Image-to-LaTeX Table Conversion on DocGenome table-1k
74.7TEDSQwen2.5-VL-7B-GRPO
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Qwen2.5-VL-7B-GRPOModel Backbone=Qwen2.5-VL, Parameter Count=7B, Reinforcement Learning Strategy=GRPO2026.04 | 74.7 | 26.9 | 68.5 | 62.7 | 38.7 | |
| Qwen2.5-VL-7B-CSPOModel Backbone=Qwen2.5-VL, Parameter Count=7B, Reinforcement Learning Strategy=CSPO2026.04 | 74.7 | 29.2 | 68.8 | 63.6 | 41.1 | |
| Qwen2.5-VL-7B-VSGRPO*Model Backbone=Qwen2.5-VL, Parameter Count=7B, Reinforcement Learning Strategy=VSGRPO2026.04 | 73.9 | 25.9 | 67.8 | 63 | 39.6 | |
| Qwen2.5-VL-3B-CSPOModel Backbone=Qwen2.5-VL, Parameter Count=3B, Reinforcement Learning Strategy=CSPO2026.04 | 72.5 | 21 | 63.6 | 54.7 | 34.5 | |
| Qwen2.5-VL-3B-VSGRPO*Model Backbone=Qwen2.5-VL, Parameter Count=3B, Reinforcement Learning Strategy=VSGRPO2026.04 | 71.5 | 17.8 | 60.3 | 53.8 | 30.9 | |
| Qwen2.5-VL-3B-GRPOModel Backbone=Qwen2.5-VL, Parameter Count=3B, Reinforcement Learning Strategy=GRPO2026.04 | 71.3 | 18.9 | 60 | 54.4 | 33.1 | |
| Gemini-2.5 FlashModel Backbone=Gemini-2.5 Flash2026.04 | 60.5 | 8 | 64.9 | 57.1 | 12.5 | |
| Qwen2.5-VL-72BModel Backbone=Qwen2.5-VL, Parameter Count=72B2026.04 | 60.3 | 5 | 61.2 | 53.9 | 8.2 | |
| GPT-4oModel Backbone=GPT-4o2026.04 | 60.2 | 5.1 | 55 | 41.2 | 8.4 | |
| Qwen2.5-VL-7BModel Backbone=Qwen2.5-VL, Parameter Count=7B2026.04 | 56.6 | 3.9 | 51.2 | 46.3 | 6 | |
| Qwen2.5-VL-3BModel Backbone=Qwen2.5-VL, Parameter Count=3B2026.04 | 50.2 | 1.1 | 43.5 | 37 | 1.7 | |
| NougatModel Backbone=Nougat2026.04 | 30.8 | 0.8 | 27.7 | 4.4 | 8.8 |