approximately 716.8 data points. We evaluate 10 state-of-the-art MLLMs across these questions and further analyze three aspects that influence model reasoning, reasoning patterns, the use of auxiliary tools such as grounding and pointing, and robustness under image perturbations (Figure 1). Results show that when computational or analyti- cal complexity increases, MLLM accuracy declines sharply, which may highlighting directions for fu- ture exploration. In summary, we aim to realize the following objectives: â˘We propose a latent graphâsupported pipeline for synthesizing multi-chart, multi-question datasets, that overcomes the challenges in the benchmarking of complex computation ability in multi-chart problems. ⢠We construct LongChart Benchmark and evaluate 10 MLLMs reasoning ability when largely extends the context and reasoning depth of multi-chart question answering. â˘We provide a comprehensive analysis of model performance from three aspects, the impact of reasoning patterns, the role of auxil- iary tools such as grounding and pointing, and robustness against common image perturba- tions and transformations. 2 Related Work 2.1 Chart-Related Benchmarks Chart is a prevalent area of focus before the prevalence of MLLMs. Early chart-related bench- marks such as ChartQA (Masry et al., 2022) and PlotQA (Methani et al., 2020) established the foun- dation of this area. Subsequent benchmarks, includ- ing ChartBench (Xu et al., 2023), ChartXiv (Wang et al., 2024b), ChartX (Xia et al., 2025), and Sci- CQA (Shen et al., 2024), further extended evalu- ation to more complex figure formats and reason- ing skills. With the MLLMâs reasoning capabil- ities radidly improves, more recent benchmarks have started to explore extended and intergrated capabilities of chart reasoning, including multi- lingual settings (Xu et al., 2025a), multiple-chart 2 Method for data curation Step 1 Define logic graph Data Computation Node Latent Space Surface Space Image lossy Step 2 Derive minimal set Synthesize data for the minimal set with human instruction Step 3 Data propagation Human Algorithm LLM Algorithm Step 4 Question Construction Step 5 Image Generation ...until images alone formulate a minimal set Step 6 Human verification Algorithm LLM Objectives Uniqueness Consistency Sufficiency Question Answer Question Answer Synthesize from sample Question Answer Human Evaluation Methods Minimal-set Full-set Chart for Data Node Chart for Derived Node ⯠You will set a set of charts, and based on the charts, answer the following questions one by one <Input all charts> Question Question Question ⯠answer answer answer Conte xt Difficulty Conte xt Difficulty Benchmark Curation Human Two Reasoning Pattern Grounding and Pointing Three Image Perturbation ⯠You will set a set of charts... <Input all charts> Question Question Question ⯠Grounding/pointing Grounding/pointing Grounding/pointing answer answer answer Resolution Noise Style Transforma tion Original You will set a set of charts... <Input all charts> ⯠Figure 2: Overview of the VQA set curation process and evaluation. The pipeline comprises graph design, data synthesis, and VQA generation with human-in-the-loop quality control. reasoning (Zhu et al., 2025), and the visual in- sights (Tang et al., 2025). Multi-image benchmarks has become prevalent in other domains. Exist- ing works in this area includes, in-depth reasoning questions (Fan et al., 2025). Despite the progress, two main challenges in multi-chart VQA bench- marks stayed insufficiently explored. First, current benchmarks may yet to push sufficiently toward the long-context settings so that there saw significantly change in the reasoning pattern from pure details, to a global-and-detail thinking pattern. Second, for such a complex Multi-Chart benchmark, it requires more holistic evaluation than one simple question to accurately measure the boundary of reasoning performance, and indicate the factors that impacts the reasoning capabilities. 2.2 MLLM Reasoning with Chart Understanding Chart understanding is an essential capacity to mea- sure MLLMâs reasoning capability. Early works in improving MLLMâs Chart understanding fo- cused on curating large scale chart images, and post-training the MLLM models, through instruc- tion tuning(Han et al., 2023; Zeng et al., 2024), and supervised fine-tuning(Meng et al., 2024). Moreover, emergent works also focused on break- ing the modality constraints of charts, such as Matcha(Liu et al., 2023) propose chart-to-table alignment, ChartMoE(Xu et al., 2024) propose alignment between chart-table-JSON-code. There are increasing attempts in MLLM reasoning aimed to improve chart understanding via enhanced tools or fine-grained capabilities, including accurate pars- ing(Li et al., 2026; Xingchen et al., 2026), ground- ing(Xu et al., 2025b), and automatic edition or generation(Zhao et al., 2025). However, most of the works focused on precisely local unit of in- formation, such as data extraction due to their prevalency in early benchmarks. It remains un- clear that when moving to multi-chart scenarios, how might MLLMs performance retain or decrease when changing from simple precision-oriented rea- soning, to complex depth-oriented reasoning. 3 Benchmark Curation This benchmark provides comphensive evaluation on MLLMsâ understanding in multi-chart complex reasoning, and investigate the performance under reasoning patterns, auxilary tools and robustness testing. To enable the extension of context and rea- soning difficulty, as well as ensure the consistency of answer, we designed the following latent-graph driven VQA set curation process. In the following, we present the design for VQA dataset curation 3.1, validation 3.2, and dataset characteristics 3.3. The full VQA curation process is provided in Ap- pendix C. Details of human-in-the-loop curation plat form is in Appendix D. Some samples of VQA are provided in Appendix E 3 Database characteristics Images Questions No. images per set No. questions per set Data volume per set Answer Type Multi-choice Numbers Image Type (main) Line Bar Scatter Table Map Pie Others/Total Question Type Extraction Computation Comparison Image Data Volume < 100 dp 100-200 dp 200-300 dp > 300 dp dp: data points main robustness density 1474(51.3%) 1402(48.7%) 87(36.4%) 82(34.3%) 19(7.95%) 17(7.11%) 6(2.51%) 6(2.51%) 13/239 13.24% 43.21% 43.55% 46.69% 35.54% 10.38% 8.38% Images (main) perturbation (b) Example of VQA Image Pool Retrieve from Retrieve from Compute sum( ) Inference to get conclusion IMGDATA DATA IMG DATA Required Skills ... up to 31.1(avg) questions to test the complex processing abilities Question 1 Question 2 ...up to 6.5 (avg) images to form a rich context If you add up road, rail, and air passengers, which long-distance year had the highest total passengers? A. 2022 B. 2023 C. 2024 D. They are the same Retrieve from DATAIMG Question 3 Question N ⎠Question sum (a) (c) (d) (e) Figure 3: Dataset characteristics of LongChart VQA. (a) Question distribution across curation categories. (b) A sample question; (c) Image and question count. (d) Distribution of number of images, questions and data points per QA set; (e) Statistics in the VQA. 3.1 Latent Space That Supports VQA Curation Graph-as-backbone.To ensure consistency and meaningfulness across charts, we design the graph as the backbone of the VQA curation process. As shown in Figure 2, the graph is manually constructed to abstract real-world data relation- shipsâfor example, between GDP, inflation, and price index. This serves as a primary step in ensur- ing coherence across multiple data sources and is later used for data synthesis and propagation. Formally, for each question set we define a di- rected acyclic graph (DAG) G = (V,E), whereVdenotes the union of data nodes, compu- tation nodes (intermediate variables derived from data), and image nodes (visual representations of other nodes). EdgesEencode information flow and transfor- mations between vertices. In practice, deterministic operators transform data nodes into computation nodes, capturing common