Knowledge-Based QA on Global-MMLU Lite (5 Unseen Languages)
71.9MMLU ScoreCAGrad
Evaluation Results
| Method | Links | |
|---|---|---|
| CAGradBackbone=Qwen3-8B-Base2026.06 | 71.9 | |
| MGDABackbone=Qwen3-8B-Base2026.06 | 71 | |
| PCGradBackbone=Qwen3-8B-Base2026.06 | 70.3 | |
| BaseBackbone=Qwen3-8B-Base2026.06 | 70 | |
| V. SFTBackbone=Qwen3-8B-Base2026.06 | 69.1 | |
| MGDABackbone=Qwen3-4B-Base2026.06 | 66 | |
| CAGradBackbone=Qwen3-4B-Base2026.06 | 65.5 | |
| PCGradBackbone=Qwen3-4B-Base2026.06 | 64 | |
| BaseBackbone=Qwen3-4B-Base2026.06 | 63.6 | |
| V. SFTBackbone=Qwen3-4B-Base2026.06 | 62.3 | |
| PCGradBackbone=Llama-3.1-8B2026.06 | 52.6 | |
| CAGradBackbone=Llama-3.1-8B2026.06 | 52.1 | |
| MGDABackbone=Llama-3.1-8B2026.06 | 51.7 | |
| MGDABackbone=Meta-Llama-3-8B2026.06 | 50.6 | |
| CAGradBackbone=Meta-Llama-3-8B2026.06 | 50.5 | |
| V. SFTBackbone=Llama-3.1-8B2026.06 | 50.5 | |
| PCGradBackbone=Meta-Llama-3-8B2026.06 | 49 | |
| V. SFTBackbone=Meta-Llama-3-8B2026.06 | 48.9 | |
| BaseBackbone=Llama-3.1-8B2026.06 | 48.9 | |
| BaseBackbone=Meta-Llama-3-8B2026.06 | 42.6 |