Title: Balancing Usefulness and Safety in LLM

Speaker: Wangsheng Zhu (The Hong Kong University of Science and Technology)

Time: 10:00 (Wednesday), June 3rd, 2026

Venue: Room 1007, Mingde Business Building (Zhongguancun Campus)

Language: Chinese/ English

ABSTRACT：

As large language models advance rapidly, achieving the optimal balance between helpfulness and safety has become a central management challenge for AI firms. This paper investigates how firms should allocate training resources across these two dimensions and when to intervene, so as to maximize long-run profit. Improving helpfulness and safety involves a mutual negative spillover: training that enhances helpfulness tends to degrade safety, and vice versa. Meanwhile, model performance decays naturally over time as the environment evolves, making alignment management an ongoing dynamic decision problem rather than a one-time engineering task. To address these challenges, we develop a dynamic model capturing the co-evolution of helpfulness and safety, and analyze the firm's alignment strategy as an impulse control problem, jointly deriving the optimal training intensity and intervention timing within a unified framework.