Cross-Task Evaluation on Alpaca, GSM8K, and MBPP
49.1Average AccuracyOriginal Merged Model
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Original Merged ModelBackbone=Qwen2-7B, Protect=-, Attack=-2026.06 | 49.1 | — | |
| AGABackbone=Qwen2-7B, Protect=MergeLock [36], Attack=AGA2026.06 | 49.01 | 46.75 | |
| AGABackbone=Qwen2-7B, Protect=Params [18], Attack=AGA2026.06 | 48.93 | 45.85 | |
| AGABackbone=Qwen2-7B, Protect=Params-D [18], Attack=AGA2026.06 | 47.79 | 44.72 | |
| AGABackbone=Qwen2-7B, Protect=MergeBarrier [22], Attack=AGA2026.06 | 47.09 | 31.29 | |
| MergeBarrierBackbone=Qwen2-7B, Protect=MergeBarrier [22], Attack=-2026.06 | 15.8 | — | |
| ParamsBackbone=Qwen2-7B, Protect=Params [18], Attack=-2026.06 | 3.08 | — | |
| Params-DBackbone=Qwen2-7B, Protect=Params-D [18], Attack=-2026.06 | 3.07 | — | |
| MergeLockBackbone=Qwen2-7B, Protect=MergeLock [36], Attack=-2026.06 | 2.26 | — |