Emotional and Cognitive State Modulation: Dynamic Simulation of Human-like Fluctuations in LLMs

TL;DR: Let’s try making LLMs more human by dynamically modulating their simulated “emotional” and “cognitive” states during long-term simulations. The experiment will test if introducing fluctuations in mood, attention, or motivation can disrupt the current bias toward stable, idealized behavior.

Research Question: Does dynamically varying simulated emotional or cognitive states in LLM-based agents increase the realism and diversity of long-horizon, cross-scenario behavior traces?

Hypothesis: Explicitly modeling fluctuating states (e.g., stress, fatigue, mood swings) in LLM prompts will induce more realistic, non-homogenized behavior patterns across extended simulations, narrowing the gap with real-world behavioral diversity.

Experiment Plan: - Design a protocol for modulating LLM agents’ “internal states” (e.g., by altering prompt instructions or persona parameters) in a stochastic or event-driven manner during long-term simulations.

• Run simulations on OmniBehavior, comparing standard static agents with those experiencing state fluctuations.
• Analyze resulting behavior traces for increased variability, emergence of non-average patterns, and alignment with observed human behavior under stress or fatigue.
• Employ metrics like intra-agent behavioral variance and cross-agent divergence.

References:

• Chen, J., Xu, R., Cao, B., Pan, R., Zhang, Y., Hu, Y., Du, Y., Gao, T., Lu, Y., Sun, Y., Han, X., Sun, L., Wu, X., & Lin, H. (2026). Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces.
• Ben-Zion, Z., Witte, K., Jagadish, A. K., Duek, O., Harpaz-Rotem, I., Khorsandian, M.-C., Burrer, A., Seifritz, E., Homan, P., Schulz, E., & Spiller, T. R. (2025). Assessing and alleviating state anxiety in large language models. npj Digital Medicine.