Token-level Addition on Addition T3 (val)
89.56Token AccuracyDual-stream consistency check method
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Dual-stream consistency check methoddelta (δ)=0.12026.05 | 89.56 | 30.46 | 98.13 | |
| REPLACEMENTtype=hard replacement, probe=Ground Truth2026.05 | 89.13 | 31.73 | 97.65 | |
| STEERINGtype=linear steering2026.05 | 88.27 | 30.58 | 96.97 | |
| Dual-stream consistency check methoddelta (δ)=02026.05 | 87.27 | 44.39 | 94.07 | |
| ORIGINALmode=baseline2026.05 | 86.26 | — | — | |
| RE-PROMPTINGmode=intervention2026.05 | 79.9 | 0.08 | 99.98 |