BigCharts-R1: Enhanced Chart Reasoning

• BigCharts-R1 is a chart-domain vision-language model that improves chart comprehension by preserving real-world visual diversity and data fidelity.
• It employs a two-stage training framework that integrates supervised fine-tuning with Group Relative Policy Optimization to enhance multi-step chart reasoning.
• Evaluations demonstrate robust performance with notable out-of-distribution gains and improved accuracy on synthetic and real-world chart benchmarks.

Searching arXiv for the specified paper and closely related chart-reasoning work. arXiv search query: (Masry et al., 13 Aug 2025) BigCharts-R1 Enhanced Chart Reasoning with Visual Reinforcement Finetuning BigCharts-R1 is a chart-domain vision-LLM and training framework introduced in “BigCharts-R1: Enhanced Chart Reasoning with Visual Reinforcement Finetuning” (Masry et al., 13 Aug 2025). It is designed to improve chart comprehension under conditions where existing vision-LLMs continue to struggle because they are trained on datasets that lack diversity and real-world authenticity, or on automatically extracted underlying data tables of charts, which can contain numerous estimation errors. The system couples a dataset creation pipeline, BigCharts, with a two-stage optimization procedure that integrates supervised fine-tuning and Group Relative Policy Optimization (GRPO)-based reinforcement learning. Its central design goal is to preserve real-world visual diversity while retaining accurate underlying data through a replotting process, then exploit chart-specific reward signals to improve robustness and generalization across chart styles and domains (Masry et al., 13 Aug 2025).

1. Problem Setting and Design Rationale

Charts are described as essential to data analysis because they transform raw data into visual representations that support human decision-making (Masry et al., 13 Aug 2025). Within this setting, BigCharts-R1 targets chart question answering and chart reasoning, including direct retrieval, visual layout recognition, arithmetic operations, and multi-step reasoning. The motivating diagnosis is twofold: first, many existing chart datasets are either synthetic and visually homogeneous or derived from noisy chart-to-table extraction; second, prior training practice is characterized as relying only on supervised fine-tuning using these low-quality datasets (Masry et al., 13 Aug 2025).

The BigCharts-R1 formulation addresses these issues by tying data fidelity to executable plotting code. The paper defines the replotting process as

$$I \xrightarrow{\;f_{\rm VLM}\;} c \quad,\quad I' = g_{\rm render}(c),$$

where $I$ is the original chart image, $f_{\rm VLM}$ maps $I$ to plotting code $c$, and $g_{\rm render}$ deterministically renders $c$ into a new image $I'$ (Masry et al., 13 Aug 2025). This structure makes the replotted chart the canonical training object: it inherits visual cues from the source chart while attaching exact data through executable code.

A plausible implication is that the method treats chart reasoning as a joint visual-and-programmatic supervision problem rather than as a purely image-to-text mapping problem. That interpretation is consistent with the paper’s emphasis on visual diversity plus data fidelity, and on the use of chart-specific rewards during reinforcement learning (Masry et al., 13 Aug 2025).

2. BigCharts Dataset Creation Pipeline

The BigCharts dataset creation pipeline begins with chart collection from three sources. The reported sources are existing academic or synthetic benchmarks—ChartGemma, FigureQA, DVQA, PlotQA, and ArXivQA—contributing 174 K images; Common Crawl via Mint-1T PDFs with a ResNet-50 two-stage filter, contributing 57 K additional charts; and targeted Google Image Search over 210 chart-style keywords, contributing 14 K charts. The total collected volume is 245,414 raw chart images (Masry et al., 13 Aug 2025).

The replotting stage uses Gemini Flash 2.0 to recover plotting code in matplotlib, plotly, or Chart.js form from the original chart image, after which deterministic rendering produces the new chart image. Charts for which code failed to render are discarded, yielding 134,950 replotted charts with guaranteed ground-truth data (Masry et al., 13 Aug 2025). The stated purpose is to combine authentic real-world appearance with exact underlying values.

Question and chain-of-thought generation are then performed on each replotted chart plus its data. Gemini Flash 2.0 produces 16 QA pairs per chart, covering direct data retrieval, visual layout questions such as “Which bar is tallest?”, arithmetic and multi-step reasoning including max, min, sum, ratio, and average, unanswerable questions marked “Not Applicable,” and multiple-choice or short dialog contexts (Masry et al., 13 Aug 2025). Each QA is paired with a chain-of-thought of variable length, with median 39 tokens and maximum approximately 1,933 tokens.

