ViCo: Visual-oriented Coding with Self-Reflection for Chart Replication

Jiaxin Duan; Dian Jiao Shuai Zhao; Jiabing Leng; Yiran Zhang; Feng Huang

cs.CLAug 31, 2026

ViCo: Visual-oriented Coding with Self-Reflection for Chart Replication

Authors: Jiaxin Duan, Dian Jiao Shuai Zhao, Jiabing Leng, Yiran Zhang, Feng Huang

Organizations: China Electronics Cloud Technology Co., Ltd.

Abstract

This paper addresses the challenge of generating high-quality academic charts that match the visual standards of human-authored papers. While existing AI agents can produce well-structured text and code, their generated visualizations often lack the stylistic and semantic fidelity of human designs. Advanced coding agents that employ self-reflection mechanisms exhibit poor visual reasoning and limited reflection following, resulting in sparse reward signals that severely undermine their reinforcement learning (RL). We propose ViCo, a training framework for visual-oriented coding that employs iterative reflections to align generated chart images progressively with the reference. We first introduce a self-supervised warm-up stage, which augments Monte Carlo Tree Search with consistency-based pruning to synthesize high-quality reflection trajectories, ensuring that each coding step strictly follows the outcomes of prior reflections. A multi-step RL algorithm is then developed, using counterfactual baselines to estimate advantage for reflection and action steps within each refinement cycle, thereby addressing the reward sparsity. To enable efficient reward in massive training, we propose an automatic, multifaceted evaluation framework that assesses charts' style, layout, and semantic consistency via a hierarchical heterogeneous layout graph structure. Experiments on three public benchmarks demonstrate that ViCo, trained on an 8B model, achieves performance close to proprietary LLMs with adequate reflection capabilities.

Figures & tables

Figure 1: Overview of visual coding with self-reflection.

Figure 2: Overview of the ViCo two-stage training.

Closed-source LLMs
Model	Size	RealChart2Code ∗		ChartMimic				Plot2Code
Model	Size	Pass (%)	Score	Direct P.(%)	Direct S.	Cust. P † .(%)	Cust. S † .	Pass (%)	Text	Rating
Claude-4.5-Sonnet Anthropic (2025)	-	85.8	7.1	100.0	90.1	99.5	91.5	93.9	73.3	8.8
Claude-4.5-Opus Anthropic (2025)	-	87.7	7.8	98.5	92.3	100.0	95.1	99.6	75.4	9.0
Gemini-2.5-Flash Google DeepMind (2026)	-	63.2	5.2	88.5	69.8	89.1	74.4	87.9	72.8	8.6
Gemini-3-Pro-Preview Google DeepMind (2026)	-	82.1	9.0	97.3	96.0	100.0	96.8	90.9	79.5	9.6

Table 1: Evaluation results on three benchmarks. *: Only the chart replication task is evaluated. The best score within each category is bolded and the second-best is underlined . †: Customized replication is achieved through a post-hoc data refinement. (

\uparrow

): the absolute performance gain of ViCo against the Qwen3-VL base models (4B/8B).

Panel of Figure 3 — Figure 3: Training dynamics of multi-step PPO with different rewarding and credit assignment mechanisms.

Rewarding Strategy (Sec. 4.3 , Group 1)
Configuration	RealChart2Code		ChartMimic
Configuration	Pass (%)	Score	Pass (%)	Score
ViCo (8B) ( Full Framework )	98.8	3.4	99.2	81.5
w/o RL (Warmup-only)	96.5	3.2	94.2	77.1
w/o SFT (Cold-start RL)	50.5	2.7	69.9	69.7
HHLG $\to$ LLM-As-A-Judge	94.1	2.8	95.0	72.3

Table 2: Ablation studies of ViCo (8B) configurations evafluated on RealChart2Code and ChartMimic (Direct Mimic). We report the Pass Rate (%) and visual Score.

