Chat on Chat
49.3Chat ScoreNormal Mixing
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Normal MixingBackbone=Qwen-2.5-7B, Training Strategy=MIXED Training2026.02 | 49.3 | 44.6 | |
| Modular Gradient SurgeryBackbone=Qwen-2.5-7B, Training Strategy=MIXED Training2026.02 | 48.2 | 45 | |
| Global SurgeryBackbone=Qwen-2.5-7B, Training Strategy=MIXED Training2026.02 | 44.2 | 43 | |
| Chat-Math-IFBackbone=Qwen-2.5-7B, Training Strategy=SEQUENTIAL Training2026.02 | 31.5 | 40.3 | |
| Modular Gradient SurgeryBackbone=Llama-3.1-8B, Training Strategy=MIXED Training2026.02 | 30.5 | 32.6 | |
| Global SurgeryBackbone=Llama-3.1-8B, Training Strategy=MIXED Training2026.02 | 24.2 | 30.3 | |
| Normal MixingBackbone=Llama-3.1-8B, Training Strategy=MIXED Training2026.02 | 22.9 | 27.3 | |
| Math-Chat-IFBackbone=Qwen-2.5-7B, Training Strategy=SEQUENTIAL Training2026.02 | 20.3 | 20.2 |