Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Chart-R1: Chain-of-Thought Supervision and Reinforcement for Advanced Chart Reasoner

Created by
  • Haebom

Author

Lei Chen, Xuanle Zhao, Zhixiong Zeng, Jing Huang, Yufeng Zhong, Lin Ma

Outline

This paper presents the Chart-R1 model, which applies an R1-style method based on reinforcement learning fine-tuning to complex inference in the chart domain. Unlike existing R1-style methods that focus on mathematical reasoning and code intelligence, Chart-R1 enhances inference capabilities for more general multimodal data, particularly chart data. To achieve this, we propose a novel programmatic data synthesis technique that generates high-quality step-by-step chart inference data containing single and multiple sub-charts. We also develop a two-step learning strategy: Chart-COT, which utilizes a Chain-of-Thought (COT) map, and Chart-RFT, which utilizes numerical sensitivity fine-tuning. Chart-COT decomposes complex inference tasks into fine-grained sub-tasks, while Chart-RFT emphasizes numerical sensitivity in the chart domain by using relatively gentle rewards for numerical responses. Experimental results show that Chart-R1 outperforms existing chart domain methods and is comparable to large-scale models such as GPT-4o and Claude-3.5.

Takeaways, Limitations

•
Takeaways:
◦
Successfully applying reinforcement learning-based R1-style methodology to complex inference problems with multimodal data, particularly chart data.
◦
Solving the problem of chart inference data shortage with new programmatic data synthesis techniques.
◦
An effective two-step learning strategy is presented that combines the counterfactual (COT) and numerical sensitivity enhancement fine-tuning (RFT).
◦
Demonstrated superior performance compared to existing methods and large-scale models.
•
Limitations:
◦
Further research is needed on the generalization performance and limitations of the proposed data synthesis technique.
◦
There is a potential bias towards certain types of chart data. Performance evaluation is needed for various types of chart data.
◦
Further research is needed on the design and optimization of the reward function used.
◦
Analysis of performance differences for specific metrics is needed when comparing with large-scale models.
👍
Made with Slashpage