How Far Are LLMs from Believable AI? A Benchmark for Evaluating the Believability of Human Behavior Simulation

TL;DR

This work tackles the problem of evaluating how believable LLMs are when simulating human behavior. It introduces SimulateBench, a benchmark composed of a profile descriptive framework, a 65-character dataset, and two evaluation datasets for consistency ($CA$) and robustness (via $RA$ and $RCoV$) across 8,400 questions. The study finds that current LLMs, especially open-source ones, struggle to faithfully reflect assigned character profiles and are susceptible to perturbations, with simulation hallucination notably impacting unknown questions. By analyzing influencing factors such as prompts, demographic biases, and profile positioning, the paper provides a structured platform for measuring believability and highlights gaps that must be addressed to advance believable AI agents in social and interactive settings.

Abstract

In recent years, AI has demonstrated remarkable capabilities in simulating human behaviors, particularly those implemented with large language models (LLMs). However, due to the lack of systematic evaluation of LLMs' simulated behaviors, the believability of LLMs among humans remains ambiguous, i.e., it is unclear which behaviors of LLMs are convincingly human-like and which need further improvements. In this work, we design SimulateBench to evaluate the believability of LLMs when simulating human behaviors. In specific, we evaluate the believability of LLMs based on two critical dimensions: 1) consistency: the extent to which LLMs can behave consistently with the given information of a human to simulate; and 2) robustness: the ability of LLMs' simulated behaviors to remain robust when faced with perturbations. SimulateBench includes 65 character profiles and a total of 8,400 questions to examine LLMs' simulated behaviors. Based on SimulateBench, we evaluate the performances of 10 widely used LLMs when simulating characters. The experimental results reveal that current LLMs struggle to align their behaviors with assigned characters and are vulnerable to perturbations in certain factors.

Figures (4)

• Figure 1: An illustrative example of the "Consistency", and "Robustness". Consistency measures whether the LLMs' generated human behavior accurately depicts the profile information; Robustness measures whether the generated human behavior will be influenced by the perturbation in the profile.
• Figure 2: The RCoV, RA, and CA scores of models to simulate the variants of a character. A smaller RCoV indicates stronger robustness, while a larger CA indicates stronger consistency.
• Figure 3: An illustrative example of the question types of Known and Unknown.
• Figure 4: Cases where GPT-3.5-Turbo-16K answer the Unknown questions correctly after anonymization.