Evaluation Metric	Spearman ( $\rho$ )		Pearson ( $r$ )		Kendall ( $\tau$ )
Evaluation Metric	Sim.	Com.	Sim.	Com.	Sim.	Com.
Baselines (LLM-as-a-Judge)
Qwen3-VL (8B)	0.425	0.184	0.451	0.203	0.318	0.125
GPT-4o	0.712	0.321	0.738	0.355	0.546	0.241
HHLG-Based (Ours)	0.724	0.458	0.701	0.412	0.503	0.317

Table 3: Correlation between different evaluation metrics and human judgment, categorized by chart complexity ( Sim. : Simple, Com. : Complex).

Model	Simple ( $1\times 1$ )		Medium ( $2\times 2$ – $2\times 3$ )		Complex ( $>6$ subplots)
Model	Pass (%)	Score	Pass (%)	Score	Pass (%)	Score
Qwen3-VL-8B	90.8	5.8	78.8	2.8	20.3	0.5
ChartCoder	89.5	5.6	72.3	3.5	28.0	0.7
ViCo (8B)	100.0	8.7	99.0	6.7	99.0	1.2
+ Divide-and-Conquer	–	–	100.0	8.2	100.0	7.6

Table 4: Performance breakdown across layout complexity tiers on RealChart2Code.

Figure 4: Cross-backbone generalization of ViCo evaluated on RealChart2Code and ChartMimic.

Panel of Figure 5 — Figure 5: Evaluation against reflection iterations.

Figure 6: Comparison of advantage estimation in PPO with varying number of counterfactual baselines.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Exact Match	pHash	CLIP Threshold	Filtered from Train (%)
RealChart2Code Zhang et al. (2026)	0 (0.00%)	Hamming $\leq 4$	Cosine sim. $\geq 0.88$	0 (0.00%)
ChartMimic Yang et al. (2025)	0 (0.00%)	Hamming $\leq 4$	Cosine sim. $\geq 0.88$	154 (0.26%)
Plot2Code Wu et al. (2025)	0 (0.00%)	Hamming $\leq 4$	Cosine sim. $\geq 0.88$	50 (0.08%)

Appendix

Table 5: Statistics of the multi-stage sample-level de-duplication pipeline across evaluation benchmarks.

Figure 7: 2D manifold projection of visual representations across datasets.

Hyperparameter	Phase 1 (SSFT)	Phase 2 (MSRL)
Backbone Model	Qwen3-VL Bai et al. (2025)
Optimizer	AdamW Loshchilov and Hutter (2019)
Learning Rate	$1\times 10^{-5}$	$5\times 10^{-6}$
Learning Rate Schedule	Cosine decay	Cosine decay
Warmup Ratio	0.03	0.05
Batch Size	128	64

Appendix

Table 6: Hyperparameter settings for Phase 1 (Self-Supervised Fine-tuning, SSFT) and Phase 2 (Multi-step Reinforcement Learning, MSRL) training of ViCo.

Hyperparameter	Symbol	Value	Explanation
Stage 1: MCTS-Based Trajectory Synthesis & Warm-Up
Max reflection steps	$K$	3	Maximum iteration depth of the multi-turn coding loop
Branching factor	$C$	8	Number of candidate code expansions generated per node
Exploration constant	$c$	1.25	Standard PUCT exploration weight
Maximum patience	$\rho_{\max}$	2	Sequence tolerance for consecutive non-improving steps
Penalty coefficient	$\beta$	0.8	Scaling factor for differential history penalty

Appendix

Table 7: Comprehensive hyperparameter settings across the MCTS warm-up stage, multi-step PPO reinforcement learning, and the HHLG-based reward evaluation framework. Overloaded notations from prior versions have been systematically disambiguated.

Single-Stage Control Baselines
Training Configuration	RealChart2Code		ChartMimic
Training Configuration	Pass (%)	Score	Pass (%)	Score
Vanilla SFT (Teacher Pseudo-labels)	85.2	2.4	90.5	68.4
MCTS Warm-up (w/o RL)	89.4	2.6	92.1	72.8
Cold-start RL (w/o Warm-up)	62.5	1.8	71.3	45.2
Full ViCo (Warm-up + RL)	98.8	3.4	99.2	81.5

Appendix

Table 8: Stage-level performance decomposition of ViCo (8B) evaluated on RealChart2Code and ChartMimic (Direct Mimic).

