ChartVerse

Overview

Abstract

Chart reasoning is a critical capability for Vision Language Models (VLMs). However, the development of open-source models is severely hindered by the lack of high-quality training data. Existing datasets suffer from a dual challenge: synthetic charts are often simplistic and repetitive, while the associated QA pairs are prone to hallucinations and lack the reasoning depth required for complex tasks.

To bridge this gap, we propose ChartVerse, a scalable framework designed to synthesize complex charts and reliable reasoning data from scratch:

To address the bottleneck of simple patterns, we first introduce Rollout Posterior Entropy (RPE), a novel metric that quantifies chart complexity. Guided by RPE, we develop complexity-aware chart coder to autonomously synthesize diverse, high-complexity charts via executable programs.
To guarantee reasoning rigor, we develop truth-anchored inverse QA synthesis. Diverging from standard generation, we adopt an answer-first paradigm: we extract deterministic answers directly from the source code, generate questions conditional on these anchors, and enforce strict consistency verification.

We curate ChartVerse-SFT-600K and ChartVerse-RL-40K using Qwen3-VL-30B-A3B-Thinking as the teacher. Experimental results demonstrate that ChartVerse-8B achieves state-of-the-art performance, notably surpassing its teacher and rivaling the stronger Qwen3-32B-Thinking.

Approach

Methodology

A two-stage framework combining complexity-aware chart synthesis with truth-anchored QA generation

📊

1. Rollout Posterior Entropy

Given a chart, we prompt a VLM to generate executable code 8 times. Simple charts yield consistent codes, while complex ones lead to divergence. We compute spectral entropy from CLIP embeddings, retaining samples with RPE ≥ 0.4.

🔄

2. Complexity-Aware Coder

We train a specialized coder via iterative self-enhancement. Starting from a cold-start dataset filtered by high RPE, we repeatedly generate, filter by complexity and similarity, and retrain to produce diverse charts from scratch.

✓

3. Truth-Anchored Inverse QA

We invert the generation flow: first extract deterministic answers from code execution, then reverse-engineer questions. A consistency check ensures logical soundness. Failure-rate filtering retains challenging samples.

Data

Datasets & Comparison

ChartVerse-SFT-600K achieves superior diversity, complexity, and reliability

Dataset	Chart Count	Data Source	Textual Data	QA Count	Total Tokens	Answer Accuracy	Reasoning Data	Rollout Posterior Entropy ↑	Color Entropy ↑	Semantic Embedding Spread ↑
ChartQA	18K	Real + Synthetic	Table	28K	100K	✗	✗	0.26	2.03	0.37
PlotQA	156K	Real + Synthetic	Table	20M	117M	✗	✗	0.21	0.83	0.30
FigureQA	100K	Synthetic	-	1.3M	2.6M	✗	✗	0.25	1.04	0.24
ReachQA	3K	Synthetic	Code	20K	1.2M	✗	✓	0.29	1.91	0.47
ECD	10K	Synthetic	Code	321K	18.6M	✗	✓	0.31	2.13	0.40
CoSyn	116K	Synthetic	Code	1.1M	5.4M	✗	✗	0.35	1.73	0.54
START	400K	Synthetic	Code	400K	143M	✗	✗	0.33	1.63	0.27
ChartGen	216K	Synthetic	Code	-	-	✗	✗	0.30	1.31	0.39
ChartCoder	163K	Synthetic	Code	-	-	✗	✗	0.33	1.48	0.38
ChartVerse-SFT-600K (Ours)	412K	Synthetic	Code	603K	3.9B	✓	✓	0.44	3.17	0.51

Table 1: Comparison of ChartVerse-SFT-600K with existing chart datasets across chart properties, QA properties, and complexity/diversity metrics.

📈 Higher Complexity

Highest RPE score (0.44) compared to all baselines, indicating genuinely challenging visual structures.

🎨 Superior Diversity

Highest Color Entropy (3.17) and Semantic Embedding Spread, covering broad visual styles.

✓ Rigorous Reliability

Guaranteed answer accuracy via inverse QA pipeline with code-execution verification.