The resulting dataset statistics reported in the paper are summarized below.

Component	Reported value
Total raw chart images	245,414
Replotted charts	134,950
QA pairs	$\sim 1.8$ M

The chart-type distribution is described as dominated by bar and line charts while also including scatter, pie, heatmap, violin, dashboards, and other forms. Topics are clustered into 20 domains, with examples including Finance at 39%, STEM at 29.7%, and Data Science at 27.5%. The answer distribution includes approximately 342 K textual answers, approximately 141 K yes/no answers, and approximately 118 K multiple-choice answers (Masry et al., 13 Aug 2025).

This pipeline differs from the programmatic synthesis strategy of Chart-R1, which starts from real-world tabular data tables extracted from arXiv papers and manually written Matplotlib examples, then synthesizes chart reasoning data with step-by-step CoTs (Chen et al., 21 Jul 2025). BigCharts-R1 instead begins from collected chart images and reconstructs code from them. This suggests that the two systems occupy adjacent but distinct points in the design space of chart-reasoning data generation.

3. Model Architecture and Two-Stage Optimization

BigCharts-R1 uses Qwen2.5-VL-Instruct vision-language backbones at 3 B and 7 B parameters (Masry et al., 13 Aug 2025). The architecture is described at a high level as a standard vision-language stack with a ViT-style vision encoder based on image patch embeddings and a transformer decoder on the language side with cross-attention to visual features. The paper does not introduce a new backbone; its contribution lies in the data pipeline and training framework.

The first training stage is supervised fine-tuning. The SFT stage uses 1.8 M BigCharts samples with CoTs and optimizes cross-entropy on token sequences to encourage step-by-step reasoning (Masry et al., 13 Aug 2025). The reported hyperparameters for both 3 B and 7 B settings are 1 epoch, learning rate $2\times 10^{-5}$, batch size 32, warmup 0.1, and cosine decay. The explicit use of CoTs places BigCharts-R1 in the same broader family as Chart-R1’s Chart-COT stage, which also trains on step-by-step decompositions before reinforcement learning (Chen et al., 21 Jul 2025).

The second stage applies reinforcement learning through the RL with Verifiable Rewards framework: $I$0 where $I$1 is the post-SFT reference policy (Masry et al., 13 Aug 2025). The RL stage uses 32 K verified QA samples, consisting of 22 K from ChartQA and 10 K from PlotQA, DVQA, and FigureQA, with hyperparameters 1 epoch, batch size 8, learning rate $I$2, and 8 candidates per sample (Masry et al., 13 Aug 2025).

The overall training logic is therefore cumulative rather than substitutive: SFT builds a reasoning prior from large-scale CoT supervision, and GRPO subsequently optimizes answer accuracy and response conformity under verifiable reward functions. The paper’s framing is that this combination strengthens robustness and out-of-distribution generalization (Masry et al., 13 Aug 2025).

4. GRPO and Chart-Specific Reward Design

BigCharts-R1 follows GRPO and avoids a separate value network by sampling $I$3 candidate responses $I$4 from $I$5 and computing normalized advantages

$I$6

The update step clips policy ratios based on group advantages, with the paper referring to Shao et al. 2024 for the GRPO mechanism (Masry et al., 13 Aug 2025).

Two chart-specific rewards are combined. The first is the Chart Error Rate Reward, or CERM, which is defined for exact numeric accuracy as

$I$7

The second is the Response Format Reward,

$I$8

The total reward is

$I$9

These formulas make numerical precision a first-class optimization target while also explicitly scoring output formatting (Masry et al., 13 Aug 2025).

Relative to Chart-R1, which uses a soft numerical reward with $f_{\rm VLM}$0 tolerance and a format reward tied to the presence of > and <answer> tags, BigCharts-R1’s reward definition is more directly phrased in terms of chart error rate and requested-answer format (Chen et al., 21 Jul 2025). A plausible implication is that BigCharts-R1 is tuned toward exact or near-exact answer production under heterogeneous benchmark protocols rather than only toward coarse numeric acceptability.

