Chart-RVR: Robust Chart Verification

• Chart-RVR is a set of techniques that verify chart reasoning using explicit logical, structural, and visual cues rather than relying solely on OCR or superficial patterns.
• It incorporates methods such as reinforcement learning with verifiable rewards and rendering-based evaluations to improve accuracy in chart question answering and chart-to-code generation.
• The framework emphasizes robustness through formal logic verification, perturbation-based stress tests, and multi-chart integration to ensure evidence-level grounding of chart elements.

Chart-RVR is a label used across recent chart-understanding research for a closely related set of problems and training paradigms: robust chart reasoning and verification on real or perturbed charts, reinforcement learning with verifiable or visual rewards for chart question answering, rendering-based reward design for chart-to-code generation, and evidence-level grounding of referring expressions over chart elements (Ahmed et al., 2023, Masry et al., 13 Aug 2025, Sinha et al., 13 Oct 2025, Tan et al., 25 Aug 2025, Niu et al., 8 May 2026). Across these usages, the common objective is to move beyond shortcut behavior based on OCR, template bias, or superficial style regularities, and instead optimize against signals that are machine-checkable at the level of chart structure, underlying data, rendered outputs, or localized visual evidence.

1. Scope and terminology

Across the cited literature, Chart-RVR is not a single benchmark or algorithmic artifact. It denotes a family of chart-centered verification regimes in which correctness is established through explicit logical, structural, visual, or executable criteria rather than by unconstrained text generation alone. RealCQA frames Chart-RVR as robust chart visual question answering under distribution shift, with real scientific charts and first-order-logic-style templates (Ahmed et al., 2023). BigCharts-R1 uses the term for “chart reasoning via visual reinforcement,” namely supervised fine-tuning followed by GRPO-based reinforcement learning with chart-specific verifiable rewards (Masry et al., 13 Aug 2025). The later “Chart-RVR” framework specializes this idea to explainable chart reasoning with rewards for chart type, table reconstruction, and process conformity (Sinha et al., 13 Oct 2025). ChartMaster, by contrast, uses the closely related phrase “Chart-RVR: Rendering-based Visual Rewards” for chart-to-code generation, where sampled code is executed and the resulting chart image is compared with the reference chart (Tan et al., 25 Aug 2025). ChartREG++ extends the notion to chart referring visual grounding/reasoning, where the task is to localize the exact chart elements denoted by a free-form expression as points, boxes, or masks (Niu et al., 8 May 2026).

Usage of Chart-RVR	Core task	Representative work
Robust chart reasoning and verification	QA on real, perturbed, or logically structured charts	RealCQA (Ahmed et al., 2023), RealCQA-V2 (Ahmed et al., 2024), Socratic Chart (Ji et al., 14 Apr 2025)
Reinforcement learning with verifiable rewards	Post-training LVLMs/VLMs for chart QA with GRPO-style objectives	BigCharts-R1 (Masry et al., 13 Aug 2025), Chart-RVR (Sinha et al., 13 Oct 2025), Chart-RL (Zhang et al., 7 Mar 2026)
Rendering- or grounding-based verification	Code execution, image comparison, or evidence-level localization	ChartMaster (Tan et al., 25 Aug 2025), ChartREG++ (Niu et al., 8 May 2026)

This spread of meanings suggests a unifying methodological commitment: chart competence should be judged against chart-grounded evidence, not only against surface-form answers.

2. Real-world robustness, formal logic, and perturbation-based stress tests

RealCQA established a real-chart benchmark built from scientific literature and positioned it as a robust test-bed for chart visual QA and formal logic verification (Ahmed et al., 2023). It uses real chart images and annotations from the ICPR 2022 CHART-Infographics challenge, includes bar, line, point, scatter, and box plots, and composes question-answer pairs by exhaustively instantiating 240 templates over closed sets of chart objects. Its test set contains 9,357 chart images, while total QA-pair counts vary by sampling regime from 203,735 to 367,139. The taxonomy extends the DVQA structure/retrieval/reasoning scheme and formalizes queries as first-order logic over chart objects, including argmax, threshold filtering, quantified statements, aggregations, comparative reasoning, and ranked or unranked list outputs. Numeric answers are evaluated with PlotQA-D’s $\pm 5\%$ tolerance; ranked lists use nDCG@K; unranked lists use exact-set correctness or set-level F1. A central result is that models pretrained on synthetic charts drop sharply on these real charts because of OCR imperfections, complex layouts, ambiguous legends, non-uniform scales, and visually cluttered scientific styles.

