Research map / Reinforcement learning / RL from feedback
MM-ReCoder: Advancing Chart-to-Code Generation with Reinforcement Learning and Self-Correction
Zitian Tang, Xu Zhang, Jianbo Yuan, Yang Zou, Varad Gunjal, Songyao Jiang, Davide Modolo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add shared-first-turn RL stage before full-trajectory RL (ablation) | MM-ReCoder RL · ChartMimic low-level (second turn) | low-level score: 83.9 -> 86.0 | Improved |
| use shared-first-turn instead of full-trajectory for self-correction (ablation) | MM-ReCoder RL · ChartMimic | percentage of improved samples: 3.4 -> 14.4 | Improved |
| add self-correction bonus eta=0.1 with gamma=0 in full-trajectory (ablation) | MM-ReCoder RL · ChartMimic first turn | first-turn exec rate: 91.8 -> 64.5 | Hurt |
| add model-based reward (ablation) | MM-ReCoder RL · ChartMimic single-turn | high-level score: 78.6 -> 83.7 | Improved |
| add rule-based reward (ablation) | MM-ReCoder RL · ChartMimic single-turn | low-level score: 78.2 -> 83.9 | Improved |
| use Qwen2.5-VL-72B instead of 7B as reward model (ablation) | MM-ReCoder RL · ChartMimic single-turn | low-level score: 81.0 -> 84.3 | Improved |
| add multi-turn cold start data after single-turn cold start (ablation) | MM-ReCoder · ChartMimic | second-turn low-level: 84.8 -> 86.0 | Improved |
| add any cold start versus none (ablation) | MM-ReCoder · ChartMimic | second-turn low-level: 84.9 -> 86.0 | Mixed |
| train full-trajectory RL for 2 epochs instead of 1 (ablation) | MM-ReCoder RL · ChartMimic | second-turn low-level: 83.9 -> 81.8 | Hurt |
| increase reward weights on model-based reward (ablation) | MM-ReCoder RL · ChartMimic single-turn | low-level score: 84.8 -> 83.9 | No clear effect |
| add more self-correction turns at inference (2 to 4) | MM-ReCoder · ChartMimic | low-level score: 84.8 -> 86.5 | Improved |
| add a fifth self-correction turn | MM-ReCoder · ChartMimic | low-level score: 86.5 -> 86.4 | No clear effect |
Closest papers by meaning
- Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.89)
- Training Language Models to Self-Correct via Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 2025 · similarity 0.86)
- RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning (ICML 2025 · similarity 0.86)
- R1-Code-Interpreter: LLMs Reason with Code via Supervised and Multi-stage Reinforcement Learning (ICLR 2026 · similarity 0.86)
- Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
