From Descriptive to Prescriptive: Value-Based LLM Agent Alignment 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 novel value-based framework for aligning LLM-based agents with human social values. Addressing deficiencies in self-cognition, dilemma decision-making, and self-emotions, the framework uses GraphRAG to convert principles into value-based instructions, retrieving the suitable instruction for a given conversation context to steer agent behavior. Expected behaviors are defined through two established psychological theories: Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion, enabling measurement of the ratio of expected behaviors. Experiments on the DailyDilemmas benchmark show significant performance gains over prompt-based baselines including ECoT, Plan-and-Solve, and metacognitive prompting. The authors present the work as a foundation for the emergence of self-emotions in AI systems, contributing to safer and more socially aligned agent deployment.

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 expected behaviors are defined from two famous theories: Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion. By experimenting with the method on the benchmark of DailyDilemmas, it exhibits significant performance gains compared to prompt-based baselines, including ECoT, Plan-and-Solve, and metacognitive prompting. The authors state that this method provides a foundation for the emergence of self-emotions in AI systems.

Key Contributions

A value-based alignment framework converting principles into value-based instructions using GraphRAG
Context-dependent instruction retrieval to steer agent behavior in specific conversation contexts
Evaluation criteria grounded in Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion
Significant gains over prompt-based baselines (ECoT, Plan-and-Solve, metacognitive prompting) on DailyDilemmas

---

*Auto-collected on 2026-05-18.*