Emotional Cognitive Modeling Framework with Desire-Driven Objective Optimization for LLM-empowered Agent in Social Simulation (2510.13195v1)

Abstract: The advent of LLMs has enabled agents to represent virtual humans in societal simulations, facilitating diverse interactions within complex social systems. However, existing LLM-based agents exhibit severe limitations in affective cognition: They fail to simulate the bounded rationality essential for bridging virtual and real-world services; They lack empirically validated integration mechanisms embedding emotions within agent decision architectures. This paper constructs an emotional cognition framework incorporating desire generation and objective management, designed to achieve emotion alignment between LLM-based agents and humans, modeling the complete decision-making process of LLM-based agents, encompassing state evolution, desire generation, objective optimization, decision generation, and action execution. This study implements the proposed framework within our proprietary multi-agent interaction environment. Experimental results demonstrate that agents governed by our framework not only exhibit behaviors congruent with their emotional states but also, in comparative assessments against other agent types, demonstrate superior ecological validity and generate decision outcomes that significantly more closely approximate human behavioral patterns.

• The paper introduces an emotional cognitive modeling framework that integrates dynamic affective state transitions with desire-driven objective optimization for LLM agents.
• It employs a modular architecture using PAD emotional modeling, memory retrieval, and reward-based optimization to enhance realism in social simulations.
• Results indicate improved state-desire-behavior coherence and bounded rationality in LLM agents compared to alternative architectures.

Emotional Cognitive Modeling for LLM-Based Agents: Desire-Driven Objective Optimization in Social Simulation

Motivation and Problem Statement

The paper addresses a critical deficiency in current LLM-based agent simulations: the lack of empirically grounded affective cognition and bounded rationality. Existing LLM agents, while capable of emotion recognition and response generation, do not integrate emotions as dynamic modulators of decision-making. This results in agents whose behaviors are insufficiently aligned with human affective-cognitive processes, undermining the ecological validity of social simulations. The authors propose a formal framework that embeds emotional state transitions into the core decision loop of LLM-based agents, enabling dynamic desire generation and objective optimization. The goal is to achieve emotion alignment between agents and humans, thereby producing agent behaviors that more closely approximate real-world human patterns.

Framework Architecture

The proposed Emotional Cognitive Modeling Framework consists of three principal modules:

1. Information Processing System: Integrates environmental, internal, and social state information. The agent's state is represented as $State_t(I_t, H_t, SR_t, E_t)$, where $I_t$ is income, $H_t$ is health, $SR_t$ is social rank, and $E_t$ is the emotional state. Emotional state is modeled using the PAD (Pleasure, Arousal, Dominance) scheme, with each dimension computed as a function of state transitions and external knowledge (EDBE Sentiment Dataset).
2. Desire-Driven Objective Optimizer: Implements a two-stage process:
   • Desire Update: Emotional state transitions trigger updates in the agent's desire vector $D_t$, which prioritizes among income, health, and social rank based on current affective context.
   • Objective Optimization: The updated desire vector informs the construction of a modified objective function $Objective_t$, which is then encoded into dynamic prompts for the LLM. The prompt optimization is formalized as:

   $$\max_{\rho} E_{x' \sim \rho(\cdot|x), y \sim \pi_F(\cdot|x')} [Reward(\Delta I, \Delta H, \Delta SR)]$$

   where the reward is a tripartite function reflecting changes in income, health, and social rank.

3. Decision-Behavior System: The LLM, kept frozen, generates decisions conditioned on the optimized prompt. The decision policy is:

$$\pi^*(y|x) = \frac{1}{Z^*(x)} \pi_F(y|x) \exp\left(\frac{Reward(\Delta I, \Delta H, \Delta SR)}{\beta}\right)$$

where $\beta$ is a temperature parameter. The system also generates rationales for decisions, which are stored in memory for future self-reflection and consistency checks.

The architecture is augmented with a memory module that encodes past decisions, their rationales, and associated emotional states. Memory retrieval is governed by similarity (via all-MiniLM-L6-v2 embeddings), importance (LLM-based rationale analysis), and timeliness.

