Reason-Mediated Behavioral Models for Auditing LLM Social Simulators
Atharva Pandey, Gautam Jajoo
Why It Matters
What makes this one worth your time
Understanding and improving how LLMs simulate human reasoning is crucial for their reliable use in social simulations and decision-making tasks.
The paper proposes a framework to audit LLMs' ability to simulate human reasoning in social simulations.
Summary
The paper explores the use of large language models (LLMs) as social simulators, specifically focusing on their ability to replicate human rationale-derived reasons in decision-making processes. It introduces an evaluation framework that assesses whether LLMs can simulate human-like reasoning without direct access to human rationales or outcomes, using a sunscreen concept test as a case study.
Key contributions
- Proposes an evaluation framework for auditing LLMs as social simulators.
- Introduces the concept of 'reason states' to assess the alignment of LLM-generated rationales with human reasoning.
Notable insights
- The study highlights the brittleness of LLM-simulated reasons, which often mimic input prompts rather than reflecting genuine human reasoning paths.
- The concept of 'reason states' is introduced as a tool to evaluate the alignment of LLM-generated rationales with human evidence.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2607.24649v2 Announce Type: replace Abstract: Large language models are increasingly used as social simulators, including as synthetic survey respondents. Most evaluations ask whether simulated outcomes resemble human outcomes. We argue that this is necessary but too weak: a simulator can match the final answer while using the wrong rationale-derived reason pattern. We study this problem through a 94-person sunscreen concept test in which each respondent evaluated three product concepts and wrote open-ended rationales. We map those rationales into signed reason states $Z$, where positive signs support adoption and negative signs block it. This gives a practical audit: holding respondent descriptors $D$, category context $K$, and concept treatment $X$ fixed, do human rationale-derived reasons help predict behavior $Y$, and can an LLM simulate the same reason state without seeing the human rationale or outcome? Human rationale-derived reasons substantially improve held-out prediction of purchase intent. LLM-simulated reasons are more brittle: they often sound plausible, but frequently echo the concept board rather than recover the respondent's acceptance or rejection path. The paper contributes an evaluation framework for social simulators. Reason states do not identify natural causal effects by themselves, but they provide an interpretable test of whether a simulator's stated reasons align with human evidence.