Do AI agents actually think like humans, or just pretend?

Most LLM benchmarks measure reasoning and planning, ignoring emotional consistency and personality stability. Researchers created HEART-Bench: 11 detailed character profiles (grounded in Big Five traits and 1,000 autobiographical memories each), tested against 64 psychologically-designed scenarios across eight dimensions like adversity, sociality, and deception. The benchmark filters these into 673 multiple-choice questions to measure whether agents actually behave like coherent humans or just pattern-match. Results reveal how far current LLMs fall from genuine psychological consistency.