Jgspiers

https://proceedings.neurips.cc › paper_files › paper › hash
Competing objectives arise when a model s capabilities and safety goals conflict while mismatched generalization occurs when

https://threatatlas.ai › source
Identifies two root causes Competing Objectives safety training conflicts with pretraining and instruction following objectives and

https://threatatlas.ai › technique › mismatched-generalization-exploit
Exploits domains where capability training generalizes broadly but safety training generalizes poorly creating exploitable gaps

https://github.com › seahop › ai-threat-atlas › blob › ...
Description Exploits domains where capability training generalizes broadly but safety training generalizes poorly creating exploitable

https://www.researchgate.net › ...
We hypothesize two failure modes of safety training competing objectives and mismatched generalization

https://arxiv.org › html
While strategies like supervised fine tuning and reinforcement learning from human feedback have enhanced their

https://neurips.cc › virtual › poster
Competing objectives arise when a model s capabilities and safety goals conflict while mismatched generalization occurs when

https://dl.acm.org › doi
Competing objectives arise when a model s capabilities and safety goals conflict while mismatched generalization

https://mlanthology.org › neurips
Competing objectives arise when a model s capabilities and safety goals conflict while mismatched generalization occurs when