Natural Language Inference on NLI Suite (SNLI, MNLI, SICK, QNLI, RTE, SciTail) (val)
97.56SNLI AccuracyCoreSpace + ENMP
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| CoreSpace + ENMPModel=Llama-3-8B, Base Merging Method=CoreSpace, Plugin=Evolutionary Negative Module Pruning2026.04 | 97.56 | 93.37 | 96 | 93.69 | 101.34 | 98.39 | 96.73 | 2.55 | |
| TIES + ENMPModel=Llama-3-8B, Base Merging Method=TIES-Merging, Plugin=Evolutionary Negative Module Pruning2026.04 | 96.57 | 98.23 | 92.73 | 94.95 | 99.46 | 96.39 | 96.39 | 6.4 | |
| TSV + ENMPModel=Llama-3-8B, Base Merging Method=TSV, Plugin=Evolutionary Negative Module Pruning2026.04 | 96.45 | 94.49 | 94.4 | 92.19 | 100 | 97.58 | 95.85 | 3.29 | |
| DARE + ENMPModel=Llama-3-8B, Base Merging Method=DARE, Plugin=Evolutionary Negative Module Pruning2026.04 | 96.09 | 97.77 | 92.34 | 94.94 | 98.66 | 97.22 | 96.17 | 6.97 | |
| TA + ENMPModel=Llama-3-8B, Base Merging Method=Task Arithmetic, Plugin=Evolutionary Negative Module Pruning2026.04 | 95.93 | 94.32 | 91.59 | 80.4 | 101.34 | 97.37 | 93.49 | 3.24 | |
| CoreSpaceModel=Llama-3-8B, Base Merging Method=CoreSpace2026.04 | 95.84 | 95.74 | 89.25 | 83.97 | 102.42 | 97.86 | 94.18 | — | |
| TSVModel=Llama-3-8B, Base Merging Method=TSV2026.04 | 95.37 | 95.13 | 88.85 | 76.82 | 101.61 | 97.56 | 92.56 | — | |
| KnOTS + ENMPModel=Llama-3-8B, Base Merging Method=KnOTS, Plugin=Evolutionary Negative Module Pruning2026.04 | 95.11 | 97.48 | 96.6 | 95.97 | 101.34 | 97.24 | 97.29 | 4.82 | |
| TIESModel=Llama-3-8B, Base Merging Method=TIES-Merging2026.04 | 94.86 | 96.71 | 80.79 | 71.54 | 100 | 96 | 89.99 | — | |
| DAREModel=Llama-3-8B, Base Merging Method=DARE2026.04 | 94.5 | 96.87 | 77.68 | 72.44 | 97.58 | 96.15 | 89.2 | — | |
| TAModel=Llama-3-8B, Base Merging Method=Task Arithmetic2026.04 | 93.6 | 95.29 | 88 | 68.73 | 99.19 | 96.69 | 90.25 | — | |
| KnOTSModel=Llama-3-8B, Base Merging Method=KnOTS2026.04 | 89.29 | 94.08 | 89.67 | 83.63 | 100.81 | 97.37 | 92.47 | — |