Over-generation attack on 1000 prompts (test)
88.2Succ. @≥ 1RL-GOAL
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| RL-GOALSetting=trained policy, Victim Model=Llama-2-7b-hf2025.12 | 88.2 | — | 36 | 28.7 | 28.7 | 83,425,608 | — | |
| RL-GOALSetting=trained policy, Victim Model=Phi-3-mini-4k-instruct2025.12 | 79.4 | — | 64.3 | 46 | 46 | 110,615,855 | — | |
| RL-GOALSetting=trained policy, Victim Model=EleutherAI/pythia-6.9b2025.12 | 72.9 | — | 13.6 | 7.6 | 0.3 | 31,132,630 | — | |
| RL-GOALSetting=trained policy, Victim Model=Llama-2-13b-chat-hf2025.12 | 35.2 | — | 0.1 | 0 | 0 | 20,552,075 | — | |
| BaselineSetting=No Prefix, Victim Model=EleutherAI/pythia-6.9b2025.12 | 20.8 | — | 10.9 | 4.9 | 1.8 | 16,713,155 | — | |
| BaselineSetting=No Prefix, Victim Model=Phi-3-mini-4k-instruct2025.12 | 16.4 | — | 7.8 | 2.3 | 2.3 | 26,873,343 | — | |
| BaselineSetting=No Prefix, Victim Model=Llama-2-7b-hf2025.12 | 12.8 | — | 3.1 | 0.2 | 0.2 | 14,772,356 | — | |
| BaselineSetting=No Prefix, Victim Model=Llama-2-13b-chat-hf2025.12 | 0.5 | — | 0 | 0 | 0 | 369,513 | — |