New GraphRAG framework aligns LLM agents with human social values

A new paper accepted at CogSci 2026 tackles a critical gap in LLM-based agents: their inability to consistently align with human social values in complex dilemmas. The authors propose a prescriptive framework that moves beyond descriptive alignment (simply reflecting training data) to actively steer agent behavior. At its core, the method uses GraphRAG (graph-based retrieval-augmented generation) to encode abstract social principles from psychology theories—specifically Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion—into concrete, context-aware instructions. When an agent encounters a conversation, it retrieves the most relevant value instruction from a structured knowledge graph, guiding its response toward expected social behaviors.

The researchers evaluated their method on the DailyDilemmas benchmark, a suite of everyday moral conflicts. Compared to prompt-based baselines such as ECoT, Plan-and-Solve, and Metacognitive prompting, the GraphRAG-driven approach achieved significant gains in aligning agent outputs with predefined value expectations. Importantly, the framework also provides a foundation for the emergence of self-emotion in AI—a step toward agents that can recognize and appropriately express emotional states in social interactions. This work represents a shift from describing what values look like to prescribing how agents should act, with implications for deploying AI in ethically sensitive domains like healthcare, customer service, and autonomous negotiation.