RealCQA-V2 reframes chart QA as Visual Premise Proving, decomposing each question into a sequence of binary premises spanning structural premises, data premises, reasoning premises, and mathematical premises (Ahmed et al., 2024). The dataset comprises over 10 million text descriptors or premises across real scientific charts from PubMed Central and RealCQA; the reported experimental counts include roughly 2,156,628 train premises and 548,564 test premises for true cases, with false cases generated at a 1:3 true:false ratio. Sequence-level verification is formalized by the strict metric

$$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$

which counts a sequence as correct only if all premises are correctly proven. The companion metric DCP measures how far an incorrect chain progresses before failing. In zero-shot evaluation with MATCHA on scientific chart QA, reasoning performance was reported at approximately $27\%$, compared with $19\%$ for structure and $14\%$ for data retrieval, raising the question of whether end-task accuracy may conceal weak visual grounding.

Socratic Chart attacks a related misconception: that strong chart QA scores necessarily reflect robust visual understanding (Ji et al., 14 Apr 2025). It introduces two perturbation-based stress tests on ChartQA. In ChartQA-RL, textual labels are removed; in ChartQA-HV, charts are stretched horizontally or vertically. Under ChartQA-RL, GPT-4o drops to 50.0 overall relaxed accuracy with a 35.7% drop, Gemini-2.0 Pro drops to 37.2 with a 50.0% drop, and Socratic Chart (multi-agent) reaches 57.6 with a 23.9% drop. Under ChartQA-HV, GPT-4o reaches 71.5 with a 14.2% drop, Gemini-2.0 Pro 52.6 with a 34.6% drop, and Socratic Chart (multi-agent) 68.1 with a 13.4% drop. The framework converts raster charts into SVG, extracts primitives such as <rect>, <path>, <line>, and <text> through specialized agents, and uses an agent-critic to merge candidate fragments into a structured symbolic representation. The perturbation results make explicit that OCR- and layout-based shortcuts remain a major confound in chart reasoning.

3. Cross-chart integration and evidence-level grounding

InterChart extends Chart-RVR from single-chart reasoning to distributed reasoning across 2–3 related charts (Iyengar et al., 11 Aug 2025). The benchmark contains 5,214 validated QA pairs across 1,012 multi-chart contexts and 2,706 unique chart images, organized into three tiers: DECAF for decomposed elementary single-chart facts, SPECTRA for synthetically aligned cross-chart reasoning, and STORM for real-world paired line charts drawn from Our World in Data. DECAF contains 2,809 QA pairs over 1,188 decomposed single-variable charts; SPECTRA contains 1,717 QA pairs over 333 context sets and 870 unique charts; STORM contains 768 QA pairs over 324 chart pairs. Evaluation uses LLM-as-a-judge majority voting rather than exact-string matching. The strongest reported model, Gemini 1.5 Pro, still falls from 65.2% on DECAF zero-shot to 59.1% on SPECTRA and 34.8% on STORM, while open models decline more sharply. Decomposition improves per-chart reasoning, but table intermediates hurt on STORM, indicating that cross-chart semantic alignment, temporal synchronization, and unit normalization are not preserved by per-chart tabularization alone.

ChartREG++ shifts verification from answer generation to direct localization of chart evidence (Niu et al., 8 May 2026). It defines chart referring visual grounding/reasoning as mapping a free-form expression and chart image to the exact set of chart elements denoted by the expression, returning points, boxes, or pixel-accurate masks. The benchmark has 3.4k samples over 850 charts, is split roughly 50/50 between ChartMimic and ECDBench, covers 18 element types and multiple granularities, and reports an average of 9.7 targets per query. Expressions use data features, textual/localization features, visual features, and hybrid cues. Correctness is format-specific: point predictions must lie inside the instance mask or within 5 pixels, boxes require IoU $\ge 0.5$, and masks require IoU $\ge 0.5$ or Boundary-IoU for thin-line elements, with Hungarian matching for multi-target cases. A code-driven Matplotlib Artist-tracing pipeline generates pixel-accurate masks, and a Mask2Former candidate generator reaches 51.59 mAP, far above GroundingSAM2’s 0.58 and SAM3-Detection’s 10.25. In two-stage mask grounding, the authors’ Mask2Former-plus-SoM setup with QwenVL-32B reports F1 values of 43.82 on ECDBench and 43.79 on ChartMimic using all candidates, or 56.77 and 56.73 with gold-category candidates. The benchmark therefore operationalizes Chart-RVR as verifiable evidence recovery rather than solely as textual QA.

