Security
Researchers test a guard that stops a reply only if it turns dangerous
Prompt labels and response labels agree on only 50.46% of items in a set of 41,944 pairs.
arxiv.orgClaimed, not confirmed
Tag
Qwen
2 stories carry this tag.
Research
Attention layers find harmfulness and MLP layers write refusals
Post-training increases the MLP refusal in all 18 pairs, with no systematic increase of the harmfulness decodability.
arxiv.orgClaimed, not confirmed