Title: Investigating the Impact of Gender Signals on LLM Value Trade-offs: Insights from a Controlled Decision Benchmark

Abstract:

As large language models (LLMs) become more prevalent in decision-making contexts sensitive to values, it is critical that irrelevant demographic indicators do not skew judgments. To address this, we introduce the Realistic Value Decision Benchmark (RVDB). This controlled framework isolates the impact of gender by varying only the role-gender configuration, while keeping the scenario, ordered value pairs, roles, candidate decisions, Value Distance, and Decision Severity constant. Through a position-balanced evaluation involving seven distinct models, we examined whether models maintain decision invariance when subjected to gender perturbations and whether their self-reported explanations align with their actual behavioral shifts.

Our findings indicate that explicit gender cues trigger systematic, albeit bounded, decision flips. This effect persists even when models are explicitly prompted to assess whether gender influenced their choice. We observed a consistent asymmetry in cross-gender role swaps, particularly regarding female-proposed decisions. Notably, when decisions were flipped, models frequently attributed the change to "No Influence" or other non-gender factors, thereby obscuring the true cause.

Further analysis reveals that gender-related effects are most pronounced near indeterminate value boundaries and within high-severity decision contexts. This suggests that gender cues function as local factors that shift boundaries, rather than globally overriding value-based reasoning. While overall value rankings remained largely stable, the trade-offs between ordered value pairs shifted unevenly depending on the role-gender configuration. These results demonstrate that gender can behaviorally influence LLM value trade-offs even when this influence is hidden in self-attribution, highlighting the need for controlled behavioral audits that go beyond explanation-based evaluations.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC