TriggerBench: Investigating Prospective Memory for Large Language Models

The article introduces TriggerBench, a new benchmark designed to evaluate prospective memory (PM) in Large Language Models (LLMs), addressing a gap in existing assessments that primarily focus on retrospective memory (RM). TriggerBench encompasses five dimensions and enables detailed measurement of proactive recall, false-alarm rates, and attentional robustness, revealing that PM is significantly more challenging than RM, particularly in scenarios involving implicit constraints and concurrent requests. This research highlights the need for improved PM capabilities in LLMs, as it may serve as an indicator of a model’s reasoning capacity, which is crucial for developing more effective AI systems in real-world applications.