relationships such as summation, difference, ratio, proportion to totals, time-shift, or spatial distance. Information preservation is a critical attribute attached to each edge. It governs graph traversal for identifying minimal sets in data propagation and is defined in two categories: â˘An edge is bidirected, denoteduâ v, if the transformation is reversible betweenuandv. For example, a bar chart fully representing the values and distribution of a data node. â˘An edge is directed, denotedu â v, if the transformation is one-way anducannot be reconstructed fromv. For instance, ratios be- tween data nodes or year-on-year changes rel- ative to the original data. Minimal set and data propagation.Unlike sim- ple chart curation, multi-chart data must satisfy two constraints: (1) Uniqueness and Consistency. In- formation across charts must remain coherent and free of conflicts; (2) Sufficiency. The information revealed must be adequate to answer all curated questions. The latent graph design supports these requirements as follows. The information-preservation property of edges ensures that graph traversal yields a minimal set containing all information conveyed across charts, questions, and data in latent space. Nodes in this minimal set are unique and independent; thus, only these nodes are synthesized, as shown in Fig- ure 2(step 2). Data synthesis is performed via LLM 4 Table 2: M2MChartBench question-type accuracy across Extended Context and In-Depth evaluation. The best-performing model in each category is typeset in boldface, and the runner-up is indicated with underline. Human baseline is assessed under full chart set (Entended Context). Model Extended ContextIn-Depth Extract Compare ComputeTotalExtract Compare ComputeTotal Baselines Human81.78%81.78%81.90%83.10%â Proprietary MLLMs doubao-seed-2.0-lite84.37% 82.28%81.14%83.32%74.25%72.28%64.19%69.20% Doubao-seed-2.0-pro 86.25%84.57%86.57%85.76% 84.49%84.03%77.16%81.22% Gemini-3.1-flash-lite 82.66%75.38%74.29%79.37% 76.99%75.29%70.94%73.74% GPT-5.4-mini69.65%79.65%68.57%72.66% 75.93%74.82%56.53%67.33% GPT-5.4-nano45.80%63.89%40.00%50.79% 54.86%55.92%35.25%47.11% Qwen3.5-flash74.33%81.18%71.14%76.10% 80.56% 76.57%59.41%69.98% Open-source MLLMs Gemma-4-26b-a4b-it 68.99%69.31%74.24%69.50% 65.02%59.09%53.68%58.03% Gemma-4-31b-it36.29%42.96%36.40%38.36% 52.78%45.43%36.81%42.92% Mimo-v2.576.67%67.37%65.52%72.60% 76.62%73.55%65.24%70.53% Qwen3.5-35b-a3b77.68%82.82%72.86%78.71% 79.86%76.73%64.17%71.94% calls guided by human-defined trends to strengthen practical relevance, while remaining nodes are de- terministically computed through predefined trans- formations. This design guarantees uniqueness and consistency while reducing computational cost. Sufficiency is enforced by assessing whether charts alone form a minimal generative set of in- formation. Once data curation is finalized, images and QA pairs are generated independently, with edges directed or bidirected to latent data nodes depending on preservation properties. If gaps re- main, the algorithm automatically compensates un- til charts alone constitute a sufficient minimal set, after which the curated VQA set undergoes human validation. 3.2 Human-in-the-Loop Benchmark Construction Although the latent-space design and algorithmic procedures ensure the validity of the complex VQA set, a key challenge for synthesized charts is the concern that they may appear disconnected from practical use. To address this, we incorporate a human-in-the-loop process to enhance meaningful- ness, usefulness, and aesthetics. VQA curation.As shown in Figure 2, human ex- perts contribute at three critical stages of curation: (1) defining the latent graph structure to ensure that connections between data nodes are realistic and grounded in real-world scenarios; (2) specify- ing and refining aesthetic parameters during chart generationâsuch as image type, subplot layout, and color schemesâto align the synthetic charts more closely with practical usage; (3) verifying and reviewing each curated QA set to confirm that questions are readable and answerable. VQA validation.After curation, external evalua- tors are invited to answer the questions using only basic tools such as a calculator. Their performance is compared against the ground truth to establish a human-level accuracy baseline. 3.3 Benchmark Characteristics Figure 2 summarizes the question and image char- acteristics of the benchmark. The dataset con- tains 557 images and 2,876 questions across the main and robustness sets. Its long-context nature arises primarily from multiple charts and multiple questions within each VQA set. In the main set, each VQA instance includes on average 6.5 images (range: 3â18), 31.2 questions (range: 8â50), and 716.8 data points (range: 116â1,536). Hops are de- fined as the number of computation steps required to reach a conclusion. The average hop count is 1.41 (range: 0â8) in the main set and 1.35 (range: 0â8) in the robustness set. 4 Experiments 4.1 Baseline Models We evaluate a diverse suite of state-of-the-art MLLMs, spanning both proprietary APIs and open- 5 0 20 40 60 80 100 Accuracy (%) 430284215118413925 Extended Context 39.6 35.9 40.3 35.4 43.4 44.3 31.7 80.1 78.2 59.3 67.8 63.9 60.3 58.2 62.5 49.6 41.6 43.4 43.4 23.8 31.3 82.6 80.1 76.9 74.1 83.1 55.6 55.2 80.8 76.2 74.9 71.3 74.7 60.3 50.7 65.2 66.0 59.0 59.4 63.7 53.7 46.1 0123456+ Total Hops to Answer 0 20 40 60 80 100 Accuracy (%) 430284215118413925 In-Depth 55.1 52.1 41.4 43.2 39.0 23.1 12.0 70.0 64.3 55.3 52.5 39.0 33.3 8.0 56.3 45.1 36.7 28.8 31.7 20.5 12.0 68.8 64.7 58.6 61.0 58.5 46.2 20.0 68.8 63.6 56.3 61.9 58.5 35.9 16.0 65.2 60.2 51.6 51.8 48.4 33.8 15.3 0-12-34+ Total Hops to Answer gemini-3.1-flash-lite qwen3.5-35b-a3b gpt-5.4-mini qwen3.5-flash mimo-v2.5 doubao-seed-2.0-pro gpt-5.4-nano gemma-4-31b 87.179.766.7 80.174.175.0 78.267.877.8 76.271.369.4 66.553.158.3 54.250.727.8 49.643.455.6 35.935.450.0 (a) Performance with Hops(b) MLLM ranks Performance with Hops Mean Q1Q3 Question count gemma-4-31b gpt-5.4-mini gpt-5.4-nano qwen3.5-35b-a3b qwen3.5-flash MLLM ranks Rank 1 (Best) Rank 2 Rank 3 Rank 4 Rank 5 Rank 6 Rank 7 Rank 8 (Worst) Figure 4: Model performance as a function of reasoning hops. (a) Accuracy degradation across increasing reasoning complexity under Extended Context (top) and In-Depth (bottom).(b) Performance ranks across model under different difficulty level of questions source state-of-the-art models. Proprietary mod- els. We include Doubao-Seed-2.0-Lite(ByteDance Seed, 2026), Doubao-Seed-2.0-Pro(ByteDance Seed, 2026), Gemini-3.1-Flash-Lite(DeepMind, 2026a), GPT-5.4-Mini(OpenAI, 2026), GPT- 5.4-Nano(OpenAI,2026),andQwen-3.5- Flash(DeepMind, 2026b). Open-source models. We additionally test Gemma-4-26B-IT(DeepMind, 2026c), Gemma-4-35B-A3B(DeepMind, 2026c), and Qwen-3.5-35B-A3B(Qwen Studio, 2026). The baseline configuration is provided in Ap- pendix G. Evaluation metric is defined in Ap- pendix F. Prompts used in the baseline evaluation are listed in Appendix H. Beyond the baseline, we design experiments to examine three dimensions that affect MLLM per- formance: (i) reasoning patterns, (i) grounding and pointing as auxiliary tools, and (i) robustness why implying image perturbations and transformation that approximate real-world settings. As illustrated in Figure 2, these dimensions are defined in detail as follows. 4.2 Reasoning Patterns Extended context.Visualizing intermediate vari- ables in charts provides a shortcut for reducing reasoning complexity. As illustrated in Figure 2(b), the Extended Context mode supplies visualizations for every computation node in the MLLM queries. This approach can be viewed as a redistribution of reasoning effort, shifting from multi-step inference to increased context length. In-depth reasoning. As shown in Figure 2(b), MLLMs can also answer all questions when pro- vided with only a minimal set of charts. As dis- cussed in Section 3, the latent-space graph ensures that such minimal sets of images are obtainable within the VQA dataset, containing the necessary information to answer each question. A compari- son between Extended Context and In-Depth Rea- soning highlights the reasoning patterns preferred by MLLMs. 