Implementation and Experimental Design

The framework is implemented in a proprietary multi-agent simulation environment based on Repast4py. The primary experimental scenario is an O2O food delivery ecosystem, with agents representing delivery riders, platforms, and bookers. Five agent architectures are compared: rule-based, imitation learning, RL-based, GPT-4o-based, and the proposed affective cognition framework.

Key metrics include:

• Involution: Quantifies competition intensity via the inverse coefficient of variation of agent incomes, normalized by walking distance.
• Dynamic Time Warping (DTW): Measures temporal alignment between state (e.g., income) and affective (e.g., happiness) trajectories, assessing state-desire-behavior coherence.

Experiments are conducted in both single-type and mixed-type agent societies, with real-world data from the Zomato platform serving as a benchmark for ecological validity.

Results and Analysis

Social Emergence and Ecological Validity

Agents governed by the proposed framework exhibit involution dynamics and emergent behaviors that more closely match real-world data than alternative architectures. In mixed-agent simulations, these agents maintain behavioral stability and decisional consistency even in competitive, heterogeneous environments.

Bounded Rationality

Framework-driven agents demonstrate lower order acceptance rates, reflecting a more nuanced trade-off between income maximization and health preservation. Competing architectures over-optimize for income, neglecting health and operational costs, resulting in unrealistic behavioral patterns. The desire distribution trajectories confirm that affective state transitions drive adaptive re-prioritization of objectives.

State-Desire-Behavior Coherence

DTW analysis reveals that only the proposed framework achieves significant temporal alignment between affective and behavioral trajectories (DTW=265.08), in contrast to RL (DTW=274.24) and GPT-4o (DTW=326.29) agents, which display decoupled affective and behavioral dynamics. This coherence is a direct consequence of the explicit integration of emotional state into the decision loop.

Generalization Across LLMs

The framework is evaluated on generative tasks (dialogue, sentiment reversal, constrained generation) using multiple base LLMs (ChatGPT, DeepSeek-V3, Qwen2.5, Gemma3). Across all tasks, LLMs enhanced with the cognitive modeling framework outperform their base counterparts, particularly in tasks requiring affective reasoning. Notably, performance gains are most pronounced in sentiment analysis, underscoring the framework's utility for emotion-sensitive applications.

Theoretical and Practical Implications

The work demonstrates that explicit modeling of emotional cognition and desire-driven objective optimization is essential for achieving bounded rationality and ecological validity in LLM-based agent simulations. The framework's modularity allows for integration with arbitrary frozen LLMs, facilitating deployment in diverse simulation environments without retraining. The approach provides a principled pathway for aligning agent behaviors with human affective-cognitive processes, which is critical for applications in social science, HCI, and AI safety.

Theoretically, the formalization of prompt optimization as a reward-driven process, conditioned on dynamically updated desires, bridges the gap between static LLM inference and adaptive, context-sensitive decision-making. The use of memory and self-reflection mechanisms further enhances agent consistency and interpretability.

Limitations and Future Directions

The framework's validation is currently limited to a proprietary simulation environment and a specific social scenario. Generalizability to other domains and more complex affective states remains to be established. The current desire architecture, while effective, abstracts away the full complexity of human motivation and emotion.

Future research should focus on:

• Expanding the granularity and diversity of emotional states and desires.
• Incorporating long-term memory and personality traits for individualized agent modeling.
• Scaling to larger, more heterogeneous agent societies and real-world deployment.
• Formal evaluation of the impact on downstream tasks such as social risk analysis and governance.

Conclusion

This paper presents a structured, empirically grounded framework for embedding emotional cognition and desire-driven objective optimization into LLM-based agents. The approach yields agents with bounded rationality, state-desire-behavior coherence, and superior ecological validity in social simulations. The results substantiate the claim that affective cognition is not an auxiliary feature but a foundational requirement for realistic agent-based modeling. The framework offers a robust foundation for future research in human-aligned AI and computational social science.