At present, fairness assessments in LLM (Large Language Model) multi-agent recruitment systems primarily hone in on the ultimate hiring rates, frequently neglecting the concealed biases within the decision-making process. The paper, which has been accepted for presentation at EMNLP 2026, introduces the SCOPED-Hiring framework—a process-aware fairness diagnostic tool. This innovative framework meticulously tracks over 311,000 structured decision trajectories, scrutinizing them across six key dimensions: outcome, counterfactual scenarios, process flow, decision paths, dynamic interactions, and system design.

The study uncovers several critical insights: career gaps tend to provoke heightened suspicion, subtle cues from agents can sway judgments on candidates' abilities, and identity-related cues can skew the allocation of investigative resources. These findings lay bare the inherent unfairness embedded in these decision trajectories, highlighting that process-aware risks significantly outweigh those focused solely on outcomes.

Drawing on these diagnostic revelations, the research team has devised the Fair Skills remediation mechanism. This mechanism is designed to intervene at high-risk stages of the decision-making process, effectively reducing the overall stratified fairness burden by a remarkable 72.3%. Notably, it achieves this while only marginally altering the final hiring rate—by a mere 1.86 percentage points—and maintaining output efficiency at a stellar level above 99.7%.

This groundbreaking framework marks a paradigm shift in fairness audits, transitioning from a mere outcome-based evaluation to a comprehensive process diagnosis. It empowers practitioners to pinpoint the sources of risk accurately and steer remediation efforts in the right direction. As such, SCOPED-Hiring stands as a valuable process fairness evaluation tool for LLM multi-agent systems involved in high-stakes decision-making.