4.3 Grounding and Pointing Bounding boxes and point annotations are strong indicators of an MLLMâs visual reasoning on charts and can serve as intermediate outputs in multi- image inference. Evaluating their accuracy pro- vides essential evidence of reasoning quality and supports multi-step inference. Leveraging the syn- thetic nature of our VQA set, ground-truth anno- tations for bounding boxes and points are readily available. For evaluation, we select 548 questions with clear visual cues. In both modes, the MLLM is prompted to return not only the final answer but also the bounding box or point corresponding to the key information. We report accuracy results and conduct bounding-box and point analyses. The prompts are provided in Appendix H. 4.4 Robustness To evaluate robustness, we select 10 VQA subsets comprising a total of 1722 questions and 318 im- ages. These subsets are stratified across curated 6 Ground Truth Encompass the Entire Chart Wrong Grounding Close to Ground Truth (GT) but not Include GT Figure 5: Performance with grounding and pointing. (a) mDIoU and Chamfer distance indicate low quality in generated grounding box and point, with significant decrease in question answer accuracy. (b) a case study showing the prevalent patterns of low-quality grounding VQA categories, with average accuracy of the best- performing models ranging from 0.6 to 0.8. Each subset is subjected to seven types of image per- turbations and transformations, with details in Ap- pendix B. The impact on MLLM performance is quantified by the accuracy difference between the original VQA set and its perturbed or transformed counterpart. 5 Results 5.1 MLLM Performance Table 2 summarizes the performance of state-of- the-art MLLMs across different question types. Figure 4 visualized the MLLMs performance dis- tribution with total computational hops, and the rank of MLLM models under difficulty levels. To provide a holistic view, we also report MLLM per- formance across different VQA set types in Ap- pendix A.1, and type of charts in Appendix A.2. Computation questions pose challenges. As shown in Table 2, accuracy in multi-hop compu- tation questions is consistently lower than overall performance. In fact, the gap between computa- tion accuracy and that of extraction or comparison tasks show as a good indicator of a modelâs reason- ing ability across multiple images. Under in-depth input settings, the best-performing model, Doubao- seed-2.0 pro, showed a drop of8.68%compared to extraction tasks, while the gap widened to35.75% in GPT-5.4 nano and30.52%in Gemma-4-31b. The drop of accuracy in computation is a shared trend across the full set of image as input (Extended Context) and minimal set of image as input (In Depth), and is further verified in the following hop- related analysis. Hops hinder performance more than context length.Table 2 further shows that models achieve higher accuracy in Extended Context settings than in In-Depth settings. Extended Context converts some multi-hop computations into intermediate vi- sualizations, suggesting that reducing hops in ex- change for longer context benefits performance. Figure 4(a) highlights this trend, as accuracy de- clines sharply once total hops grows. In Extended Context, average accuracy drops to53.7%at five hops and further falls to46.1%beyond six hops, whereas in In-Depth settings accuracy drops more significantly to33.8%at five hops and15.3%be- yond six. Data computation and conversion remain the task-level bottleneck. These difficulties mani- fest at the task level as persistent performance gaps. Tables 3 and 4 show that MLLMs perform best on tasks with direct and explicit data, with- out complex inference. For example, table-based VQA tasks under Extended Context achieve the highest accuracies, reaching93.43%for the best- performing models. In contrast, tasks involving less direct information show significant drops, such as82.99%under In-Depth settings, or68.57%on time-sequence data with the same best-performing models. 7 DataNoiseTextNoiseVisualNoiseBlurring Compression Trans doubao-seed-2.0-lite doubao-seed-2.0-pro gemini-3.1-flash-lite-preview gemma-4-31b-it gpt-5.4-mini gpt-5.4-nano qwen3.5-35b-a3b qwen3.5-flash -4.2-1.4-1.40.0-1.4-5.9 -2.1-1.40.0-0.3-1.1-2.8 -1.4+2.1+0.7-0.7-3.1+0.3 +9.1-8.0-16.9+5.9-11.2-6.2 +0.3+3.5+0.3+0.7+3.5-2.1 -20.6+3.1-3.10.0+1.1+0.3 -4.2-4.2-0.7-1.4-0.7-4.2 -6.6-4.2-0.7-7.0-3.8-6.3 Extended Context -20 -15 -10 -5 0 5 10 15 20 Accuracy Change (%) Figure 6: Impact of robustness perturbations. The value shows equals the accuracy in the robustness set minus the accuracy for the corresponding original syn- thesized charts. Trans=Image Transformation. 5.2 Analysis Grounding and pointing are inaccurate. Fig- ure 5(a) illustrates both the low grounding qual- ity and performance decline when MLLMs are re- quired to generate bounding boxes or point anno- tations to cross-validate their answers. All models yield negative mDIoU scores, with mean Chamfer Distance ranging from0.316to0.380. Case stud- ies in Figure 5 show that errors arise from outputs that are either completely irrelevant, approximate but imprecise, or overly broad such as circling the entire plot rather than the key area. Moreover, pro- ducing bounding-box or point coordinates often reduces overall performance significantly. For in- stance, generating grounding box causes best per- forming models drop from85.5%of accuracy to 57.6%. These findings highlight the difficulty of employing self-adjusted image tools as intermedi- ate steps in agentic pipelines to improve multi-hop QA reasoning. Accuracy deviation under robustness settings. Heterogeneous effects of image perturbation and transformation imposed across MLLMs. As shown in Figure 6, robustness transformation impose not only negative effects, but also interestingly, some positive effects on model performance. Overall leading models present strong robustness against common image perturbations. Data noise and im- age transformation poses strong to model perfor- mance. Cost-performance frontier. Figure 7 illustrates the costâperformance frontier of MLLMs under breadth and depth modes. Two clear efficient front- tier were shown both in the token measurement and time measurement, indicating space for im- provement when extending the inference time and context. Switching from Extended Context mode to In-Depth mode, lead to 0.55x to 0.76x times re- duction in tokens, and 0.36x to 3.68x change in pro- cesing time. At present, the sacrifice of times and tokens in exchange for accuracy pattern is present. 1000020000300004000050000 Total Time (seconds) 40 50 60 70 80 Accuracy (%) mimo-v2.5 mimo-v2.5 doubao-seed-2.0-pro doubao-seed-2.0-pro gpt-5.4-mini gpt-5.4-mini doubao-seed-2.0-lite doubao-seed-2.0-lite gpt-5.4-nano gpt-5.4-nano qwen3.5-35b-a3b qwen3.5-35b-a3b qwen3.5-flash qwen3.5-flash Extended Context In-Depth 0.81.01.21.41.61.82.02.2 Total Tokens 1e7 40 50 60 70 80 Accuracy (%) mimo-v2.5 mimo-v2.5 doubao-seed-2.0-pro doubao-seed-2.0-pro gpt-5.4-mini gpt-5.4-mini doubao-seed-2.0-lite doubao-seed-2.0-lite gpt-5.4-nano gpt-5.4-nano qwen3.5-35b-a3b qwen3.5-35b-a3b qwen3.5-flash qwen3.5-flash Extended Context In-Depth Figure 7: Cost-performance distribution of evaluated models. Two seperate efficient frontier curves of Ex- tended Context and In-Depth presents, with both token count (top) and runtime (bottom) versus accuracy. 