3 ms·
The alignment has certainly become stronger though. Llama 3.1 is trivial to decensor with abliteration and Heretic's optimizer will rapidly converge to paramete
by p-e-w 11mo ago
The alignment has certainly become stronger though. Llama 3.1 is trivial to decensor with abliteration and Heretic's optimizer will rapidly converge to parameters that completely stomp out refusals, while for gpt-oss and Qwen3, most parameter configurations barely have an effect and it takes much longer to reach something that even slightly lowers the refusal rate.
- shikon7 11mo agoIt seems to me that thinking models are harder to decensor, as they are trained to think whether to accept your request.
- int_19h 11mo agoIt goes both ways. E.g. unmodified thinking Qwen is actually easier to jailbreak to talk about things like Tiananmen by convincing it that it is unethical to refuse to do so.