CDL Generation on Formalgeo Rec-CoT 1.0 (test)
97.2Img CDL RecallQwen3-VL 4B (RL)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Qwen3-VL 4B (RL)Training Data=5.5k, Training Mode=RL, Backbone=Qwen3-VL 4B2026.01 | 97.2 | 99.1 | 99.1 | 97.2 | 95.4 | 95.1 | |
| Qwen2.5-VL 7B (RL)Training Data=5.5k, Training Mode=RL, Backbone=Qwen2.5-VL 7B2026.01 | 97 | 99.1 | 99.1 | 96.9 | 92.7 | 92.1 | |
| Qwen3-VL 8B (RL)Training Data=5.5k, Training Mode=RL, Backbone=Qwen3-VL 8B2026.01 | 97 | 99.2 | 99.3 | 96.9 | 95.9 | 95.8 | |
| Qwen2.5-VL 3B (RL)Training Data=5.5k, Training Mode=RL, Backbone=Qwen2.5-VL 3B2026.01 | 96.5 | 98.9 | 98.9 | 96.9 | 87.1 | 87.6 | |
| Qwen3-VL 4B (SFT)Training Data=5.5k, Training Mode=SFT, Backbone=Qwen3-VL 4B2026.01 | 96.4 | 97.1 | 97.1 | 95.1 | 80.9 | 79.4 | |
| Qwen3-VL 8B (SFT)Training Data=5.5k, Training Mode=SFT, Backbone=Qwen3-VL 8B2026.01 | 96 | 97.2 | 97.2 | 95 | 80.9 | 79.6 | |
| Qwen2.5-VL 7B (SFT)Training Data=5.5k, Training Mode=SFT, Backbone=Qwen2.5-VL 7B2026.01 | 95.1 | 96.5 | 96.6 | 93.5 | 70 | 68.3 | |
| Qwen2.5-VL 3B (SFT)Training Data=5.5k, Training Mode=SFT, Backbone=Qwen2.5-VL 3B2026.01 | 94.7 | 94.7 | 94.7 | 92.2 | 62.3 | 58.7 | |
| Diagram FormalizerPub. & Year=ICASSP’25, Training Data=1M, Input Mode=Diagram-only2026.01 | 92.9 | — | — | — | 90.3 | — | |
| FgeoParser (Diag.)Pub. & Year=Symmetry’24, Training Data=14.7k, Input Mode=Diagram-only2026.01 | 77.5 | — | — | — | 87 | — | |
| FgeoParser (Text)Pub. & Year=Symmetry’24, Training Data=14.7k, Input Mode=Text-only2026.01 | — | 96.5 | — | — | — | — |