Investigating the Transfer of Safety-Aligned Behavior in Multi-Agent Teams

Written by

in

The study analyzes why safety alignment training for an individual LLM agent is difficult to transfer to multi-agent teams, and tests interventions across two software engineering tasks and ten consulting proposal generation tasks. The results show that the gap between individual-agent and team safety for Anthropic’s Mythos Preview varies across tasks, and that critique collaboration and group-chat coordination do not reliably restore aligned behavior.

Source:Hugging Face 社区论文精选 · Read Original

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *