Dynamics of Cognitive Heterogeneity: Investigating Behavioral Biases in Multi-Stage Supply Chains with LLM-Based Simulation

A team of researchers has published a groundbreaking study at ACL 2026 that uses Large Language Models (LLMs) as simulated decision-makers in complex supply chain environments. The paper, titled 'Dynamics of Cognitive Heterogeneity: Investigating Behavioral Biases in Multi-Stage Supply Chains with LLM-Based Simulation,' introduces a scalable experimental paradigm that moves beyond traditional behavioral research limitations. By employing both DeepSeek and GPT agents within a Hierarchical Reasoning Framework, the researchers systematically varied reasoning sophistication across different supply chain tiers—manufacturers, distributors, and retailers—to study how cognitive diversity impacts collective outcomes.

Through rigorously replicated simulations with statistical validation, the study found that LLM-based agents naturally exhibit human-like behavioral biases, including myopic decision-making and self-interested behaviors that exacerbate systemic inefficiencies like the bullwhip effect. However, the research demonstrated that structured information sharing between agents effectively mitigates these adverse effects, reducing inefficiency by measurable margins. This work represents a significant advancement in using AI agents as proxies for human decision-making in operational research, offering both a new methodological approach and practical insights for designing more resilient AI-enabled organizations. The findings suggest that while LLMs can replicate problematic human behaviors, they also provide a testbed for interventions that could improve real-world supply chain coordination.