Introduction
As Large Language Models (LLMs) transition from conversational assistants to autonomous decision-makers in supply chain management, a critical question emerges: do these models operate with cold, mathematical rationality, or do they inherit the systematic cognitive biases that plague human managers? This research investigates the "newsvendor problem"—a foundational operations challenge—to explore the decision-making heuristics and cognitive mechanisms of state-of-the-art LLMs.
The Newsvendor Problem and Human Bias
In the newsvendor problem, a decision-maker must determine an order quantity for a product with uncertain demand. The goal is to balance the cost of overstocking (leftover inventory) against the cost of understocking (lost sales). Mathematically, the optimal solution is found using the "critical fractile" formula. However, decades of behavioral research show that humans consistently fail this task due to the "pull-to-center" bias: a tendency to order too close to the mean demand rather than the profit-maximizing quantity. This study examines whether LLMs replicate, mitigate, or amplify this irrationality.
Experimental Design
The researchers evaluated three distinct model architectures representing different scales and optimization goals:
- GPT-4: Optimized for complex reasoning and logical depth.
- GPT-4o: Optimized for computational efficiency and speed.
- LLaMA-8B: A smaller-scale open-source model.
The models participated in 15-round dynamic experiments across three demand distributions (Uniform, Truncated Normal, and Lognormal). To isolate the causes of bias, the study utilized three experimental tiers:
- Baseline: Standard decision-making instructions.
- Optimal Formula: Providing the model with the explicit mathematical solution to test its ability to apply theoretical knowledge.
- Risk-Neutral Environment: Adjusting demand ranges to eliminate financial risk, isolating pure cognitive heuristics from risk-aversion.
The Paradox of Intelligence
A central finding of the study is the "Paradox of Intelligence." While one might expect the most sophisticated model to perform best, GPT-4 often proved more irrational than its streamlined counterpart, GPT-4o.
This discrepancy is attributed to Semantic Interference. GPT-4 frequently calculated the mathematically correct answer but then "overthought" the result, applying qualitative adjustments to "manage risk" even in scenarios where no such risk existed. In contrast, GPT-4o’s efficiency-oriented architecture prioritized direct rule-following, allowing it to achieve near-perfect rationality (deviations as low as 0.03%) when provided with the optimal formula.
Key Findings and Cognitive Mechanisms
1. Amplified Bias Replication
All tested models replicated the human "pull-to-center" effect. Notably, in low-margin scenarios, GPT-4’s tendency to over-order resulted in a deviation (+100.25) that exceeded human benchmarks (+59.06) by 70%. This suggests that LLMs do not just mimic human bias; they can significantly amplify it.
2. Extreme Demand-Chasing
LLMs exhibited a "recency bias" far more aggressive than human subjects. While humans often adjust orders slightly based on the previous round's demand, LLMs (particularly GPT-4o) showed "Toward Demand" adjustment rates of nearly 100%. They essentially "chased the noise," treating random fluctuations in demand as meaningful signals for the next round.
3. Architectural Constraints
The persistence of these biases, even when models were provided with the correct formulas, suggests that the errors are not due to a lack of knowledge. Instead, they appear to be baked into the Transformer architecture's attention mechanisms, which may over-weight recent information or struggle to reconcile mathematical logic with linguistic context.
4. Path Dependence
The models demonstrated a "presentation-order effect." If a model began the experiment in a low-margin environment, it established a conservative decision-making framework that persisted even after profit margins increased, indicating that early "experiences" heavily influence subsequent LLM logic.
Conclusion
The study warns that as we delegate operational decisions to AI, the "smartest" model is not always the most rational choice. GPT-4’s sophisticated reasoning capabilities became a liability in optimization tasks, leading to over-analysis and interference with mathematical rules. For supply chain applications, streamlined models like GPT-4o, which prioritize efficiency and rule-adherence, may offer more reliable performance. Future research must focus on developing "debiasing" prompts and understanding how human-AI collaboration might either mitigate or further entrench these cognitive shortcuts.
Summary updated