5. Benchmarks, Main Results, and Ablations

The evaluation suite includes synthetic subsets—FigureQA-Sub, DVQA-Sub, and PlotQA-Sub, each with 1 K samples—together with real-world benchmarks ChartQA and CharXiv (Masry et al., 13 Aug 2025). Metrics vary by benchmark: exact accuracy is used for FigureQA-Sub and DVQA-Sub, “relaxed” accuracy for PlotQA-Sub and ChartQA, and LLM-scored accuracy for CharXiv (Masry et al., 13 Aug 2025).

For BigCharts-R1-3B with SFT and RL, the paper reports an average of approximately 72.1%. The benchmark-specific values are FigureQA-Sub Val1 80.1% and Val2 81.0%; DVQA-Sub ValE 81.2% and ValH 80.6%; PlotQA-Sub T1 78.5% and T2 59.9%; ChartQA average 88.16% over augmented and human splits; and CharXiv Reasoning 37.4% and Description 62.4% (Masry et al., 13 Aug 2025). BigCharts-R1-7B is reported at approximately 74.5% average. The paper further states that notable closed-source baselines include GPT-4o at 61.2% average and Gemini-Flash at 59.3%, while the best open-source model listed is Intern-VL2.5-8B at 70.4% average (Masry et al., 13 Aug 2025).

The ablation studies isolate both data and optimization effects. On the same 3 B backbone, SFT on BigCharts attains 68.7% average, compared with 48.8% for TinyChart and 58.1% for ChartGemma (Masry et al., 13 Aug 2025). A separate comparison between original real-chart images and replotted charts with code shows that replotted data gives a gain of 2.5 percentage points overall (Masry et al., 13 Aug 2025). These results directly support the argument that replotting improves supervision quality.

The paper also reports an out-of-distribution generalization study in which training is conducted only on ChartQA and testing is performed on PlotQA-Sub and DVQA-Sub. In this setting, learning curves show that the GRPO-trained model outperforms pure SFT by approximately 5–8 percentage points on out-of-distribution splits (Masry et al., 13 Aug 2025). This is a central empirical claim of the BigCharts-R1 framework, because it ties RL not merely to in-domain optimization but to improved transfer.

6. Relation to Adjacent Work and Reported Limitations

BigCharts-R1 belongs to a 2025 line of chart-reasoning systems that combine CoT supervision with RL-style post-training. Chart-R1, for example, introduces Chart-COT and Chart-RFT on top of Qwen2.5-VL-7B-Instruct, using a programmatic synthesis pipeline and GRPO-based reinforcement fine-tuning for chart reasoning (Chen et al., 21 Jul 2025). BigCharts-R1 shares the two-stage structure but differs materially in data construction: it emphasizes visually diverse chart images conditioned on real-world charts sourced from multiple online platforms, together with a replotting process that guarantees underlying data fidelity (Masry et al., 13 Aug 2025). Within the evidence given, this is the paper’s principal methodological distinction.

At a broader level, chart-image learning is not unique to chart question answering. An earlier example is the CNN-BI trading model, which used 2-D stock bar chart images directly as inputs to a convolutional neural network for buy-hold-sell prediction (Sezer et al., 2019). That work is not a chart-reasoning benchmark model, but it demonstrates that chart images can serve as primary machine-learning objects rather than mere renderings of tabular data. This suggests a broader research landscape in which chart understanding spans decision support, multimodal reasoning, and direct visual prediction.

The BigCharts-R1 paper explicitly lists several limitations and future directions. It reports marginal RL gains on descriptive, non-reasoning questions, suggesting that reward design could diversify beyond numeric error (Masry et al., 13 Aug 2025). It also observes diminishing returns on larger backbones, specifically from 7 B versus 3 B, which points to the need for stronger image encoders or pretraining (Masry et al., 13 Aug 2025). Planned extensions include tables, geometric diagrams, and richer reward functions for summarization and fact-checking (Masry et al., 13 Aug 2025).

The paper’s concluding interpretation is that carefully curated, visually diverse chart datasets plus targeted reinforcement objectives yield large gains in chart comprehension and reasoning (Masry et al., 13 Aug 2025). A plausible implication is that future progress in this area may depend less on scaling backbone size alone than on the interaction between executable data grounding, visual heterogeneity, and reward design.