Reproduce “Multi-level Value Alignment in Agentic AI Systems: Survey and Perspectives” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wei Zeng, Hengshu Zhu, Chuan Qin, Han Wu, Yihang Cheng, Sirui Zhang, Xiaowei Jin, Yinuo Shen and 3 more

A comprehensive survey details a 'Multi-level Value framework' for understanding and addressing value alignment in LLM-based multi-agent (agentic AI) systems from a socio-governance perspective. The framework systematically categorizes value principles, application scenarios, methods, and evaluation, revealing the dynamic and hierarchical nature of values and the need for improved coordination and data resources.

Key insights

  • Values are multi-level and dynamic, structured hierarchically (macro, meso, micro), and are continuously interpreted and reshaped, requiring agents to internalize them through design rather than being static rules.
  • The complexity of value alignment directly correlates with a system's generalizability; highly generalizable systems need macro-level alignment, while specialized systems require simultaneous adherence to macro-, meso-, and micro-level norms.
  • In multi-agent systems, the challenge shifts from aligning individual agents to coordinating values among multiple interacting entities, necessitating robust interaction mechanisms and organizational models to resolve conflicts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.