Evaluating LLM Adaptation Strategies in Physical Security Operations - Reducing Cognitive Load and Alarm Fatigue in Control Rooms

Abstract

Physical security control rooms demand rapid decisions under high alarm volumes, producing cognitive load and alarm fatigue that degrade operator performance. This thesis evaluates three LLM adaptation strategies (zero-shot, few-shot, and retrieval-augmented generation) against a rule-based baseline on two tasks: alarm dispatch and guard chat. Evaluation was conducted in a simulation environment modelling GuardTools Command and Control, a SaaS platform for manned guarding operations. For dispatch, the rule-based strategy was fastest and cheapest. Among LLM strategies, few-shot achieved the lowest execution failure rate (1.82%), while zero-shot degraded sharply under clustered alarms, reaching 26.62% failures. RAG consumed roughly twice the tokens of few-shot without improving reliability. For chat, the decisive factor was context management rather than prompting technique: filtered few-shot and RAG isolated the relevant location description before prompting, achieving the highest quality scores while minimizing token usage.