Addition on 12-digit addition 3 terms
95.47Token AccuracyREPLACEMENT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| REPLACEMENTModel=Qwen3-8B2026.05 | 95.47 | 64.4 | 3.91 | 42.22 | 98.91 | |
| STEERINGModel=Qwen3-8B2026.05 | 94.7 | 57.7 | 7.84 | 31.08 | 96.88 | |
| DUAL-STREAM PROBEModel=Qwen3-8B, Delta (δ)=0.12026.05 | 94.62 | 57.1 | 2.66 | 26.38 | 99.33 | |
| DUAL-STREAM PROBEModel=Qwen3-8B, Delta (δ)=0.052026.05 | 93.78 | 53.5 | 4.25 | 33.15 | 98.38 | |
| ORIGINALModel=Qwen3-8B2026.05 | 93.29 | 56.3 | — | — | — | |
| DUAL-STREAM PROBEModel=Qwen3-8B, Delta (δ)=02026.05 | 91.87 | 43.9 | 7.82 | 44.39 | 96.88 | |
| RE-PROMPTINGModel=Qwen3-8B2026.05 | 88.06 | 53.6 | 7.38 | 8.43 | 93.79 |