6 Conclusion This paper introduces LONG-CHART BENCH, a new pipeline and benchmark designed to evalu- ate MLLMsâ visual reasoning and performance in multi-chart settings with complex computational relationships. Our evaluation of state-of-the-art MLLMs reveals a widening performance gap be- tween MLLMs as data become less apparent due to multi-chart interaction and computational difficulty. We further observe that current models remain vul- nerable when required to produce grounding out- puts and when exposed to external perturbations. These findings underscore the need for future ad- vances in MLLMs to strengthen visual reasoning in chart contexts, particularly as between-chart re- lationships grow more intricate. Limitations We acknowledge the following limitations of this work and outline directions for future improve- ment. 8 Scope of the QA Benchmark. Compared with other synthetic chart benchmarks, the number of images and questions in our dataset is relatively limited. This constraint arises from the human-in- the-loop design and the difficulty of validating each question set, which requires approximately 30 min- utes per set (e.g., 5 images with 30 questions) and a total curation period of about 15 days. While the dataset size is modest, we hope to provide a pre- liminary yet holistic view of multi-chart scenarios with complex relationships, complementing with diverse comparative evaluations under difference baseline settings and image scenarios. Coverage of Models. Due to time constraints, the number of evaluated MLLMs is limited relative to the broad range of models currently available. We selected representative models from leading developers, but still the evaluation does not capture the full spectrum of existing SOTA models. We plan to continue expanding coverage to a wider set of models to provide a more comprehensive assessment. References ByteDance Seed. 2026.Seed2.0.https:// seed.bytedance.com/en/seed2. [Accessed 05-05- 2026]. Zixin Chen, Sicheng Song, Kashun Shum, Yanna Lin, Rui Sheng, Weiqi Wang, and Huamin Qu. 2025. Unmasking deceptive visuals: Benchmarking mul- timodal large language models on misleading chart question answering. In Proceedings of the 2025 Con- ference on Empirical Methods in Natural Language Processing, pages 13767â13800. Google DeepMind. 2026a.Gemini 3.1 Flash-Lite â deepmind.google.https://deepmind.google/ models/gemini/flash-lite/. [Accessed 05-05- 2026]. Google DeepMind. 2026b.Gemini 3.1 Pro â deepmind.google.https://deepmind.google/ models/gemini/pro/. [Accessed 05-05-2026]. Google DeepMind. 2026c.Gemma 4 - our most intelligent open models, built from gemini 3 re- search and technology to maximize intelligence-per- parameter.https://deepmind.google/models/ gemma/gemma-4/. [Accessed 05-05-2026]. Wan-Cyuan Fan, Yen-Chun Chen, Mengchen Liu, Alexander Jacobson, Lu Yuan, and Leonid Si- gal. 2025. In-depth and in-breadth: Pre-training multimodal language models customized for com- prehensive chart understanding.arXiv preprint arXiv:2507.14298. Negar Foroutan, Angelika Romanou, Matin Ansaripour, Julian Martin Eisenschlos, Karl Aberer, and RĂŠmi Lebret. 2025. Wikimixqa: a multimodal benchmark for question answering over tables and charts. In Findings of the Association for Computational Lin- guistics: ACL 2025, pages 24941â24958. Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xi- awu Zheng, Ke Li, Xing Sun, and 1 others. 2023. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, and 1 others. 2025. Deepseek- r1 incentivizes reasoning in llms through reinforce- ment learning. Nature, 645(8081):633â638. Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. 2023. Chartllama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483. Wenxuan Huang, Yu Zeng, Qiuchen Wang, Zhen Fang, Shaosheng Cao, Zheng Chu, Qingyu Yin, Shuang Chen, Zhenfei Yin, Lin Chen, and 1 others. 2026. Vision-deepresearch: Incentivizing deepresearch ca- pability in multimodal large language models. arXiv preprint arXiv:2601.22060. Anirudh Iyengar Kaniyar Narayana Iyengar, Srija Mukhopadhyay, Adnan Qidwai, Shubhankar Singh, Dan Roth, and Vivek Gupta. 2025. Interchart: Bench- marking visual reasoning across decomposed and dis- tributed chart information. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Lin- guistics, pages 2046â2067. Jihyung Kil, Zheda Mai, Justin Lee, Arpita Chowdhury, Zihe Wang, Kerrie Cheng, Lemeng Wang, Ye Liu, and Wei-Lun Chao. 2024. Mllm-compbench: A com- parative reasoning benchmark for multimodal llms. Advances in Neural Information Processing Systems, 37:28798â28827. Jinsong Li, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Jiaqi Wang, and Dahua Lin. 2026. Visual self-refine: A pixel-guided paradigm for accurate chart parsing. arXiv preprint arXiv:2602.16455. Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Eisenschlos. 2023. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. In Proceed- ings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12756â12770. 9 Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chun- yuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai- Wei Chang, Michel Galley, and Jianfeng Gao. 2023. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255. Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai- Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in neural information processing systems, 35:2507â2521. Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. 2019. Ok-vqa: A visual ques- tion answering benchmark requiring external knowl- edge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195â3204. Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. 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 2263â 2279. Fanqing Meng, Wenqi Shao, Quanfeng Lu, Peng Gao, Kaipeng Zhang, Yu Qiao, and Ping Luo. 2024. Char- tassistant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning. In Findings of the Association for Computational Linguistics: ACL 2024, pages 7775â 7803. Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. 2020. Plotqa: Reasoning over sci- entific plots. In Proceedings of the ieee/cvf winter conference on applications of computer vision, pages 1527â1536. OpenAI. 2026. Introducing GPT-5.4 mini and nano â openai.com.https://openai.com/index/ introducing-gpt-5-4-mini-and-nano.[Ac- cessed 05-05-2026]. Qwen Studio. 2026. Qwen3.5: Towards native multi- modal agents.https://qwen.ai/blog?id=qwen3. 5. [Accessed 05-05-2026]. Lingdong Shen, Kun Ding, Gaofeng Meng, Shiming Xiang, and 1 others. 2024. Rethinking compre- hensive benchmark for chart understanding: A per- spective from scientific literature. arXiv preprint arXiv:2412.12150. Liyan Tang, Grace Kim, Xinyu Zhao, Thom Lake, Wenxuan Ding, Fangcong Yin, Prasann Singhal, Manya Wadhwa, Zeyu Leo Liu, Zayne Sprague, and 1 others. 2025. Chartmuseum: Testing visual rea- soning capabilities of large vision-language models. arXiv preprint arXiv:2505.13444. Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, and 1 others. 2025. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599. Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. 2024a. Measuring multimodal mathematical reason- ing with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095â95169. Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, and 1 others. 2024b. Charxiv: Charting gaps in realistic chart understand- ing in multimodal llms. Advances in Neural Informa- tion Processing Systems, 37:113569â113697. Renqiu Xia, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Botian Shi, Junchi Yan, and Bo Zhang. 2025. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. IEEE Transactions on Image Pro- cessing. ZENG Xingchen, Zhewei Su, Hengming Zhang, Juyong Jiang, Jiazhi Xia, and Wei Zeng. 2026. Davinci: Re- inforcing visual-structural syntax in mllms for gener- alized scientific diagram parsing. In The Fourteenth International Conference on Learning Representa- tions. Yichen Xu, Liangyu Chen, Liang Zhang, Jianzhe Ma, Wenxuan Wang, and Qin Jin. 2025a. Poly- chartqa: Benchmarking large vision-language models with multilingual chart question answering. arXiv preprint arXiv:2507.11939. Zhengzhuo Xu, SiNan Du, Yiyan Qi, Siwen Lu, Chengjin Xu, Chun Yuan, and Jian Guo. 2025b. Chartpoint: Guiding mllms with grounding reflection for chart reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 426â436. Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo. 2023. Chartbench: A bench- mark for complex visual reasoning in charts. arXiv preprint arXiv:2312.15915. Zhengzhuo Xu, Bowen Qu, Yiyan Qi, Sinan Du, Chengjin Xu, Chun Yuan, and Jian Guo. 2024. Chartmoe: Mixture of diversely aligned expert con- nector for chart understanding.arXiv preprint arXiv:2409.03277. Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, and 1 others. 2024. Mmmu: A massive multi-discipline multimodal un- derstanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 9556â 9567. Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, 10 Ge Zhang, Huan Sun, and 1 others. 2025. Mmmu- pro: A more robust multi-discipline multimodal un- derstanding benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15134â 15186. Xingchen Zeng, Haichuan Lin, Yilin Ye, and Wei Zeng. 2024. Advancing multimodal large language mod- els in chart question answering with visualization- referenced instruction tuning. IEEE Transactions on Visualization and Computer Graphics, 31(1):525â 535. Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, and 1 others. 2024. Math- verse: Does your multi-modal llm truly see the dia- grams in visual math problems? In European Confer- ence on Computer Vision, pages 169â186. Springer. Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. 2023. Multi- modal chain-of-thought reasoning in language mod- els. arXiv preprint arXiv:2302.00923. Xuanle Zhao, Xuexin Liu, Yang Haoyue, Xianzhen Luo, Fanhu Zeng, Jianling Li, Qi Shi, and Chi Chen. 2025. Chartedit: How far are mllms from automating chart analysis? evaluating mllmsâ capability via chart edit- ing. In Findings of the Association for Computa- tional Linguistics: ACL 2025, pages 3616â3630. Zifeng Zhu, Mengzhao Jia, Zhihan Zhang, Lang Li, and Meng Jiang. 2025. Multichartqa: Benchmarking vision-language models on multi-chart problems. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu- tational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 11341â11359. A Extended Evaluation Results and Analysis A.1 Extended MLLM Performance Across Data Curation Taxonomy This subsection extends the main benchmark com- parison by reporting model performance across the detailed data curation taxonomy. We separate the results into complementary views so that readers can inspect model behavior on both the main and robustness splits in a more fine-grained manner. A.2 Performance Across Figure Types We further analyze performance variation across figure types. To provide an intuitive summary of category-level strengths and weaknesses, we visu- alize the results with radar plots for the two evalua- tion splits and reserve a dedicated table for a more detailed category-by-category comparison. A.3 Model Cost-Performance Distribution Finally, we examine the trade-off between model performance and token cost on the main and ro- bustness benchmarks. These scatter plots provide a compact view of which models achieve stronger accuracy under different cost budgets. A.4 Comparison of MLLM Answers Before and After Reasoning This subsection compares model accuracy before and after explicit reasoning across different ques- tion types. The results help reveal whether per- formance gains mainly come from improved rea- soning depth or from better answer grounding on specific question formats. 11 Table 3: MLLM performance across VQA curation categories under Extended Context. The best-performing model in each category is typeset in boldface, and the runner-up is indicated with underline. Model MainRobustness GEOMETRY MULTICHART STATISTICS TABLE TIMESEQTotalIMAGESTYLE NOISE RESOLUTIONTotal doubao-seed-2.0-pro90.34%84.64%79.25%93.13%66.86%84.15%85.37%86.99%87.46%86.88% doubao-seed-2.0-lite86.90%83.70%77.36%90.15%66.86%82.29%80.49%84.09%85.71%84.03% gemini-3.1-flash-lite-preview79.31%84.64%71.70%86.57%61.14%78.75%80.49%80.60%78.22%79.79% gemma-4-31b-it20.00%56.74%41.98%28.66%1.71%33.56%42.59%40.28%49.20%43.85% gpt-5.4-mini73.79%73.35%61.32%82.39%53.71%70.91%70.73%74.22%74.91%73.87% gpt-5.4-nano48.28%57.05%52.83%61.79%28.00%52.28%53.31%46.11%53.48%49.77% qwen3.5-35b-a3b86.21%74.92%75.00%88.06%56.57%77.32%78.05%79.21%81.18%79.67% qwen3.5-flash83.45%77.43%74.06%82.09%55.43%75.63%74.91%77.35%75.78%76.42% Table 4: MLLM performance across VQA curation categories under In-Depth evaluation. The best-performing model in each category is typeset in boldface, and the runner-up is indicated with underline. Model MainRobustness GEOMETRY MULTICHART STATISTICS TABLE TIMESEQTotalIMAGESTYLE NOISE RESOLUTIONTotal doubao-seed-2.0-lite67.59%66.46%64.62%77.31%42.29%65.77%61.54%77.49%82.39%77.15% doubao-seed-2.0-pro68.97%72.41%66.51%82.99%49.71%70.57%86.76%88.73%89.20%88.56% gemma-4-31b-it57.24%61.13%42.45%46.27%22.86%47.47%39.72%37.05%43.90%39.78% gpt-5.4-mini56.55%61.76%62.74%71.64%33.71%59.95%70.73%72.36%73.34%72.42% gpt-5.4-nano46.21%49.53%34.43%54.03%20.57%43.42%50.17%47.27%52.96%49.65% mimo-v2.560.00%66.14%59.43%76.12%32.00%61.97%77.35%76.77%75.44%76.42% qwen3.5-35b-a3b61.38%57.05%60.85%77.91%41.14%61.80%78.75% 79.33%78.40%78.92% qwen3.5-flash61.38%58.31%52.83%80.90%36.00%60.79%67.94%77.24%79.09%76.31% Bar Box Plot Heatmap Histogram Line Map Pie Range Plot Scatter Stacked Bar Table 20% 40% 60% 80% 100% Accuracy by Chart Type (Extended Context) doubao-seed-2.0-pro gemini-3.1-flash-lite-preview gemma-4-31b-it gpt-5.4-mini gpt-5.4-nano mimo-v2.5 qwen3.5-35b-a3b qwen3.5-flash Bar Heatmap Histogram Line Map Pie Range Plot Stacked Bar Table 20% 40% 60% 80% 100% Accuracy by Chart Type (In-Depth) doubao-seed-2.0-lite gemma-4-31b-it gpt-5.4-mini gpt-5.4-nano qwen3.5-35b-a3b qwen3.5-flash Figure 8: Performance comparison across figure types on the two evaluation splits. Table 5: MLLM performance across figure types under Extended Context. The best-performing model in each category is typeset in boldface, and the runner-up is indicated with underline. ModelBarBox plot Heatmap HistogramLineMapPieRange plot Scatter Stacked bar Table doubao-seed-2.0-pro85.83% 100.00% 77.78%96.43%83.77% 74.29% 85.00%86.05%66.67%72.73%92.22% doubao-seed-2.0-lite85.43%66.67%77.78%91.07%82.45%71.43% 85.00%90.70%58.33%54.55%85.56% gemini-3.1-flash-lite-preview 81.10% 100.00% 74.07%82.14%82.45%68.57% 85.00%76.74%50.00%90.91%83.33% gemma-4-31b-it40.14%0.00%46.67%46.43%37.06% 28.57% 50.00%33.78%20.83%45.45%14.29% gpt-5.4-mini71.26% 33.33%62.96%76.79%72.19% 71.43% 75.00%69.77%58.33%72.73%77.78% gpt-5.4-nano51.97% 33.33%48.15%64.29%53.64% 54.29% 50.00%44.19%12.50%63.64%62.22% mimo-v2.574.71% 33.33%72.97%91.07%72.19% 66.67% 65.00%76.60%62.50%76.47%77.48% qwen3.5-35b-a3b77.56% 66.67%81.48%91.07%79.14% 74.29% 80.00%53.49%66.67%63.64%81.11% qwen3.5-flash75.59% 33.33%81.48%91.07%73.51% 74.29% 85.00%72.09%66.67%54.55%80.00% Total71.68% 51.85%69.83%79.64%70.89% 65.76% 74.12%67.53%50.83%65.35%72.30% 12 Table 6: MLLM performance across figure types under In-Depth evaluation. The best-performing model in each category is typeset in boldface, and the runner-up is indicated with underline. ModelBarHeatmap HistogramLineMapPieRange plot Stacked barTable doubao-seed-2.0-lite 64.98%65.91%56.45%69.55%73.17%71.43%59.26%87.50%67.84% doubao-seed-2.0-pro 72.06%63.64%53.23%80.66% 75.61% 66.67%29.41%100.00%70.07% gemma-4-31b-it44.12%52.27%24.19%53.91% 70.73% 38.10%17.65%62.50%42.86% gpt-5.4-mini56.37%47.73%53.23% 64.20% 70.73% 61.90%29.41%62.50%57.82% gpt-5.4-nano44.12%45.45%17.74%40.74% 60.98% 47.62%17.65%37.50%49.66% mimo-v2.561.27%52.27%40.32%66.67% 65.85% 66.67%23.53%75.00%66.67% qwen3.5-35b-a3b69.12%40.91%46.77%68.72% 73.17%38.10%23.53%87.50%63.27% qwen3.5-flash65.69%59.09%37.10%65.84% 70.73% 61.90%29.41%100.00%68.03% Total59.72%53.41%41.13%63.79% 70.12% 56.55%28.73%76.56%60.78% Table 7: Accuracy change under visual perturbations. Abbreviation: Trans = Transformation. (D) denotes In-Depth evaluation mode. ModelData NoiseText NoiseVisual NoiseBlurringCompressionTrans doubao-seed-2.0-liteâ4.18%â1.39%â1.39%0.00%â1.39%â5.92% doubao-seed-2.0-lite (D)+24.59%+25.41%â3.70%+17.05%+15.91%+3.08% doubao-seed-2.0-proâ2.09%â1.39%0.00%â0.35%â1.05%â2.79% doubao-seed-2.0-pro (D)+20.21%+19.16%+18.82%+19.16%+20.56%+17.42% gemini-3.1-flash-lite-previewâ1.39%+2.09%+0.70%â0.70%â3.14%+0.35% gemma-4-31b-it+9.09%â8.02%â16.92%+5.88%â11.23%â6.17% gemma-4-31b-it (D)â5.57%â4.53%â19.86%â0.70%â5.57%â7.32% gpt-5.4-mini+0.35%+3.48%+0.35%+0.70%+3.48%â2.09% gpt-5.4-mini (D)+12.20%+10.10%+7.67%+7.32%+14.63%+8.36% gpt-5.4-nanoâ20.56%+3.14%â3.14%0.00%+1.05%+0.35% gpt-5.4-nano (D)â6.97%+5.57%0.00%+8.01%+2.44%+2.44% mimo-v2.5 (D)+11.50%+14.98%+12.54%+11.85%+11.50%+13.59% qwen3.5-35b-a3bâ4.18%â4.18%â0.70%â1.39%â0.70%â4.18% qwen3.5-35b-a3b (D)+14.98%+9.06%+16.38%+6.97%+18.12%+12.89% qwen3.5-flashâ6.62%â4.18%â0.70%â6.97%â3.83%â6.27% qwen3.5-flash (D)+16.72%+19.86%+20.56%+23.00%+18.82%+9.76% Extract CompareCompute 0 20 40 60 80 100 Accuracy (%) Extended Context 77.7 32.3 82.0 36.3 83.5 40.4 82.8 43.0 73.4 29.8 76.1 36.4 Accuracy Before Reasoning After Reasoning doubao-seed-2.0-lite doubao-seed-2.0-pro gemini-3.1-flash-lite gemma-4-31b gpt-5.4-mini gpt-5.4-nano mimo-v2.5 qwen3.5-35b-a3b qwen3.5-flash Extract CompareCompute 0 20 40 60 80 100 Accuracy (%) In-Depth 80.6 51.6 80.6 52.8 77.0 43.0 76.7 45.4 63.3 24.5 64.2 35.2 Accuracy Before Reasoning After Reasoning doubao-seed-2.0-lite doubao-seed-2.0-pro gemini-3.1-flash-lite gemma-4-31b gpt-5.4-mini gpt-5.4-nano mimo-v2.5 qwen3.5-35b-a3b qwen3.5-flash Figure 9: Comparison of model accuracy before and after reasoning across question types. Table 8: Performance before reasoning under Extended Context. The best-performing model in each category is typeset in boldface, and the runner-up is indicated with underline. An em dash (â) denotes unavailable results. Model MainRobustness GEOMETRY MULTICHART STATISTICS TABLE TIMESEQTotalIMAGESTYLE NOISE RESOLUTIONTotal doubao-seed-2.0-pro90.34%84.64%78.30%92.84%66.86%85.37%86.99%87.46%83.90%86.88% doubao-seed-2.0-lite84.83%83.39%77.83%89.85%66.86%81.18%83.97%85.54%81.96%84.03% gemini-3.1-flash-lite-preview76.55%77.74%68.87%77.01%53.14%71.43%73.40%70.21%72.18%72.01% gemma-4-31b-it20.00%51.41%35.85%26.57%1.14%30.35%38.89%35.91%44.12%39.33% gpt-5.4-mini68.28%61.76%50.94%65.67%48.57%59.78%60.63%59.70%63.94%61.27% gpt-5.4-nano41.38%42.01%39.15%40.30%26.29%38.62%43.21%35.77%39.20%38.15% qwen3.5-35b-a3b86.21%74.92%75.00%88.06%56.57%77.32%80.14%80.14%77.32%80.14% qwen3.5-flash83.45%77.12%73.11%83.88%55.43%75.89%74.22%76.66%75.78%75.96% 13 Table 9: Performance before reasoning under In-Depth evaluation. The best-performing model in each category is typeset in boldface, and the runner-up is indicated with underline. Model MainRobustness GEOMETRY MULTICHART STATISTICS TABLE TIMESEQTotalIMAGESTYLE NOISE RESOLUTIONTotal doubao-seed-2.0-lite66.90%65.52%64.15%75.52%41.71%64.76%60.00%77.12%82.39%76.76% doubao-seed-2.0-pro68.97%71.79%66.04%82.09%49.71%70.07%86.41%88.50%89.20%88.39% gemma-4-31b-it56.55%54.23%40.09%40.00%17.14%42.50%34.49%31.94%38.50%34.55% gpt-5.4-mini53.79%53.29%51.89%54.63%24.00%49.16%57.84%61.44%62.72%61.27% gpt-5.4-nano40.69%37.30%26.89%40.30%17.14%33.73%39.02%37.75%41.64%39.26% mimo-v2.560.00%66.14%59.43%76.12%30.86%61.80%77.35%76.77%75.44%76.42% qwen3.5-35b-a3b62.07%57.05%60.85%75.82%41.14%61.30%78.75%78.51%79.44%78.86% qwen3.5-flash58.62%57.05%52.83%76.12%36.00%58.77%67.94%77.82%79.44%76.71% 14 B Task Descriptions The benchmark is organized into two complemen- tary splits. The main split comprises VQA sets curated under diverse scenario settings and node- relationship configurations, spanning normal tab- ular data, time-series data, and spatial data. The robustness split reuses the same questionâanswer pairs from the main split but applies controlled visual perturbations to the images, ensuring that answer semantics and human readability remain intact so that any accuracy change can be attributed solely to the perturbation. Main subset. The main subset is categorized by the underlying scenario and the linkage functions that connect nodes in the latent dependency graph. We curate five categories: ⢠SingleDense: VQA sets built from a single chart type that contains a high volume of data points, stressing dense-context reading and fine-grained value extraction. â˘MultiChart: The base category in which each VQA set combines heterogeneous chart types, requiring cross-chart aggregation and compar- ison. â˘StatAnalysis: Sets focused on statistical measurement and distributional characteris- tics, featuring plots such as box plots, range plots, and histograms, with questions targeting mean, standard deviation, and other summary statistics. â˘TimeSequence: Sets that introduce a special- ized SHIFT function to encode temporal differ- ences along the time axis, evaluating modelsâ ability to track evolution and detect temporal trends. â˘Coordinates: Sets that capture geometric and spatial relationships among multiple subjects, testing coordinate-aware and positional rea- soning. These categories reflect distinct real-world scenar- ios and structural settings. To ensure fair compari- son, difficulty level, data volume, and question type are distributed independently and evenly across cat- egories. Robustness subset. To measure the stability of MLLMs under realistic visual degradation, we ap- ply four families of perturbation to images from the main track without altering the underlying answers or human readability: ⢠Noise: Image-wise noise, such as Gaussian noise, text-level noise, such as watermarks, and data-level noise, such as misleading back- ground text printed alongside the chart. ⢠Resolution: Degradation through blurring, rescaling, and lossy compression, PNG-to- JPEG transformation. â˘Style Transfer: Redrawing charts with a generative image model to emulate non- computer-generated styles, enabling assess- ment of whether performance shifts when charts depart from standard rendering tem- plates. By holding the questionâanswer pairs fixed, any performance variation in the robustness split can be directly attributed to the applied perturbation. 15 C Data Curation Theory The latent graph-based data dependencies is criti- cal to ensure the consistency and uniqueness of a multi-chart set. This section outlines the construc- tion process of long-chart VQA set, in which each questionâanswer pair is explicitly traceable to data nodes and visual evidence. C.1 Latent Dependency Graph The latent structure is represented by a directed acyclic graph (DAG), G = (V,E) whereVis the set of nodes andE â V Ă Vis the set of edges. The node set is partitioned into three types of nodes, V = D⪠Y ⪠I. Data nodes. The root layer consists of source data nodes, D =D 1 ,...,D n d , where eachD j denotes an independent and non- redundant raw data object. These nodes have no parents in the graph, so that Pa(D j ) = â , j = 1,...,n d . Computation nodes.The computation node con- sists of derived variables, Y =Y 1 ,...,Y n y , where each nodeY k is obtained as a deterministic mathematical transformation of its parent set: Y k = f k Pa(Y k ) , k = 1,...,n y . Here,f k denotes a deterministic function and Pa(Y k )is the set of immediate predecessors of Y k in G. Image nodes.The image node consists of visual representations, I =I 1 ,...,I n i , where each visualization node is generated by a mapping from data or derived variables into a vi- sual space: I â = Ď â Pa(I â );θ â , â = 1,...,n i . In this expression,Ď â is a visualization map andθ â denotes the specifies aesthetic parameters such as scale, layout, color, or geometry. C.2 Edges Mapping Information Flow Each directed edge in the network has an injectiv- ity indicator, which characterizes the information- preserving capacity of the corresponding transfor- mation. Letu,v â Vand suppose that an edge uâ vcorresponds to an information flow fromu into v. We distinguish the following cases. Bidirected Edges. An edge is bidirected, de- noted informally by uâ v, if the transformation is one-to-one and reversible. In this case,uandvare informationally equivalent, and the conditional entropies satisfy H(u| v) = 0and H(v | u) = 0. Directed Edges.An edge is directed, denoted by u inj ââ v, if the target nodevcan be transformed fromu, althoughvin reverse, cannot reconstructu. Equiv- alently, H(u| v) = 0. C.3 The Minimal Set and Subgraph Theory The latent graph design is introduced to satisfy the following objectives. First, all data are inter-related and consistent, ensuring