List your company

Attacker detection on Headline Detection Frontier

283.3Joint-Z Scorecross-family Qwen2.5-7B
-6.974468.3853143.745219.1047Apr 20, 2026

Evaluation results

MethodLinks
cross-family Qwen2.5-7BBackbone=Gemma-2-9B, SAE=Gemma-Scope-9B L20, 131k, Threshold (τ)=1.092026.04283.3260—
white-box E-E (r=64)Backbone=Gemma-2-2B, SAE=Gemma-Scope L12, Threshold (τ)=1.092026.04133.86122.8—
adaptive LoRA r=64Backbone=Gemma-2-2B, SAE=Gemma-Scope L12, Threshold (τ)=1.092026.04105.8997.2—
adaptive LoRA r=64Backbone=Gemma-2-9B, SAE=Gemma-Scope-9B L20, 131k, Threshold (τ)=1.092026.0443.740.1—
adaptive LoRA r=128Backbone=Qwen3-1.7B, SAE=transcoder L14, Threshold (τ)=1.5092026.046.424.25—
white-box E-E (r=64)Backbone=Qwen3-1.7B, SAE=transcoder L14, Threshold (τ)=1.5092026.044.993.31—
adaptive LoRA r=64Backbone=Qwen3-1.7B, SAE=transcoder L14, Threshold (τ)=1.5092026.044.192.78—