4. Reinforcement learning with verifiable rewards for chart question answering

BigCharts-R1 formulates Chart-RVR as “visual reinforcement finetuning” for chart reasoning (Masry et al., 13 Aug 2025). Starting from Qwen2.5-VL-Instruct 3B and 7B models, it performs supervised fine-tuning on chart chain-of-thought data and then applies GRPO-based reinforcement learning with two chart-specific rewards: a Chart Error Rate Reward for numeric answers and a response format reward. For numeric answers, the error rate is

$\mathrm{ER}(\hat y, y) = \frac{|\hat y - y|}{|y|},$

and the reward is $R_{\mathrm{CERM}} = 1/(1+\mathrm{ER})$; for non-numeric answers, exact match is used. The total reward is $$R_{\mathrm{total}} = R_{\mathrm{CERM}} + R_{\mathrm{Fmt}}$$. The underlying BigCharts dataset is built from 245,414 chart images collected from existing datasets, Common Crawl, and Google Search, then filtered through “chart re-plotting” to yield 134,950 charts with accurate underlying data and around 1.8 million questions. GRPO training uses $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$0 sampled responses per prompt. BigCharts-R1-7B reports an overall average of 74.48 across FigureQA-Sub, DVQA-Sub, PlotQA-Sub, ChartQA, and CharXiv, exceeding several larger open and closed models on that consolidated suite.

The later framework explicitly titled “Chart-RVR” introduces a more structured explainability-oriented reward stack for LVLMs (Sinha et al., 13 Oct 2025). The base model emits a > block containing <type>, <table>, and rationale steps, followed by an <answer> block. Rewards combine schema checks, surrogate perception tasks, and process conformity: chart-type exact match, faithful chart-table reconstruction, and stepwise alignment of the rationale with an oracle process skeleton using MiniLM-L6-v2 embeddings. The total reward is

$$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$1

with $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$2 and $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$3. Training uses TRL GRPO on Qwen2.5VL-3B-Instruct with prompt length 4096, completion length 768, $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$4 rollouts, 4 epochs, FP16, and 4× H100 GPUs. On six benchmarks, Chart-RVR-3B reports 84.56 on ChartQA, 78.68 on PlotQA, 77.62 on ChartFC, 53.36 on EvoChart, 28.38 on ChartQAPro, and 68.32 on ChartBench; the “Hard” variant improves these to 85.76, 77.90, 80.07, 54.24, 28.64, and 69.46. Relative to SFT, OOD gains are especially large: +7.28 on EvoChart, +4.82 on ChartQAPro, and +3.68 on ChartBench. The framework also reports improved rationale fidelity via positive $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$5 on several OOD datasets.

Chart-RL generalizes the verifiable-reward approach to chart comprehension benchmarks such as MultiChartQA, ChartInsights, and RobustCQA (Zhang et al., 7 Mar 2026). It uses Qwen2.5-VL-3B-Instruct with LoRA of rank 64, scaling $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$6, dropout 0.05, unfrozen vision modules, 8× H100 GPUs, and GRPO with $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$7 samples per prompt. The reward is the weighted sum of a binary final-answer accuracy term and a binary format reward enforcing a <thinking> block plus a JSON answer inside <answer>. The paper emphasizes that task difficulty is more important than data volume: training on 448 hard CharXiv examples outperforms training on 6,200 easy PlotQA examples, and even 10 hard examples beat models trained on over 6,000 simple examples. Reported gains over baseline are +16.7% on MultiChartQA and +11.5% on ChartInsights, with improved performance in 18 of 25 RobustCQA perturbation categories and a +55.6% relative transfer gain on MathVerse Vision.

5. Rendering-based visual rewards in chart-to-code generation

ChartMaster adapts Chart-RVR to chart-to-code generation by making the reward depend on the rendered chart rather than on token-level similarity alone (Tan et al., 25 Aug 2025). The task is: given a chart image $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$8 and an instruction $$Acc_{VPP} = \frac{1}{S}\sum_{s=1}^{S}\left(\prod_{p=1}^{P_s} C_{s,p}\right),$$9, generate executable plotting code $27\%$0 whose rendered chart $27\%$1 visually and semantically matches $27\%$2. Supervised fine-tuning is defined by

$27\%$3

but the paper argues that SFT alone often fails to maintain visual consistency. After SFT on ReChartPrompt-240K, the model undergoes ChartSimRL, a GRPO-based RL stage in which sampled code is executed in Python matplotlib and compared with the original chart image. For each training example, $27\%$4 candidate codes are sampled; non-executable outputs receive zero reward. Executable outputs are scored by

$27\%$5

where the attribute term is a Jaccard similarity over extracted chart attributes,

$27\%$6

and the visual term averages cosine similarities across four blocks of a pretrained ResNet-18 feature extractor.

The data pipeline is equally central. ReChartPrompt begins from 30,071 arXiv papers, downloads LaTeX sources and images, uses Qwen2.5-VL-72B to classify images into 12 chart categories, and then prompts the same model with 20 diverse instructions to produce Python matplotlib code. From 288,992 extracted images, execution filtering yields 242,479 triplets, forming ReChartPrompt-240K. Data generation uses vLLM with top-p = 0.9 and temperature = 0.1. The backbone model is Qwen2.5-VL-7B; SFT runs for 1 epoch on the full dataset with learning rate $27\%$7 and batch size 128, while RL uses 10% of the dataset with learning rate $27\%$8, batch size 128, and sampling temperature 1.0, top-p 1.0, top-k 80.

