FairGap Benchmark Exposes Hidden Fairness Gaps in LLM Recommenders

FairGap is a new benchmark that challenges the idea that consistent recommendations from LLM-based systems mean their internal processes are stable. Developed by researchers and detailed in arXiv:2608.08284, it evaluates recommendation fairness on two fronts: observable output shifts (OBS) and hidden representation shifts (IBS). By using controlled counterfactual identity probes looking at gender, age, and race, it introduces a concept called Representation-Output Alignment (ROA) to analyze these shifts. When applied to six open-weight LLM families in three different areas, FairGap reveals significant hidden-output discrepancies, with ROA rarely exceeding 0.22. Interestingly, many users see stable outputs while internal changes occur, a problem overlooked in output-only audits. The study also suggests that activation steering could lessen IBS by up to 8 times. You can find the full research on arXiv with the ID 2608.08284.