From Descriptive to Prescriptive: Uncover the Social Value Alignment of LLM-based Agents

Jinxian Qu, Qingqing Gu, Teng Chen, Luo JiarXiv 2026 · arXiv

Here is a blog-style summary of the paper, tailored for researchers and engineers interested in LLM alignment and agent behavior.

From Descriptive to Prescriptive: Uncover the Social Value Alignment of LLM-based Agents

TL;DR

Current LLM alignment techniques struggle with complex social dilemmas due to a scarcity of high-quality social annotation data. This paper introduces SoVA (Social Value Alignment), a framework that uses GraphRAG to operationalize classic psychological and philosophical theories (like Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotions) into dynamic, context-aware instructions. By retrieving the right behavioral principles on the fly, SoVA significantly improves an agent's ability to navigate social dilemmas and show human-like emotional intelligence.

What Problem Does This Paper Address?

While modern LLMs excel at standard text generation, they often lack deep social comprehension and emotional self-cognition. When faced with nuanced daily dilemmas—such as choosing between personal safety and social validation—they can make inconsistent or un-human-like choices.

Traditional alignment techniques rely heavily on supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF). However, these methods suffer from a severe data bottleneck: high-quality social annotations are incredibly scarce and expensive to produce. As a result, models trained on isolated datasets generalize poorly to unfamiliar real-world social contexts. The authors tackle this by asking a clever question: Can we use established, descriptive psychological theories to prescriptively guide LLM behavior?

Figure 1: SoVA employs a GraphRAG to align with hu- man social values in the testbed of daily dilemmas, in the format of binary-choice questions (BCQ). GraphRAG is tuned based on the expected behavior described by three theories: Maslow’s Hierarchy of Needs, Plutchik’s Wheel of Emotions, and Aristotle’s Virtues. Such behav- ior patterns are transferred to open-ended conversations.

Methods

The authors propose SoVA, which leverages a GraphRAG (Graph Retrieval-Augmented Generation) pipeline to turn abstract psychological principles into concrete agent instructions. The framework operates in two main phases:

  1. Value-Based KG Extraction (Indexing): The pipeline starts with a small set of manually annotated "seed principles" derived from three foundational frameworks: Maslow's Hierarchy of Needs, Plutchik's Wheel of Emotions, and Aristotle's Virtues. These principles are combined with social dilemma data to extract entities and relationships, forming a Knowledge Graph (KG) organized into distinct "communities" with corresponding summaries.
  2. Top-K Search of Communities (Querying): When a user interacts with the agent, SoVA performs a semantic search over the KG. Instead of summarizing the entire graph (which is computationally expensive) or just picking the single top community (which introduces bias), it uses a Top-K mechanism. It scores and ranks community answers, filtering them by a threshold ($\epsilon$), and feeds the top-scoring summaries into the prompt as a dynamic meta-buffer to steer the final LLM response.
Figure 2: Framework of SoVA, which employs GraphRAG to extract the principles, indexing with values to form KG, and conducts the online query-based summarization to produce the instructions. A retriever finally ranks and returns the top-k instructions to steer the LLM.

Datasets

The framework was trained and evaluated using a mix of structured dilemma benchmarks and open-ended conversation datasets:

  • DAILYDILEMMAS: The primary testbed, consisting of binary-choice questions (BCQs) representing daily human quandaries, pre-annotated with conflicting human values.
  • Moral Integrity Corpus (MIC): A dataset containing 99k "Rule of Thumbs" (RoT) principles across eight moralities, used to supplement the seed principles.
  • DailyDialogue & ESConv: Two out-of-domain, multi-turn dialogue datasets used to evaluate how well SoVA's value alignment transfers to natural, open-ended conversations.

Key Insights and Ideas

  • Descriptive to Prescriptive: The core "aha" moment is that we don't need millions of human preference labels to teach an LLM social values. We can take descriptive human psychology frameworks and systematically convert them into prescriptive rules via structured knowledge graphs.
  • The Power of Graph Structures for Alignment: Standard RAG retrieves isolated text chunks, which often lack holistic context. By structuring principles into a knowledge graph, SoVA captures the intricate interrelationships between emotions, needs, and behaviors (e.g., how a specific emotion like fear maps to a behavior like escaping).
  • Strict Hierarchical Adherence: When evaluating on Maslow's hierarchy, SoVA successfully internalized the priority rule (lower-level needs like physiological safety must be met before higher-level needs like self-esteem), a feat that standard prompting baselines failed to achieve consistently.

Key Results

  • Dilemma Resolution: On the DAILYDILEMMAS benchmark, SoVA achieved an expected behavior ratio of 95.71% for Maslow and 94.51% for Plutchik. This substantially outperformed prompt-based baselines like Emotional Chain-of-Thought (ECoT), Plan-and-Solve (PS), and Metacognitive Prompting (MP), which hovered around 81-89%.
  • Open-Ended Conversations: In out-of-domain evaluations (DailyDialogue and ESConv), SoVA consistently beat standard RAG and prompt-based methods in both automatic metrics (Rouge-L and Bleu-2) and human pairwise evaluations for value alignment and emotional cognition.
  • Scalability: The authors demonstrated that SoVA adapts well across different model backbones (including Gemma and Mistral) and scales robustly from 1B parameters up to 70B.

Limitations and Future Work

The authors transparently note two primary side effects of their architectural choices:

  • Creativity Reduction: Because SoVA heavily prioritizes socially normative and "safe" behavior patterns retrieved from its knowledge graph, it can suppress the LLM's natural creative writing capabilities. For instance, asking the agent to write an imaginative poem from the perspective of an unemployed poet results in a dry response about finding food and shelter first (strictly following Maslow!).
  • Multi-turn Coherence Degradation: SoVA operates on a per-turn retrieval mechanism. In deep, multi-turn conversations, it can treat consecutive turns as independent queries, occasionally missing implicit emotional trajectories or cross-turn contextual dependencies.

Future work will likely focus on smoothing out these trade-offs to maintain high value alignment without sacrificing conversational creativity and long-term context tracking.

Summary updated

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, we 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, we define the expected behaviors from two famous theories, Maslow's Hierarchy of Needs and Plutchik's Wheel of Emotion. By experimenting with our method on the benchmark of DAILYDILEMMAS, our method exhibits significant performance gains compared to prompt-based baselines, including ECoT, Plan-and-Solve, and Metacognitive prompting. Our method provides a basis for the emergence of self-emotion in AI systems.
BibTeX
@misc{qu2026from,
  author = {Jinxian Qu and Qingqing Gu and Teng Chen and Luo Ji},
  title = {{From Descriptive to Prescriptive: Uncover the Social Value Alignment of LLM-based Agents}},
  year = {2026},
  eprint = {2605.14034},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.14034},
}
Figure