Concept Suppression on COCO (test)
40.5Concepts Removed (%)CSAE
Evaluation Results
| Method | Links | |
|---|---|---|
| CSAEModel=Qwen3-VL-4B-Instruct, LLM Judge=Gemini-2.5-Flash, Evaluation Scale=1,000 evaluations, Steering scale=-3us2026.06 | 40.5 | |
| SingleModel=Qwen3-VL-4B-Instruct, LLM Judge=Gemini-2.5-Flash, Evaluation Scale=1,000 evaluations, Steering scale=-3us, Description=Best individual CSAE Level-1 concept2026.06 | 35.6 | |
| MSAEModel=Qwen3-VL-4B-Instruct, LLM Judge=Gemini-2.5-Flash, Evaluation Scale=1,000 evaluations, Steering scale=-3us2026.06 | 25.2 | |
| Rand.Model=Qwen3-VL-4B-Instruct, LLM Judge=Gemini-2.5-Flash, Evaluation Scale=1,000 evaluations, Steering scale=-3us2026.06 | 12.4 | |
| NSModel=Qwen3-VL-4B-Instruct, LLM Judge=Gemini-2.5-Flash, Evaluation Scale=1,000 evaluations, Steering scale=none2026.06 | 0 |