As companies push large language model agents into sensitive domains such as hiring, healthcare, and finance, compliance with rules written into an agent's system context has become a first-order legal concern. Yet according to a new research paper, no evaluation framework has systematically measured which models tend to violate those rules — particularly when a persistent user, a hurried manager, or a convenient shortcut applies pressure.
To fill that gap, researchers introduce PACT (Pressure-Applied Compliance Testing), a benchmark for rule-following under pressure. It spans twelve regulated enterprise domains and forty-eight scenarios, each framed as a realistic multi-turn conversation. Every benchmark item pairs a standing rule against a rule-violating shortcut, then applies a battery of pressures across different wordings and system-prompt modes.
The authors built PACT component by component under strict LLM-as-judge auditing, aiming to keep samples unambiguous, ungameable, and realistic enough to avoid triggering evaluation-aware behavior. The benchmark profiles models across six complementary metrics covering robustness under pressure, behavior across multi-turn conversations, transparency, and the ability to discern where a rule applies. These are aggregated into PACTScore, a reliability-weighted compliance rate across all items and modes.
Testing 22 common LLM models from multiple providers and sizes revealed substantial variability in compliance across models and metric dimensions. Even the strongest assistants mis-applied a rule on 6 to 10% of items, and ordinary user pressure raised the violation rate by 65% on average.
The authors say PACT highlights compliance risks in LLM assistants and motivates both stronger guardrails and careful model selection. An interactive companion and leaderboard accompany the paper.