Phase 1: SFT Warm-Up Stage
Training Configuration	Pass (%)	Score	$\mathbf{S_{\rm sem}}$	$\mathbf{S_{\rm lay}}$	$\mathbf{S_{\rm style}}$	$\mathbf{S_{\rm color}}$
ViCo-4B (Warmup-only)	89.1	69.4	75.4	69.5	63.8	65.2
w/o Semantic	88.2	44.8	32.2	70.8	64.3	65.6
w/o Layout	88.3	46.5	74.8	38.6	58.1	60.2
w/o Style	88.5	58.2	75.6	69.2	36.4	64.9
w/o Color	88.7	57.3	75.1	69.4	63.4	34.8

Appendix

Table 9: Leave-One-Component-Out (LOO) ablation of the HHLG reward formulation across SFT warm-up and RL stages using ViCo-4B evaluated on ChartMimic.

Figure 8: Trajectory of internal HHLG rewards versus independent GPT-4o evaluation scores under extended training horizons on RealChart2Code. Both metrics align synchronously, demonstrating no reward hacking up to 2,000 steps.

Parsing Perturbation	Extent	Sub-metric Impact ( $\Delta S_{\rm sem}/\Delta S_{\rm lay}/\Delta S_{\rm style}/\Delta S_{\rm color}$ )	$\mathbf{\Delta\rho_{\rm human}}$
OCR token corruption	10%	( $-0.10$ , $-0.02$ , $\phantom{-}0.00$ , $\phantom{-}0.00$ )	$-0.015$
Visual-node dropout	10%	( $\phantom{-}0.00$ , $-0.08$ , $-0.01$ , $-0.07$ )	$-0.035$
Mask erosion/dilation	10%	( $\phantom{-}0.00$ , $-0.01$ , $\phantom{-}0.00$ , $-0.04$ )	$-0.012$
Text-box jitter	5%	( $\phantom{-}0.00$ , $-0.03$ , $\phantom{-}0.00$ , $\phantom{-}0.00$ )	$-0.008$
Subplot-boundary jitter	5%	( $-0.02$ , $-0.06$ , $-0.05$ , $-0.03$ )	$-0.042$

Appendix

Table 10: Sensitivity analysis of HHLG sub-metrics and human judgment correlation (

\Delta\rho_{\rm human}

) under controlled upstream parsing noise.

Model	Pass@1 (Single-shot)	Reflect $\times$ 1	Reflect $\times$ 2
Qwen3-VL-8B	65.6	63.4	65.2
Qwen3-VL-30B	89.5	90.5	90.8
Qwen3-VL-235B	95.5	98.3	99.5
ChartCoder	90.9	83.2	79.5
ViCo-8B	88.5	96.7	99.2

Appendix

Table 11: Code execution pass rates (%) under strictly controlled inference budgets on ChartMimic.

RL Method	RealChart2Code		ChartMimic
RL Method	Pass (%)	Score	Pass (%)	Score
Warmup + GRPO (Terminal reward)	49.7	3.0	89.2	69.7
Warmup + PPO (Terminal reward)	57.1	3.0	89.4	77.1
ViCo (Warmup + MSRL)	98.8	3.4	99.2	81.5

Appendix

Table 12: Performance comparison between holistic trajectory-level RL baselines and ViCo multi-step counterfactual RL (MSRL) on RealChart2Code and ChartMimic.

Method	GPU Setup	GPU Hours (h)	Peak Mem. (GiB)	Throughput (/s)
Warmup SFT	8 $\times$ A800	17.5	554.6	5.94 samples
Multi-turn GRPO	8 $\times$ A800	14.2	527.4	10.80 rollouts
Multi-turn PPO	8 $\times$ A800	17.2	551.2	1.42 samples
ViCo MSRL	8 $\times$ A800	24.4	523.4	0.84 samples

Appendix

Table 13: Training resource profiling for the 4B parameter model on 8

\times

NVIDIA A800 GPUs.

