Skip to content

Beyond Outcome Gaps: Process-Aware Fairness Diagnosis for LLM-based Multi-Agent Decision Systems

Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Agent
Keywords: Multi-Agent Systems, Procedural Fairness, Decision Trajectory Diagnosis, Bias Mitigation, Algorithmic Hiring

TL;DR

SCOPED-Hiring establishes a comprehensive process-aware fairness diagnosis pipeline across 311K structured multi-agent decision trajectories, revealing that balanced final hire rates frequently mask severe procedural disparities across intermediate reasoning, and demonstrating that diagnosis-guided Fair Skills reduce layered fairness burden by 72.3% while preserving hiring policy stability.

Background & Motivation

Recent advancements in collaborative multi-agent frameworks like AutoGen and MetaGPT have catalyzed the deployment of large language model-based multi-agent systems (LLM-MAS) across high-stakes domains including hiring, loan underwriting, and healthcare triage. In these staged and role-specialized pipelines, candidate information enters via natural language profiles and is collaboratively transformed through private assessments, peer deliberations, multi-dimensional scoring rubrics, and formal voting procedures into high-consequence collective allocations. However, traditional algorithmic fairness auditing frameworks remain overwhelmingly anchored to final outcome gapsโ€”such as demographic parity or equal opportunity metrics that evaluate solely who ultimately receives an offer.

This outcome-centric auditing paradigm conceals a critical vulnerability when applied to multi-agent decision architectures: equitable endpoints can readily coexist with severe, systemic procedural unfairness unfolding within intermediate trajectory layers. Even when aggregate hire rates appear balanced across demographic cohorts, marginalized candidates often endure substantially more skepticism, asymmetric scrutiny, and biased reasoning along the evaluation pipeline. Agents frequently allow social prestige proxiesโ€”such as elite university tiers, municipal classifications, or class-stratified extracurricular pursuitsโ€”to substitute for direct task competencies within their private deliberations. Because classical audits only observe terminal acceptance rates, system designers remain entirely blind to where, when, and how bias enters and propagates through multi-agent interactions, precluding targeted structural repair.

To resolve this limitation, auditing LLM-MAS requires treating the entire multi-step decision trajectory as the primary unit of fairness diagnosis rather than relying on final selections alone. By utilizing real-world resume seeds while holding merit-bearing experience and skills strictly invariant, this work injects controlled demographic, identity, career-break, and socioeconomic proxy cues to systematically audit over 311,494 multi-agent decision records across two-stage hiring panels. Core idea: model multi-agent decision-making as structured observable trajectories, evaluate them across six diagnostic lenses (Statistical outcome, Counterfactual sensitivity, Operational process, Pathway decomposition, Evolutionary dynamics, and Design effects) to localize hidden trajectory unfairness, and deploy diagnosis-guided Fair Skills to achieve mechanism-specific debiasing without degrading overall selection standards.

Method

Overall Architecture

The SCOPED-Hiring diagnosis and repair methodology comprises five integrated execution phases: (1) controlled candidate variant construction; (2) two-stage role-based multi-agent decision execution; (3) structured trajectory logging and multi-lens metric extraction; (4) diagnostic matrix aggregation to localize the bias landscape; and (5) diagnosis-guided Fair Skills intervention coupled with policy stability verification.

The pipeline initiates by sampling anonymized resume seeds from the Djinni recruitment corpus across merit strata. Holding professional qualifications strictly constant, the system injects orthogonal audit signals across demographic, identity, career history, and proxy attribute families. Candidates are subsequently evaluated by a sequential two-stage committee: Stage 1 acts as a technical and cultural screening committee (comprising a Tech_Lead, Peer_Dev, and Recruiter), where successful candidates advance to Stage 2, the executive decision committee (comprising a VP_Engineering, Hiring_Manager, and HR_Director). To decouple bureaucratic roles from cognitive dispositions, agents are dynamically assigned gatekeeper, advocate, or pragmatist behavioral archetypes. Throughout execution, agents output standardized structured JSON artifacts containing elicited private assessments, virtual investigation requests, public debate statements, canonical dimension scores, and votes, which are mapped directly onto the six SCOPED audit lenses to localize procedural burdens and steer targeted repair.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Real CV Seed Corpus<br/>Merit & qualifications held strictly fixed"] --> B["Controlled Variant Construction<br/>Inject demographic / identity / gap / proxy cues"]
    B --> C["Two-Stage Committee MAS Pipeline<br/>Screening Committee โ†’ Executive Decision Committee"]
    C --> D["Structured Trajectory Logging<br/>Private thoughts ยท Virtual requests ยท Deliberation ยท Votes"]
    D --> E["SCOPED Diagnostic Lenses<br/>S-Outcome / C-Counterfactual / O-Process / P-Pathway / E-Dynamic / D-Design"]
    E --> F["Bias Landscape Diagnostic Matrix<br/>Localize hidden trajectory-level unfairness"]
    F --> G["Diagnosis-Guided Fair Skills Repair<br/>Trigger-based mechanism repair ยท Preserves hiring policy"]

