Majority Vote on Majority-vote 432 prompts (test)
100AccuracySoft DPS
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Soft DPSlogit offset α=52026.05 | 100 | — | — | |
| Soft DPSlogit offset α=102026.05 | 100 | — | — | |
| Soft DPSlogit offset α=502026.05 | 100 | — | — | |
| Baseline (vanilla Qwen3-8B)Condition=vanilla Qwen3-8B2026.05 | 51.4 | — | — | |
| Random directionlogit offset α=10, seeds=202026.05 | 49.4 | — | — | |
| Best probe (layer 2)layer=22026.05 | — | 0.003 | 0.9998 |