Prompt-Based Value Steering of Large Language Models

2025-11-14

AI Summary

This study addresses **value alignment without model fine-tuning or dynamic prompt optimization**: how to guide large language models (LLMs) to generate value-congruent text via *static prompt design* alone. Methodologically, it formalizes target human values—grounded in Schwartz’s Theory of Basic Human Values—and constructs a structured dialogue dataset; it further proposes a reproducible, model-agnostic prompt evaluation framework that quantifies both the *presence* and *incremental gain* of target values in generated outputs. Experiments using a Wizard-Vicuna variant demonstrate that prompts explicitly conditioning on target values significantly improve value consistency over baseline prompts, with statistically significant alignment gains. The core contribution is the first *quantifiable, static-prompt evaluation paradigm* for dynamic value alignment—enabling lightweight, interpretable, and controllable value guidance in LLMs.

Technology Category

• Humans and AI: Learning Human Values and Preferences
• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Large language models are increasingly used in applications where alignment with human values is critical. While model fine-tuning is often employed to ensure safe responses, this technique is static and does not lend itself to everyday situations involving dynamic values and preferences. In this paper, we present a practical, reproducible, and model-agnostic procedure to evaluate whether a prompt candidate can effectively steer generated text toward specific human values, formalising a scoring method to quantify the presence and gain of target values in generated responses. We apply our method to a variant of the Wizard-Vicuna language model, using Schwartz's theory of basic human values and a structured evaluation through a dialogue dataset. With this setup, we compare a baseline prompt to one explicitly conditioned on values, and show that value steering is possible even without altering the model or dynamically optimising prompts.

Problem

• Steering LLM outputs toward human values
• Evaluating prompt effectiveness for value alignment
• Achieving value alignment without model fine-tuning

Innovation

• Prompt-based steering method for value alignment
• Model-agnostic procedure evaluating prompt effectiveness
• Quantitative scoring for target values in responses