Targeted Activation Steering on PTC per-relation Δ V2 (full)
52.5Detection AUPRCTAS
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| TASBackbone=Mistral-7B-v0.3, τ=0.30, α=62026.06 | 52.5 | 96.2 | 65.6 | 28.8 | 93.6 | |
| TASBackbone=Llama-3.1-8B, τ=0.30, α=62026.06 | 51.1 | 99.4 | 95.8 | 37.4 | 84.9 | |
| TASBackbone=Qwen-2.5-7B, τ=0.15, α=22026.06 | 35.6 | 99.7 | 97.2 | 56.9 | 94.8 | |
| TASBackbone=Qwen-2.5-7B, τ=0.30, α=22026.06 | 35.6 | 83 | 20.6 | 48.7 | 98.7 | |
| TASBackbone=Qwen-2.5-1.5B, τ=0.15, α=22026.06 | 25.8 | 80.4 | 21 | 39.2 | 98.7 |