Researchers have developed a new framework called the Preference-Behavior Model (PBM) to better understand how large language models (LLMs) make decisions in complex scenarios. This model combines an LLM's stated preferences with mechanisms that account for imperfect utility maximization, allowing for joint inference of preferences and behavioral parameters from observed choices. Experiments in social dilemmas like the Public Goods Game and Collective-Risk Dilemma showed that while standard elicitation instruments did not consistently predict LLM behavior, the PBM demonstrated predictive value across different dilemmas. The PBM also revealed that LLM preferences generally align with persona instructions but can diverge across dimensions, models, and specific games. AI
IMPACT Provides a framework for understanding and aligning LLM decision-making in complex, multi-objective scenarios.
RANK_REASON Academic paper introducing a new modeling framework for LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Collective-Risk Dilemma
- Common-Pool Resource dilemma
- large-language models
- Preference-Behavior Model
- Public Goods Game
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
