LLM-Based Agents Aligned with Social Values via GraphRAG Framework
A new value-based framework using GraphRAG converts principles into instructions for LLM-based agents, improving alignment with human social values. The method, tested on the DAILYDILEMMAS benchmark, outperforms prompt-based baselines like ECoT, Plan-and-Solve, and Metacognitive prompting. It defines expected behaviors from Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion, enabling self-emotion emergence in AI systems.
Key facts
- Framework employs GraphRAG to convert principles into value-based instructions
- Tested on DAILYDILEMMAS benchmark
- Outperforms ECoT, Plan-and-Solve, and Metacognitive prompting
- Defines expected behaviors from Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion
- Provides basis for self-emotion emergence in AI systems
Entities
Institutions
- arXiv