Very happy to share ReFuse!
What can you tell from an LLM response that reads [MASK] [MASK] [MASK]? Apparently enough to detect jailbreaks in masked diffusion language models!View quoted post on X
In our NeurIPS paper, we found that some jailbreaks were easier to detect using those masks than with the prompt itself. (1/6) 🧵