The reported gains separate the contribution of data diversity from that of rendering-based rewards. On ChartMimic, the raw Qwen2.5-VL-7B baseline reports 65.5 execution, 39.9 low-level, and 34.2 high-level. Adding ReChartPrompt alone yields 91.1, 73.7, and 73.3. Adding ChartSimRL further improves these to 93.8, 78.2, and 77.3. On Plot2Code, the same progression is 67.4/43.8/4.60 to 80.3/59.3/5.34 to 88.2/62.6/5.65, and on ChartX from 2.18 to 2.36 to 2.46. The full ChartMaster-7B system attains 93.8 execution on ChartMimic and 88.2 pass rate on Plot2Code, rivaling GPT-4o on several benchmark dimensions while remaining a 7B-parameter model. Reward ablations further show that attribute-only and visual-only rewards underperform the combined reward, and that Jaccard plus ResNet-18 features outperform several metric alternatives including precision, recall, F1, MSE, SSIM, PSNR, AlexNet, VGG, and MLLM-based scoring.

6. Limitations, recurrent failure modes, and open directions

A recurrent finding across Chart-RVR work is that high answer accuracy does not guarantee chart-grounded reasoning. RealCQA shows large out-of-distribution drops when models trained on synthetic charts are evaluated on real scientific charts, especially for binary first-order-logic questions and list outputs (Ahmed et al., 2023). RealCQA-V2 sharpens this diagnosis by showing that zero-shot reasoning performance can exceed structure and data-retrieval performance, implying that models may solve some questions through generic mathematical or linguistic priors rather than through reliable chart parsing (Ahmed et al., 2024). Socratic Chart reaches the same conclusion from a different angle: removing labels or perturbing aspect ratios causes large drops for GPT-4v, GPT-4o, and Gemini, which indicates reliance on OCR and positional regularities (Ji et al., 14 Apr 2025).

Another common limitation is that verification is itself modality-specific and brittle. In ChartMaster, renderer sensitivity to matplotlib versions, fonts, DPI, or antialiasing can affect the visual reward; attribute extraction errors can corrupt Jaccard similarity; and non-executable outputs receive $27\%$9, which provides no gradient signal for near-correct code with minor runtime issues (Tan et al., 25 Aug 2025). In the explainable Chart-RVR framework, the reward depends on ground-truth chart types, chart tables, and oracle-generated rationale skeletons; regex- and JSON-based checks can be brittle, and extending beyond the 10 canonical chart families requires updating the verification schema (Sinha et al., 13 Oct 2025). BigCharts-R1 reports that RL helps reasoning-heavy tasks but does not help the CharXiv descriptive subset and may slightly decrease performance there, which indicates that numeric reward design does not automatically improve retrieval-like behavior (Masry et al., 13 Aug 2025). Chart-RL likewise requires deterministic, mathematically verifiable targets and observes regressions on two RobustCQA perturbation categories, log_scale and scaling (Zhang et al., 7 Mar 2026).

The grounding and multi-chart literature exposes additional bottlenecks. InterChart shows that decomposition and structured intermediates can raise single-chart performance, yet cross-chart integration remains weak on real-world STORM pairs because temporal alignment, unit mismatches, irregular axes, and semantic glue across charts are not preserved by isolated table conversion (Iyengar et al., 11 Aug 2025). ChartREG++ reports that data-driven clues remain harder than perceptual cues, thin and small primitives remain difficult even with boundary-aware metrics, performance drops as the number of referred targets grows, and oracle-candidate experiments reveal substantial remaining headroom in language-mask alignment (Niu et al., 8 May 2026).

These limitations have already shaped the stated research agenda. RealCQA points toward stronger chart parsers, layout models, and neuro-symbolic computation (Ahmed et al., 2023). RealCQA-V2 proposes premise-level verification as a way to curb hallucinations and inspect reasoning depth (Ahmed et al., 2024). InterChart emphasizes explicit alignment and aggregation modules for multi-chart reasoning (Iyengar et al., 11 Aug 2025). ChartREG++ argues for stronger line-aware and small-object segmentation and better chain-of-grounding (Niu et al., 8 May 2026). BigCharts-R1 suggests structure-aware rewards and broader tasks such as summarization and fact-checking (Masry et al., 13 Aug 2025). Chart-RVR and Chart-RL both indicate that the next stage of progress will likely depend less on larger generic LVLMs than on verifiers that are specific to chart structure, executable semantics, and evidence-level faithfulness (Sinha et al., 13 Oct 2025, Zhang et al., 7 Mar 2026).