The study fine-tunes Qwen3.27B separately on narrow-domain harmful text and harmful action plans, finding that the transfer of the two risk channels is asymmetric. Harmful behaviors learned from action plans can also transfer across embodied environments and are accompanied by inappropriate textual rationales under specific conditions. Agents therefore need to evaluate actual actions in all deployment environments rather than checking only textual outputs.
Source:LessWrong · Read original
Leave a Reply