Context-Value-Action Architecture for Value-Driven Large Language Model Agents

The core problem of behavioral rigidity in Large Language Model (LLM)-based agents and the motivation for the proposed Context Value Action (CVA) architecture. While LLMs have shown promise in simulating human behavior across various domains—ranging from virtual avatars and social simulacra to embodied Vision Language Action systems and task-oriented assistants existing agents often fail to capture the complexity, diversity, and stochasticity inherent in real human behavior. A critical flaw identified is behavioral rigidity, where agents produce stereotypical or exaggerated responses due to latent model biases amplified by current design paradigms. This issue is frequently masked by evaluation methods that rely on “LLM-as-a-judge” metrics, which suffer from self-referential bias because the judge model shares similar pre-training data and biases with the agent being evaluated. As a result, such evaluations may incorrectly validate polarized or caricatured behaviors as high-quality simulations.

To address this, grounding agent decision-making in established psychological theories, specifically the Stimulus-Organism-Response (S-O-R) model and Schwartz’s Theory of Basic Human Values. They conceptualize human behavior as a dynamic process of value activation influenced by context, rather than a static output based on fixed personas. For example, an individual may generally value self-direction but prioritize hedonism after a tiring workday, seeking relaxation over productivity. Existing agents fail to model this context-dependent value activation, leading to rigid and unrealistic behaviors. The proposed CVA architecture aims to bridge this gap by explicitly decoupling action generation from cognitive reasoning through a novel Value Verifier trained on authentic human data. This verifier models how values are dynamically activated by context and ensures that generated actions align with these activated values, enhancing both behavioral fidelity and interpretability.