To break the plateau, the authors implement a two-stage Reinforcement Learning (RL) process [11].
) to ensure the generated code matches the visual intent [11]. 2.8M GMAIL.txt
: Uses 22k data pairs focusing on textual accuracy ( To break the plateau, the authors implement a
: The SFT stage requires 60 hours of training on 16 H800 GPUs . The RL stages take an additional 34 hours on 24 H800 GPUs [11]. The RL stages take an additional 34 hours
) used in the RL stages or the used to measure the success of the 2.8M dataset?
The paper addresses the "SFT plateau," a phenomenon where Supervised Fine-Tuning (SFT) performance on Large Language Models (LLMs) stops improving even as the dataset size increases [11, 22]. The authors use a specific of chart-to-code data to demonstrate this limitation and propose Multimodal Structured Reinforcement Learning (MSRL) as a solution [11, 22]. 2. Methodology Supervised Fine-Tuning (SFT) Phase : Baseline Model : Qwen2.5-VL-7B-Instruct [11, 22].
: Qwen2.5-VL-72B-Instruct is used as the judge model for calculating visual rewards during training [11]. 4. Experimental Results