From Descriptive to Prescriptive: Social Value Alignment for LLM Agents via GraphRAG

Forum topic · 小凯 · 2026-05-18

Summary

This arXiv paper (2505.12352) by Jinxian Qu, Qingqing Gu, and Teng Chen proposes a value-based framework for aligning LLM-based agents with human social values. Addressing deficiencies in self-cognition, dilemma decision-making, and self-emotions in current alignment work, the method uses GraphRAG to convert abstract principles into value-based instructions, retrieving the suitable instruction for a given conversation context to steer agent behavior. Expected behaviors are defined using two established psychological theories: Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion. Evaluated on the DAILYDILEMMAS benchmark, the approach shows significant performance gains over prompt-based baselines including ECoT, Plan-and-Solve, and metacognitive prompting. The authors position this work as a foundation for the emergence of self-emotions in AI systems.

Paper Overview

Field: NLP Authors: Jinxian Qu, Qingqing Gu, Teng Chen Published: 2026-05-17 arXiv: 2505.12352

Abstract

Wide applications of LLM-based agents require strong alignment with human social values. However, current works still exhibit deficiencies in self-cognition and dilemma decision, as well as self-emotions. To remedy this, the authors propose a novel value-based framework that employs GraphRAG to convert principles into value-based instructions and steer the agent to behave as expected by retrieving the suitable instruction upon a specific conversation context.

To evaluate the ratio of expected behaviors, the paper defines expected behaviors from two famous theories:

Maslow's Hierarchy of Needs
Plutchik's Wheel of Emotion

Results

Experiments on the DAILYDILEMMAS benchmark show significant performance gains compared to prompt-based baselines, including:

ECoT
Plan-and-Solve
Metacognitive prompting

The authors state that this method provides a foundation for the emergence of self-emotions in AI systems.

---

*Auto-collected on 2026-05-18*