Search papers, labs, and topics across Lattice.
Korea University, Georgia Institute of Technology
3
0
5
Non-English users pay a significantly higher price for safety alignment in AI models, revealing systemic inequities in current practices.
Aligning soft-token representations across languages can boost multilingual reasoning accuracy by over 17 points, especially for low-resource languages.
HarmAmp reveals that LLMs can amplify harm in multi-turn interactions, while TrajSafe effectively curbs this risk without sacrificing performance.
