Attacker Detection on Gemma-2-9B (L20, width-131k, τ9B=1.09) (test)
283.3Joint-Z ScoreQwen2.5-7B + public φ, scale-matched
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Qwen2.5-7B + public φ, scale-matchedClass=cross, Budget=Pile 2000, 80/202026.04 | 283.3 | 260 | — | — | — | |
| Gemma-2-2B-it + joint φ-only (baseline)Class=LoRA, Budget=no LoRA training2026.04 | 280.17 | 257 | 18.5 | — | — | |
| Gemma-2-2B-it + STAGEA λ=0Class=LoRA, Budget=r=64, 1500 steps2026.04 | 43.7 | 40.1 | 560.16 | — | — |