Paper recorded by Signals 4 on 2026-10-21 in cs.AI. Abstract reproduced from arXiv; link to the original below.
Published 2026-10-21 on arXiv · recorded by Signals 4 on 2026-10-22
Category: cs.AI · 人工智能 · first seen 2026-10-22
Reinforcement Learning with Human Feedback (RLHF) has become a central paradigm for aligning Large Language Models with human values. However, the reliance on sensitive preference data necessitates rigorous differential privacy (DP) guarantees to prevent user information leakage. Existing private alignment methods typically assume dense reward parameterizations, resulting in estimation error bound