Social fairness evaluation on FACET (test)
92.58AccuracyMS-PoE + BAIR
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| MS-PoE + BAIRBackbone=Qwen2.5-VL-3B2026.05 | 92.58 | 70.71 | 3.53 | |
| MS-PoE + BAIRBackbone=DeekSeek-VL-7B2026.05 | 91.94 | 51.58 | 3.7 | |
| MS-PoEBackbone=DeekSeek-VL-7B2026.05 | 91.82 | 51.05 | 3.78 | |
| Standard RAG + BAIRBackbone=DeekSeek-VL-7B2026.05 | 91.02 | 30.95 | 2.48 | |
| MAD-RAG + BAIRBackbone=DeekSeek-VL-7B2026.05 | 90.96 | 31.68 | 2.63 | |
| Standard RAGBackbone=DeekSeek-VL-7B2026.05 | 90.89 | 31.05 | 2.64 | |
| MAD-RAGBackbone=DeekSeek-VL-7B2026.05 | 90.86 | 30.84 | 2.65 | |
| Visual-focus Instr. + BAIRBackbone=DeekSeek-VL-7B2026.05 | 90.81 | 30.84 | 2.71 | |
| Visual-focus Instr.Backbone=DeekSeek-VL-7B2026.05 | 90.47 | 28.53 | 2.83 | |
| BaselineBackbone=DeekSeek-VL-7B2026.05 | 90.24 | — | — | |
| Standard RAG + BAIRBackbone=Qwen2.5-VL-3B2026.05 | 88.56 | 50.31 | 4.64 | |
| LongLLMLingua + BAIRBackbone=DeekSeek-VL-7B2026.05 | 87.8 | 33.79 | 6.36 | |
| LongLLMLinguaBackbone=DeekSeek-VL-7B2026.05 | 87.76 | 34.21 | 6.45 | |
| LongLLMLingua + BAIRBackbone=Qwen2.5-VL-3B2026.05 | 87.45 | 45.97 | 4.89 | |
| Visual-focus Instr. + BAIRBackbone=Qwen2.5-VL-3B2026.05 | 87.28 | 44.93 | 5.2 | |
| MAD-RAG + BAIRBackbone=Qwen2.5-VL-3B2026.05 | 85.1 | 36.08 | 6.19 | |
| BaselineBackbone=Qwen2.5-VL-3B2026.05 | 84.91 | — | — | |
| MAD-RAGBackbone=Qwen2.5-VL-3B2026.05 | 84.41 | 25.53 | 5.11 | |
| Visual-focus Instr.Backbone=Qwen2.5-VL-3B2026.05 | 84.38 | 28.25 | 5.65 | |
| LongLLMLinguaBackbone=Qwen2.5-VL-3B2026.05 | 84.36 | 31.59 | 6.26 | |
| Standard RAGBackbone=Qwen2.5-VL-3B2026.05 | 84.34 | 26.62 | 5.41 | |
| MS-PoEBackbone=Qwen2.5-VL-3B2026.05 | 82 | 29.07 | 8.6 |