: Uses 22k data pairs focusing on textual accuracy (
: Qwen2.5-VL-72B-Instruct is used as the judge model for calculating visual rewards during training [11]. 4. Experimental Results
To break the plateau, the authors implement a two-stage Reinforcement Learning (RL) process [11]. 2.8M GMAIL.txt
: Uses 11k pairs with a balance of textual and visual rewards (
The paper demonstrates that MSRL significantly outperforms pure SFT models by optimizing for both textual structure and visual fidelity, effectively surpassing the performance limit reached at 2.8M SFT samples [11, 25]. MSRL Stage Max Dataset Size 2.8 million samples [11, 22] 33k curated samples [11] GPU Requirement 16 H800 GPUs [11] 24 H800 GPUs [11] Training Goal Min. Negative Log-Likelihood [22] Hybrid Text-Visual Reward [11] Outcome Performance Plateaus [22] Breaks SFT Performance Limit [11] : Uses 22k data pairs focusing on textual
: Increasing data from 2M to 2.8M results in no further performance gains, confirming the plateau [22]. Multimodal Structured Reinforcement Learning (MSRL) :
: The model is tested on subsets ranging from 200k to 2.8 million samples. : Uses 11k pairs with a balance of
) to ensure the generated code matches the visual intent [11].