Method	GPU Setup	GPU Hours (h)	Peak Mem. (GiB)	Throughput (/s)
Warmup SFT	8 $\times$ A800	18.0	571.6	5.80 samples
Multi-turn GRPO	8 $\times$ A800	18.3	525.5	6.92 rollouts
Multi-turn PPO	8 $\times$ A800	18.8	573.5	1.39 samples
ViCo MSRL	8 $\times$ A800	25.6	535.7	0.79 samples

Appendix

Table 14: Training resource profiling for the 8B parameter model on 8

\times

NVIDIA A800 GPUs.

AI Coding Agents

AI coding agents use language models and software tools to plan, edit, test, and maintain code with varying levels of autonomy.

See all 329 papers

UniCoder: Unified Visual-to-Code Generation via Symbolic Rewards and Reference-Guided Code Optimization
Jun 30, 2026Yaozhi Zheng, Yilei Jiang, Manyuan Zhang +5RL for Code Generation Reinforcement Learning
The Chinese University of Hong Kong · Shanghai Artificial Intelligence Laboratory
ChartJudgeBench: Evaluating LMM Judges for Chart-to-Code Generation
Sep 21, 2026Lijian Wu, Henry Hengyuan Zhao, Zijian Zhang +3VLM Evaluation LLM Evaluation
CSU-JPG, Central South University · National University of Singapore
Self-Distillation Policy Optimization via Visual Feedback: Bridging Code and Visual Artifacts
Jun 9, 2026Haoyu DongRL for Code Generation Code Generation
Microsoft

Jun 30, 2026J/K move · Enter open · S save

UniCoder: Unified Visual-to-Code Generation via Symbolic Rewards and Reference-Guided Code Optimization

Yaozhi Zheng, Yilei Jiang, Manyuan Zhang, Yuxuan Wan+4

The Chinese University of Hong Kong · Shanghai Artificial Intelligence Laboratory

RL for Code Generation Reinforcement Learning Reward Shaping Code Generation Multimodal Large Language Models

Open paper AskPDF arXiv

Visual-to-Code generation, which transforms scientific plots, vector graphics, and webpages into executable scripts, demands a level of pixel-precise alignment that standard Multimodal Large Language Models (MLLMs) fail to achieve through Supervised Fine-Tuning (SFT) alone. While Reinforcement Learning (RL) offers a theoretical pathway to bridge this gap, its application is hindered by two fundamental obstacles: (1) Reward Coarseness, where semantic metrics like CLIP scores fail to penalize fine-grained element deviations, and (2) Exploration Stagnation, where the sparse, heterogeneous code search space prevents the policy from bootstrapping valid trajectories. To overcome these limitations, we introduce UniCoder, a unified RL framework that integrates two novel mechanisms. First, we propose Symbolic Attribute Alignment, which employs a lightweight auxiliary LLM to parse generated code into discrete visual attributes (e.g., hex colors, coordinate limits), enabling dense, element-wise reward computation. Second, to escape local optima, we devise Reference-Guided Code Optimization, a strategy that dynamically injects ground-truth trajectories into low-performing rollout groups, transforming blind exploration into guided policy improvement. Extensive experiments on ChartMimic, UniSVG, Design2Code and ScreenBench benchmarks demonstrate that our 8B-parameter model not only surpasses all open-source baselines but also achieves state-of-the-art performance comparable to proprietary models, establishing a new paradigm for generalized visual-to-code synthesis.

Abstract

AI Coding Agents

Explore similar work

UniCoder: Unified Visual-to-Code Generation via Symbolic Rewards and Reference-Guided Code Optimization

ChartJudgeBench: Evaluating LMM Judges for Chart-to-Code Generation

Self-Distillation Policy Optimization via Visual Feedback: Bridging Code and Visual Artifacts