that questions from any data node yield a unique answer. Second, the final set of images must collectively encode all informa- tion required to answer the complete set of ques- tions. To formalize these objectives, we introduce the following concept for dataset curation. Information equivalence classes.Define a rela- tionâź on V by declaring uâź v if and only if there exists a path fromutovcon- sisting only of bidirected transformations. Equiva- lently,uandvare informationally equivalent, writ- tenu⥠v, whenever each is recoverable from the other with zero information loss. In entropic terms, this implies H(u| v) = 0and H(v | u) = 0. The equivalence classes induced byâźpartition the vertex setVinto maximal subsets of mutually 16 information-equivalent nodes. We denote the class containing u by [u] =v â V : v âź u. Within such a class, any node serves as a perfect proxy for any other, regardless of whether it repre- sents raw data, a derived variable, or a visualization. Two nodes connected by bidirected edges form a trivial equivalence class. For example, an image node bijectively linked to a data node contains all information necessary to answer any question asso- ciated with the class. C.4 The Information Basis We seek a subset S â V that acts as a minimal generative set for the full framework. Intuitively,Sis an information basis if it contains exactly the information required to reconstruct every node in the network, while ex- cluding redundant or conflicting specifications. Formally,Sis required to satisfy the following properties. Completeness. The setSis complete if it spans the entire framework in the sense that every node in V is recoverable from S: H(V | S) = 0. Equivalently, the joint state of all nodes in the graph is fully determined once the elements ofSare known. Independence. The setSis independent if no element of S is informationally redundant relative to the others. That is, for every s i â S, H s i | S\s i > 0. Thus, removal of any element fromScauses a genuine loss of information. Conflict-free structure. The setSis conflict- free if it does not contain multiple nodes that en- code the same information that potentially cause conflicts in data. Two important forbidden patterns are: 1.selecting more than one node from the same information equivalence class; and 2.selecting both a parent node and a determin- istic descendant whose value is already fixed by that parent through the structural mapsf k or Ď â . The first introduces redundancy, while the second yields an overdetermined system vulnerable to in- consistency if supplied values deviate from deter- ministic transformations. Finally, a minimal generative set is a subset S â Vsatisfying completeness, independence, and conflict-freeness simultaneously. C.5 Long-Chart VQA Set Curation Process Building on the above theory, we design a struc- tured curation pipeline. The synthetic process be- gins by manually defining the latent dependency graph and uses algorithm to identify a minimal generative set within it. Data in this minimal set is then synthesized by LLM according to human- specified requirements. Subsequently, the rest of data or computation node on the the graph is propa- gated, which ensure all data and images are global consistent. After data curation is complete, questions are generated from data nodes, while images are pro- duced from visualization nodes in parallel. Finally, the process performs a backward propagation step to verify sufficiency to answer each question. An algorithm verify that curated images along collec- tively form a minimal information set, thereby guar- anteeing that all questions can be answered solely from the provided charts. Finally, a human verifi- cation will examine all questions are answerable, answer are accurate and images are aesthetically well-built. C.6 Global Data Structures The implementation is organized around a small collection of persistent objects. Graph specification. Each VQA-set contains human-specified graph structure initially. The graph is an abstracted dependency graph some real- life scenarios. The graph specification stores the node set, edge set, node types, structural equations, and attribute schemas: GraphSpec = (V,E,T ,E). Here,Trecords node types,Erecords the operator and information flow attributes between data nodes. 17 Minimal-set result. The algorithm of graph traversal provide the following results MinimalSetResult = (S â ,S alt ), whereS â is the unique selected minimal set,S alt is the set of admissible alternatives. Requirement Set. Requirements in data synthe- sis are necessary to ensure that the resulting base data is meaningful. When drafting the latent graph, requirements are specified by humans to guide the simulation of data in the minimal set. Formally, requirements are stored as RequirementSet =r 1 ,...,r n , where eachr j denotes a seperate text instruction that input to LLM at the data synthesis stage. Data Synthesis. Node data are generated based on the graph specification and the requirement set. For data nodes without parents, values are syn- thesized directly by the LLM according to the re- quirements: D d = f LLM (GraphSpec, RequirementSet). For computation nodes with parents, values are propagated using the predefined operators encoded in the edge connections: D c = f propagate (GraphSpec,D d ). Finally, all node data are collected into a uni- fied set, which is then passed in parallel to chart generation and QA generation: NodeData =D d ,D c . VQA Generation. Once data are synthesized, questionâanswer pairs are generated from the node data. Questions are instructed to be evenly dis- tributed across data nodes and computation nodes, ensuring a balanced difficulty profile across com- putational complexity: QA = f LLM (GraphSpec, NodeData). Candidate QA sets are then manually validated. In parallel, images are generated from the same node data: Images = f LLM (GraphSpec, NodeData). Final dataset package. The output of the full pipeline is Dataset = (GraphSpec, NodeData, Images, QA), containing the graph, generated data, images, se- lected question-answer pairs, and all meta informa- tion indicating the question characteristics. C.7 Stepwise Dataset Curation Pipeline Building on the curation techniques described above, we present the long-chart VQA curation pipeline, summarized in Algorithm 1. D Annotation and Quality Control We developed a dedicated annotation interface to support efficient QA construction and systematic quality control. The interface allows annotators to inspect candidate images, write and revise questionâ answer pairs, and verify whether each item is fully grounded in the provided visual evidence. This setup helps standardize the annotation workflow, reduce formatting inconsistencies, and make it eas- ier to trace potential errors during the verification stage. Figure 10: Annotation interface developed to annotate and verify the QA & images. E Examples of VQA Sets E.1 VQA Set Sample 1 This sample VQA set includes diverse image types such as pie charts, line charts, bar charts, and tables. It represents a relatively complex case, containing more than 200 data points. E.2 VQA Set Sample 2 This sample VQA set features stacked bar charts, histograms, range plots, and box plots, illustrating varied chart structures. 