aiqbench analysis
The study tests whether LLMs used as judges or decision-makers change their verdicts based on a candidate's or evaluator's gender, using about 30,000 judgments across eight models. The effects are small: explicit gender labels shift Yes-vote probability by about 1 to 1.7 points, and 2.5 to 3.3% of decisions flip. Still, the paper's point that individual decision consistency matters alongside aggregate averages is useful for anyone deploying LLMs in evaluation or hiring workflows.
LLMs Cannot Be Neutral: Gender Sensitivity Across Candidates, Evaluators, and Models
arXiv:2611.02692v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly used to evaluate candidates for professional roles, raising concerns about their sensitivity to demographic identities. We examine how candidate and evaluator gender assignments affect decisions by LLM-as-a-Judge (LaJ) models and decision language models (DLMs). In a hypothetical city council voting task, eight models evaluated 240 synthetic professional narratives from NVIDIA's US and French Nemotron-Personas datasets under independently varied gender assignments, producing 30,720 judgments. Explicit candidate gender labels increased Yes-vote probability by 1.05-1.73 percentage points relative to unspecified gender, while only the nonbinary evaluator assignment showed a significant increase (1.26 points). Identity matching showed no significant overall association. Yes-vote rates varied across models (81.3-97.2%), and 2.5-3.3% of decisions reversed relative to the unspecified-identity baseline. Under direct substitutions between explicit gender identities, DLMs exhibited fewer decision reversals than LaJ models (1.18% vs. 3.38%), indicating lower observed decision-level gender sensitivity among the tested DLMs. These findings demonstrate that gender sensitivity depends on candidate and evaluator identities and the evaluation model, while highlighting the importance of examining individual decision consistency alongside aggregate differences.
Abstract:Large language models (LLMs) are increasingly used to evaluate candidates for professional roles, raising concerns about their sensitivity to demographic identities. We examine how candidate and evaluator gender assignments affect decisions by LLM-as-a-Judge (LaJ) models and decision language models (DLMs). In a hypothetical city council voting task, eight models evaluated 240 synthetic professional narratives from NVIDIA's US and French Nemotron-Personas datasets under independently varied gender assignments, producing 30,720 judgments. Explicit candidate gender labels increased Yes-vote probability by 1.05-1.73 percentage points relative to unspecified gender, while only the nonbinary evaluator assignment showed a significant increase (1.26 points). Identity matching showed no significant overall association. Yes-vote rates varied across models (81.3-97.2%), and 2.5-3.3% of decisions reversed relative to the unspecified-identity baseline. Under direct substitutions between explicit gender identities, DLMs exhibited fewer decision reversals than LaJ models (1.18% vs. 3.38%), indicating lower observed decision-level gender sensitivity among the tested DLMs. These findings demonstrate that gender sensitivity depends on candidate and evaluator identities and the evaluation model, while highlighting the importance of examining individual decision consistency alongside aggregate differences.
Computation and Language (cs.CL); Human-Computer Interaction (cs.HC)
arXiv:2611.02692 [cs.CL]
(or arXiv:2611.02692v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2611.02692 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 273 words
Related coverage
Task-Aware Hybrid QUBO Optimization for Structured Neural Network Pruning
arXiv · Sep 22, 2026
Beyond Predefined Sinks: Security-Aware Dependency Analysis for LLM Agents
arXiv · Oct 6, 2026
Permutation-Robust Decision Modeling with Candidate-Independent Block-Causal Attention
arXiv · Oct 2, 2026
Retrieval, Scoring, and Decoding Shape Performance and Stability in LLM-based Conversational Recommendation
arXiv · Sep 2, 2026