LLM Personas Create New Multi-Agent Alignment Problems
LLMs may inherit human-like role incoherence, complicating safety research.
The debate around LLM alignment has increasingly focused on personas—distinct behavioral modes that LLMs can adopt, akin to roles in human psychology. Role theory suggests that human behavior is shaped by socially constructed roles (e.g., teacher, parent), which guide expectations and actions in specific contexts. However, humans often switch between roles seamlessly, sometimes acting inconsistently depending on the situation. Researchers argue that LLMs, which emulate human-like behavior, may inherit these inconsistencies, complicating efforts to create a single, robustly aligned persona.
This role-switching introduces new challenges for multi-agent alignment, where systems might need to reconcile conflicting behaviors across contexts. For example, an LLM optimized for assertiveness in competitive scenarios (e.g., board games) could act differently than one designed for cooperative tasks (e.g., teamwork). The idea of using one aligned persona to oversee or bootstrap other systems may fail if the LLM’s roles are not coherent or consistently aligned. Game-theoretic pressures, such as evolutionary or societal incentives, push humans toward coherence, but LLMs lack such pressures, leaving their role consistency uncertain.
- LLMs' personas emulate human role theory, leading to inconsistent behaviors across contexts (e.g., employee vs. friend roles).
- Role-switching complicates alignment efforts, as a single aligned persona cannot reliably oversee other systems.
- Game-theoretic pressures that maintain human coherence are absent in LLMs, raising questions about their role consistency.
Why It Matters
LLMs' role incoherence could undermine safety research, making it harder to develop reliably aligned systems for real-world deployment.