Addition on 10-digit addition 3 terms
89.56Token AccuracyDUAL-STREAM PROBE
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| DUAL-STREAM PROBEModel=Qwen3-4B, Delta (δ)=0.12026.05 | 89.56 | 32.4 | 6.46 | 30.46 | 98.13 | |
| REPLACEMENTModel=Qwen3-4B2026.05 | 89.13 | 30 | 7.1 | 31.73 | 97.65 | |
| DUAL-STREAM PROBEModel=Qwen3-4B, Delta (δ)=0.052026.05 | 88.87 | 29 | 8.44 | 35.51 | 96.75 | |
| STEERINGModel=Qwen3-4B2026.05 | 88.27 | 28.2 | 7.75 | 30.58 | 96.97 | |
| DUAL-STREAM PROBEModel=Qwen3-4B, Delta (δ)=02026.05 | 87.27 | 23.2 | 12.4 | 44.39 | 94.07 | |
| ORIGINALModel=Qwen3-4B2026.05 | 86.26 | 28.1 | — | — | — | |
| DUAL-STREAM PROBEModel=Gemma-3-4B-IT, Delta (δ)=0.052026.05 | 85.98 | 37.3 | 9.67 | 32.64 | 98.37 | |
| DUAL-STREAM PROBEModel=Gemma-3-4B-IT, Delta (δ)=02026.05 | 85.86 | 34.7 | 12.97 | 40.46 | 96.43 | |
| REPLACEMENTModel=Gemma-3-4B-IT2026.05 | 84.99 | 37.8 | 10.85 | 34.05 | 97.32 | |
| DUAL-STREAM PROBEModel=Gemma-3-4B-IT, Delta (δ)=0.12026.05 | 84.92 | 35.8 | 8.41 | 27.26 | 98.97 | |
| STEERINGModel=Gemma-3-4B-IT2026.05 | 82.1 | 30.1 | 14.17 | 33.19 | 95.31 | |
| RE-PROMPTINGModel=Qwen3-4B2026.05 | 79.9 | 26.1 | 10.4 | 0.08 | 99.98 | |
| ORIGINALModel=Gemma-3-4B-IT2026.05 | 72.86 | 31.7 | — | — | — | |
| RE-PROMPTINGModel=Gemma-3-4B-IT2026.05 | 71.41 | 32.1 | 10.65 | 17.57 | 91.33 |