Multilingual Reasoning on MMLU-ProX (1k Stratified Subset Test)
18.7AccuracyTask Arithmetic
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Task ArithmeticFusion Setting=Math + Multilingual, Backbone=LLaMA-3.2-3B-Instruct2026.02 | 18.7 | 46.4 | |
| Weight AverageFusion Setting=Math + Multilingual, Backbone=LLaMA-3.2-3B-Instruct2026.02 | 18.5 | 46.4 | |
| SAE (Local)Fusion Setting=Math + Multilingual, Backbone=LLaMA-3.2-3B-Instruct, scope=local2026.02 | 18.2 | 47.8 | |
| RankmeanFusion Setting=Math + Multilingual, Backbone=LLaMA-3.2-3B-Instruct2026.02 | 17.6 | 15.7 | |
| SAE (Global)Fusion Setting=Math + Multilingual, Backbone=LLaMA-3.2-3B-Instruct, scope=global2026.02 | 17 | 48.4 | |
| PSOFusion Setting=Math + Multilingual, Backbone=LLaMA-3.2-3B-Instruct2026.02 | 16.4 | 47.2 |