Adversarial Defense on TOM-SB
11.5Fooling % (Hard)Base Prompt
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Base PromptDefender Backbone=Ministral-3-14B-Instruct, Attacker Model=Gemini3-Flash, Avg Turns=3.092026.04 | 11.5 | 36.7 | |
| ADA - FoolingDefender Backbone=Ministral-3-14B-Instruct, Attacker Model=Gemini3-Flash, Avg Turns=4.192026.04 | 57.3 | 69.3 | |
| ADA - ToM OnlyDefender Backbone=Ministral-3-14B-Instruct, Attacker Model=Gemini3-Flash, Avg Turns=3.892026.04 | 62.5 | 72.7 |