Emotional Reasoning on EmoSet_Emotion6_WebEmo
74.06In-domain AccuracyEMO-R3
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| EMO-R3Backbone=Qwen2.5-VL-3B-Instruct, Roll-out=8, Post-training=true2026.02 | 74.06 | 53.6 | 60.42 | |
| SFTBackbone=Qwen2.5-VL-3B-Instruct, Roll-out=-, Post-training=true2026.02 | 73.34 | 29.09 | 43.84 | |
| EMO-R3Backbone=Qwen2.5-VL-3B-Instruct, Roll-out=4, Post-training=true2026.02 | 73.1 | 54.2 | 60.5 | |
| GRPOBackbone=Qwen2.5-VL-3B-Instruct, Roll-out=4, Post-training=true2026.02 | 72.74 | 53.59 | 59.97 | |
| GRPOBackbone=Qwen2.5-VL-3B-Instruct, Roll-out=8, Post-training=true2026.02 | 72.66 | 52.42 | 59.16 | |
| DAPOBackbone=Qwen2.5-VL-3B-Instruct, Roll-out=4, Post-training=true2026.02 | 68.78 | 53.04 | 58.28 | |
| DAPOBackbone=Qwen2.5-VL-3B-Instruct, Roll-out=8, Post-training=true2026.02 | 66.3 | 52.22 | 56.91 | |
| SEPM*Backbone=LLaVA1.5-7B, Roll-out=-, Post-training=false2026.02 | 55.13 | 48.76 | 50.88 | |
| Vanilla*Backbone=Qwen2.5-VL-3B-Instruct, Roll-out=-, Post-training=false2026.02 | 50.77 | 45.71 | 47.4 | |
| Vanilla*Backbone=LLaVA1.5-7B, Roll-out=-, Post-training=false2026.02 | 50.55 | 38.05 | 42.22 |