Knowledge-Based QA on Global-MMLU-Lite (8 Seen Languages)
65.2MMLU ScoreCAGrad
Evaluation Results
| Method | Links | |
|---|---|---|
| CAGradBackbone=Qwen3-8B-Base2026.06 | 65.2 | |
| MGDABackbone=Qwen3-8B-Base2026.06 | 64.8 | |
| PCGradBackbone=Qwen3-8B-Base2026.06 | 64.6 | |
| V. SFTBackbone=Qwen3-8B-Base2026.06 | 63.4 | |
| BaseBackbone=Qwen3-8B-Base2026.06 | 61.7 | |
| PCGradBackbone=Qwen3-4B-Base2026.06 | 58.4 | |
| MGDABackbone=Qwen3-4B-Base2026.06 | 58.1 | |
| CAGradBackbone=Qwen3-4B-Base2026.06 | 58 | |
| BaseBackbone=Qwen3-4B-Base2026.06 | 56.9 | |
| V. SFTBackbone=Qwen3-4B-Base2026.06 | 55.9 | |
| CAGradBackbone=Llama-3.1-8B2026.06 | 45.6 | |
| MGDABackbone=Meta-Llama-3-8B2026.06 | 45.5 | |
| PCGradBackbone=Meta-Llama-3-8B2026.06 | 45.2 | |
| MGDABackbone=Llama-3.1-8B2026.06 | 45.2 | |
| PCGradBackbone=Llama-3.1-8B2026.06 | 44.8 | |
| V. SFTBackbone=Llama-3.1-8B2026.06 | 44.2 | |
| CAGradBackbone=Meta-Llama-3-8B2026.06 | 43.9 | |
| V. SFTBackbone=Meta-Llama-3-8B2026.06 | 42.5 | |
| BaseBackbone=Llama-3.1-8B2026.06 | 41.8 | |
| BaseBackbone=Meta-Llama-3-8B2026.06 | 37.2 |