Results

Performance

Data quality triumphs over model scale—smaller ChartVerse models outperform much larger baselines

📊 Main Results

Model	ChartQA-Pro	CharXiv (Reasoning)	CharXiv (Descriptive)	ChartMuseum	ChartX	EvoChart	ChartBench (GPT-acc)	Average
ECD-7B	45.1	41.7	74.9	24.5	59.3	56.0	48.6	50.0
START-7B	43.5	46.3	76.8	29.7	57.3	63.8	50.0	52.5
Chart-R1-7B	45.6	47.7	70.0	33.4	60.9	67.1	50.5	53.6
ChartVerse-2B	48.2	46.9	71.2	37.5	60.5	66.8	49.1	54.3
InternVL3.5-38B	47.8	45.5	86.5	34.2	55.5	65.0	49.0	54.8
InternVL3.5-241B-A28B	50.7	47.5	88.1	36.1	59.1	67.4	48.6	56.8
Qwen3-VL-8B-Thinking	53.9	53.0	85.9	44.3	59.6	74.1	49.1	60.0
ChartVerse-4B	55.2	56.2	84.1	45.9	63.7	75.0	52.9	61.9
Qwen3-VL-30B-A3B-Thinking	55.8	56.6	86.9	49.2	62.3	77.2	52.4	62.9
ChartVerse-8B	56.2	60.8	88.0	49.2	63.9	76.2	54.2	64.1
Qwen3-VL-32B-Thinking	58.8	65.2	90.2	55.9	64.1	80.8	54.3	67.0
Qwen3-VL-235B-A30B-Thinking	60.0	66.1	90.5	60.0	64.5	79.9	54.5	67.9

Table 2: Comparison of different models on 7 chart reasoning benchmarks. Shaded rows denote our ChartVerse models.

🔥 Small > Large

ChartVerse-4B (61.9%) outperforms Qwen3-VL-8B-Thinking (60.0%) with half the parameters.

🎓 Student > Teacher

ChartVerse-8B (64.1%) surpasses its teacher Qwen3-VL-30B-A3B-Thinking (62.9%).

🏆 Top-Tier Performance

Our 8B model rivals the much larger Qwen3-VL-32B-Thinking on chart reasoning benchmarks.

🔄 Training Stages Analysis

Model	ChartQA-Pro	CharXiv (Descriptive)	CharXiv (Reasoning)	ChartMuseum	ChartX	EvoChart	ChartBench	Chart Avg
Qwen3-VL-Instruct-2B	42.1	62.3	26.8	23.9	49.8	53.6	38.8	42.5
+ SFT	44.4	69.1	40.8	30.0	56.9	61.0	46.4	49.8
+ RL (ChartVerse-2B)	48.2	71.2	46.9	37.5	60.5	66.8	49.1	54.3
Qwen3-VL-Instruct-4B	53.7	76.2	39.7	37.2	57.2	68.2	45.1	53.9
+ SFT	52.9	84.5	52.8	42.5	61.9	73.3	50.3	59.7
+ RL (ChartVerse-4B)	55.2	84.1	56.2	45.9	63.7	75.0	52.9	61.9
Qwen3-VL-Instruct-8B	54.4	83.0	46.4	39.6	58.2	70.2	46.6	56.9
+ SFT	55.5	88.3	56.2	47.5	61.0	76.7	52.2	62.5
+ RL (ChartVerse-8B)	56.2	88.0	60.8	49.2	63.9	76.2	54.2	64.1

Table 3: Performance evolution across training stages. SFT establishes strong foundations; RL further enhances performance on challenging samples.

🧪 Generalization to STEM Tasks

The reasoning skills acquired from ChartVerse data transfer effectively to out-of-domain STEM reasoning tasks, demonstrating the general value of our high-quality chart reasoning data.

Model	MathVista	DynaMath	MathVerse	LogicVista	VisuLogic	STEM Avg
Qwen3-VL-Instruct-2B	61.3	52.1	54.2	35.8	11.5	43.0
+ SFT	58.7	47.6	55.8	43.2	17.5	44.6
+ RL (ChartVerse-2B)	60.4	49.2	56.8	48.5	23.1	47.6
Qwen3-VL-Instruct-4B	73.7	46.8	65.3	53.2	19.0	51.6
+ SFT	70.9	60.7	71.1	57.3	25.0	57.0
+ RL (ChartVerse-4B)	72.8	62.0	71.8	59.3	27.0	58.6
Qwen3-VL-Instruct-8B	77.2	62.1	67.7	55.3	22.5	56.7
+ SFT	75.0	67.4	75.6	61.3	26.5	61.2
+ RL (ChartVerse-8B)	75.6	69.0	76.5	62.6	27.1	62.2

Table 4: Performance on STEM-related benchmarks. ChartVerse-8B improves from 56.7 to 62.2 (+5.5%), demonstrating strong transfer to mathematical and logical reasoning tasks.