18 Algorithm 1 Pseudocode for the VQA Set Curation Pipeline Require: Specifed latent graph GraphSpec, simulation requirementsR. Ensure: Long-Chart VQA Curation 1: MinimalSetâ f traversal (GraphSpec) 2: DataMinimalSetâ f LLM (GraphSpec,R) 3: NodeDataâ f propagate (GraphSpec, DataMinimalSet) 4: QACandidatesâ f LLM (GraphSpec, NodeData) 5: Imagesâ (RenderImages)(GraphSpec, NodeData) 6: DependencyGraphâ UPDATEGRAPH(GraphSpec, QACandidates, Images) ⡠Verify images can sufficiently support each question; assign structural difficulty hop(q) 7: VerifiedVQAâ HUMANVERIFY(GraphSpec, NodeData, Images, QACandidates)⡠Human- in-the-loop: confirm readability, answer uniqueness, coherent easy-to-hard path, and analytical validity E.3 VQA Set Sample 3 This sample demonstrates image style transfer. Line plots and tables are rendered in a whiteboard style while preserving the readability of all data and information. E.4 VQA Set Sample 4 This sample illustrates text noise injection. Water- marks are applied to all images while maintaining readability, ensuring that the underlying data and information remain accessible. F Evaluation Metrics F.1 Accuracy The accuracy metric measures the proportion of questions answered correctly: Accuracy = P N i=1 I Ëa i = a i N , whereNis the total number of questions,Ëa i is the predicted answer for thei-th question,a i is the ground-truth answer, andI(¡)is the indicator function that returns 1 when the prediction matches the ground truth and 0 otherwise. The exact matching criterion depends on the question type: â˘Multiple-choice questions. An answer is re- garded as correct only if the selected option exactly matches the ground-truth option (e.g., A, B, C, or D). â˘Numerical questions. An answer is regarded as correct if the relative error is within 1% of the standard answer, i.e., |Ëa i âa i | |a i | ⤠0.01. F.2 mDIoU: Modified Distance IoU for Grounding Boxes For evaluating bounding-box grounding quality, we adopt a distance-aware intersection-over-union variant. Let Ë Bdenote the predicted bounding box andBthe ground-truth box. The modified Distance IoU (mDIoU) is defined as mDIoU( Ë B,B) = IoU( Ë B,B)â Ď 2 c Ë B ,c B d 2 , whereIoU( Ë B,B)is the standard intersection-over- union of the two boxes,c Ë B andc B are the cen- ter points of the predicted and ground-truth boxes, Ď(¡,¡)is the Euclidean distance between centers, anddis the diagonal length of the smallest enclos- ing box that contains both Ë BandB. The second term penalizes predictions that are correctly sized but spatially displaced. We choose DIoU over plain IoU because model- generated bounding boxes are often much larger than the ground-truth regions; in such cases the in- tersection area is small relative to the union, yield- ing a near-zero IoU that is hard to discriminate across models. By adding a center-distance penalty, DIoU rewards predictions whose centers are close to the ground-truth center even when the box sizes differ, making the metric more comparable across systems. The metric is implemented from scratch. F.3 Chamfer Distance for Point Sets For evaluating point-level grounding, we measure the bidirectional Chamfer distance between the pre- dicted point set Ë P =Ëp 1 ,..., Ëp m and the ground- truth point setP =p 1 ,...,p n : CD( Ë P,P) = 1 m X Ëpâ Ë P min pâP âĽËpâp⼠2 + 1 n X pâP min Ëpâ Ë P âĽËpâp⼠2 . 19 (a) Sample VQA set 1. (b) Sample VQA set 2. (c) Sample VQA set 3. (d) Sample VQA set 4. Figure 11: Examples of VQA sets covering mixed chart types, statistical views, style transfer, and text-noise perturbation. These figures may need to enlarge to read. The first term averages the nearest-neighbor dis- tance from each predicted point to the ground- truth set, and the second term averages the nearest- neighbor distance from each ground-truth point to the predicted set. This symmetric formulation pe- nalizes both false-positive predictions and missed ground-truth points. The metric is implemented from scratch. G Baseline Configurations Table 10: main context ModelRelease Date Max Tokens Temperature Tok-P doubao-seed-2.0-lite2026-02-251000001.0 doubao-seed-2.0-pro2026-02-251000001.0 gemini-3.1-flash-lite2026-02-201000001.0 gpt-5.4-mini2026-03-181000001.0 gpt-5.4-nano2026-03-181000001.0 qwen3.5-flash2026-03-041000001.0 gemma-4-26b-a4b-it2026-04-041000001.0 gemma-4-31b-it2026-04-041000001.0 mimo-v2.52026-04-231000001.0 qwen3.5-35b-a3b2026-03-041000001.0 claude-sonnet-4-52025-09-301000001.0 20 H Prompts H.1 Baseline Assessment Prompts Prompts for baseline assessment Please carefully examine the images provided above. They contain charts, tables, or data visualizations. Based on the information shown in these images, answer the following questions one by one. For each question, respond with a JSON object containing âLLM_answerâ. The final answer: a number, A/B/C/D, or True/False and âreasoningâ is a brief explanation of your reasoning. Input charts and provides the following per-question prompt, appended sequentially for each question: Question idx: question_text Options:options_text Please answer in the following JSON format exactly: "LLM_answer_initial": "your initial answer here (a number, A/B/C/D, or True/False)", "reasoning": "brief explanation of how you arrived at this answer", "LLM_answer_final": "your final answer here (a number, A/B/C/D, or True/False)" Prompts for the retrieval mode Please carefully examine the images provided above. They contain charts, tables, or data visualizations.Your first task is to reconstruct ALL the underlying data shown in these images as completely and accurately as possible. For each chart or visualization, extract: ⢠The chart title ⢠All axis labels, categories, and legends ⢠All numerical values with their corresponding row/column/category labels Present the reconstructed data in a clear structured format (JSON or markdown table). This reconstruction will be used to answer subsequent questions. Prompts for the grounding mode You are answering questions about chart images. For each question, provide your best answer AND identify the rectangular regions in each image that support your answer. Coordinate convention: ⢠Use pixel coordinates with top-left origin: (0, 0) is the top-left pixel. ⢠x increases rightward; y increases downward. ⢠Each bbox is [x_min, y_min, x_max, y_max] as integers. ⢠Legal ranges: x in [0, width_px]; y in [0, height_px]. Image labels: Each image below is labeled with its ID and title. Use the exact image_id in your output. image_descriptions Instructions: ⢠First, give your initial answer and the bounding boxes that support it. ⢠Then explain your reasoning. ⢠Finally, give your final answer and revised bounding boxes (they may be the same as initial). If an image is NOT relevant to the answer, you MUST include it with an empty bbox list: "bboxes": []. Do not omit any image. Respond in this exact JSON format: "initial_answer": "your initial answer (number, A/B/C/D, or True/False)", "initial_bbox": "example_image_id": "bboxes": [[x_min, y_min, x_max, y_max], ...] , "reasoning": "brief explanation of how you arrived at this answer and why you selected these regions", "final_answer": "your final answer (number, A/B/C/D, or True/False)", "final_bbox": "example_image_id": "bboxes": [[x_min, y_min, x_max, y_max], ...] 21 Prompts for the point mode You are answering questions about chart images. For each question, provide your best answer AND identify the exact point locations in each image that support your answer. Coordinate convention: ⢠Use pixel coordinates with top-left origin: (0, 0) is the top-left pixel. ⢠x increases rightward; y increases downward. ⢠Each point is [x, y] as integers. ⢠Legal ranges: x in [0, width_px]; y in [0, height_px]. Image labels: Each image below is labeled with its ID and title. Use the exact image_id in your output. image_descriptions Instructions: ⢠First, give your initial answer and the point coordinates that support it. ⢠Then explain your reasoning. ⢠Finally, give your final answer and revised point coordinates (they may be the same as initial). For each relevant data value, provide ONE point coordinate that best represents its location in the image. For bar charts, this is typically the center or top of the bar. For line charts, this is the data point marker. For pie charts, this is the center of the slice. If an image is NOT relevant to the answer, you MUST include it with an empty point list: "points": []. Do not omit any image. Respond in this exact JSON format: "initial_answer": "your initial answer (number, A/B/C/D, or True/False)", "initial_points": "example_image_id": "points": [[x, y], ...] , "reasoning": "brief explanation of how you arrived at this answer and why you selected these points", "final_answer": "your final answer (number, A/B/C/D, or True/False)", "final_points": "example_image_id": "points": [[x, y], ...] 22