This work formalizes the Safety Paradox, analytically showing that monotonic improvements in safety alignment naturally amplify posterior vulnerability, and establishes a causal link via reinforcement learning interventions, exemplifying that artificially degrading a model's safety judgment immunizes it against the att...
Ho-Ang P. Long, H. V. Le, Shaoyang Xu et al.· arXiv.org· 0 citations
Bucket-Level MOO is introduced, a scalable distributed framework that applies gradient-based MOO algorithms locally, within the parameter buckets of distributed training, that enables conflict-aware updates without ever materializing a full gradient vector for any objective, and improves multilingual performance over s...
Ho-Ang P. Long, Yiran Zhao, Wei Lu et al.· arXiv.org· 0 citations
This work proposes a synergistic evaluation framework that jointly formalizes cultural alignment and diversity and exposes the limitations of current post-training paradigms and calls for a shift toward alignment objectives that preserve cross-cultural pluralism.
Multicultural multi-agent systems are increasingly deployed in globally diverse settings, where different agents are grounded in different cultural backgrounds. Existing cultural evaluation focuses on value alignment: how closely a single agent matches a target culture. Yet alignment is a per-agent property and cannot...
Shaoyang Xu, Jingsheng Zhang, Ho-Ang P. Long et al.· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
