Multiple-Choice Question on PhyX mini (test)
46.2Overall AccuracyGRPO (Fmt + Acc + ASM)
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| GRPO (Fmt + Acc + ASM)Training=GRPO, Reward Configuration=Format + Accuracy + ASM, Backbone=Granite Vision 3.3 (2B)2026.04 | 46.2 | 50.6 | 49.1 | 41.2 | 46.1 | 41.6 | 48.5 | |
| GRPO (Fmt + Acc)Training=GRPO, Reward Configuration=Format + Accuracy, Backbone=Granite Vision 3.3 (2B)2026.04 | 46 | 38.2 | 53.3 | 38.2 | 61.8 | 34.3 | 50.3 | |
| GRPO (Rubric)Training=GRPO, Reward Configuration=Rubric, Backbone=Granite Vision 3.3 (2B)2026.04 | 44 | 30.6 | 38.5 | 44.2 | 62.4 | 38 | 50.9 | |
| SFTTraining=Supervised Fine-Tuning, Backbone=Granite Vision 3.3 (2B)2026.04 | 43.3 | 32 | 37.1 | 33.7 | 52.9 | 52.4 | 51.9 | |
| GRPO (ASM)Training=GRPO, Reward Configuration=Attention Score Mask (ASM), Backbone=Granite Vision 3.3 (2B)2026.04 | 35.2 | 57.1 | 26 | 13.9 | 16.4 | 54.2 | 43 | |
| GRPO (Fmt)Training=GRPO, Reward Configuration=Format (Fmt), Backbone=Granite Vision 3.3 (2B)2026.04 | 30.4 | 34.1 | 33.7 | 25.5 | 27.3 | 30.7 | 30.9 | |
| BaselineBackbone=Granite Vision 3.3 (2B)2026.04 | 21.7 | 20 | 35.5 | 20 | 10.3 | 15.7 | 28.5 |