Researchers have developed ReguSim, a controlled environment for testing the ability of large language models (LLMs) to follow rules in financial markets. The system separates four aspects: stated reasoning, attempted action, execution enforcement, and monitor evidence. Experiments with two popular LLMs show that while visible rules reduce rejected actions, they do not eliminate them entirely. The study suggests that evaluating compliance is more complex than a single score, requiring an audit of rule-grounded actions and evidence use. --- Why it matters: This matters to AI researchers because it highlights the limitations of current LLMs in financial markets and provides a framework for evaluating their compliance with rules. Source: https://arxiv.org/abs/2608.19974