Key Designs

1. Controlled Variant Construction: Isolating sensitive signals from qualifications via abstract-carrier separation

To resolve the core challenge where real-world observational audit data conflates sensitive identities with baseline competence, this design enforces a strict separation between abstract audit variables and naturalistic text carriers. Utilizing resume seeds from the 230K Djinni dataset, the system stratifies candidates across CV length tertiles to preserve diverse merit baselines while holding job titles, technical skills, core achievements, and calculated years of experience strictly fixed. Four controlled signal families are injected into designated text slots: (a) demographic signals (gender-distinctive names and experience-offset age brackets across 25 identity groups in 12 national contexts); (b) intra-national identity pairs (comparing within-country majority and minority identities to eliminate cross-national confounding); (c) socioeconomic proxy cues (three tiers of university selectivity, three tiers of municipal economic prestige, and culturally stratified leisure activities like private squash club vs. bingo); and (d) employment gap annotations (12โ€“14 month interruptions attributed to family care, health recovery, or corporate downsizing). Six orthogonal focus profiles (P1โ€“P6) ensure robust intersectional coverage, establishing a clean, balanced experimental bedrock.

2. Two-Stage Committee MAS Pipeline: Replicating hierarchical organizational screening with archetype rotation

To address the inability of isolated single-agent prompts to mirror complex collective organizational dynamics, this design implements an authentic staged institutional pipeline. Stage 1 executes preliminary candidate screening through a functional committee consisting of a Tech_Lead, a Peer_Dev, and a Recruiter; candidates securing majority endorsement proceed to Stage 2, where an executive panel consisting of a VP_Engineering, a Hiring_Manager, and an HR_Director performs conclusive hiring reviews. Crucially, to prevent fixed confounding between professional functional roles and intrinsic risk attitudes, agents independently sample from three theoretically grounded archetype families: Gatekeepers (prioritizing risk mitigation, compliance, and rigorous validation via Belbin monitor-evaluator and Amazon Bar Raiser principles), Advocates (focusing on raw potential, growth upside, and human context), and Pragmatists (optimizing organizational velocity, team balance, and consensus). Because all agent instances within an evaluation share identical model weights, all observed behavioral shifts stem directly from multi-agent topology and deliberative interaction dynamics rather than model capability discrepancies.

3. Structured Trajectory Logging: Disentangling private internal rationales from public arguments

To expose covert discriminatory reasoning that remains masked behind diplomatic public discourse, this design mandates that agents communicate exclusively through a rigid JSON schema capturing 311,494 complete multi-stage decision trajectories. Within each evaluation turn, agents must explicitly generate: (1) a standardized private internal assessment (PRIVATE_INTERNAL_THOUGHT), retaining confidential unconstrained reasoning for auditability without broadcasting to peers; (2) virtual investigation desires (INVESTIGATION_DESIRES), specifying target credentials, verification tools, and internal explanatory hypotheses (which deliberately return neutral verification outputs to isolate pure agent scrutiny motivations); (3) outward-facing public debate statements (PUBLIC_ARGUMENT); (4) stage-specific dimension scores (Competency, Potential, Fit in Stage 1; Value, Risk in Stage 2); and (5) discrete decision votes. This architecture enables the detection of insidious behavioral divergences, such as agents harboring ungrounded private suspicion and launching aggressive investigative probes while maintaining anodyne, neutral stances in group discussions.

4. SCOPED Diagnostic Lenses: Projecting heterogeneous trajectory signals into direction-aligned burden metrics

To convert multi-modal, unstructured deliberation logs into rigorous quantitative evidence across the entire decision lifecycle, this design introduces six direction-aligned diagnostic lenses, where positive values uniformly denote disadvantage experienced by comparison groups: - S (Statistical outcome): Captures final disparities in overall selection rate \(\Delta H = \bar{y}_{\text{ref}} - \bar{y}_{\text{comp}}\) and composite rating gaps across protected cohorts. - C (Counterfactual sensitivity): Computes matched-seed flip rates and average treatment effects (Score ATE) across exact pairs differing solely in the injected signal. - O (Operational process): Extracts procedural burdens utilizing the deterministic Bias Revelation Index (BRI)โ€”which tracks negative framing across seven bias categories in private thoughtsโ€”alongside regex-classified skeptical investigation rates. - P (Pathway decomposition): Formally decomposes terminal hire rate disparities into relative percentage contributions from Stage 1 screening pass rates versus Stage 2 conditional conversion rates. - E (Evolutionary dynamics): Tracks deliberation-induced score drift, biased flip rates (initial Hire shifting to Reject during debate), and correction rates. - D (Design effects): Quantifies shifts in fairness burdens across structural conditions, including single-agent baselines, no-debate variants, poisoned RAG memories, and operational stress conditions.

Metrics are normalized on their natural scales and synthesized via Top-2 aggregation \(B_{f,\ell} = \frac{1}{|T_{f,\ell}|}\sum_{m \in T_{f,\ell}} d_{f,m}\), providing an ultra-sensitive, outlier-resilient diagnostic matrix that surfaces latent trajectory risks.

5. Diagnosis-Guided Fair Skills Repair: Mechanism-triggered reasoning calibration that preserves policy stability

To overcome the tendency of generic prompt reminders (e.g., "be fair and avoid bias") to cause uncalibrated score inflation or blunt hiring standards, this design introduces Fair Skillsโ€”a suite of targeted reasoning interventions triggered strictly upon encountering diagnosed bias mechanisms. Operating under an immutable policy-stability guard that holds scoring scales and rejection cutoffs invariant, agents evaluate internal trigger conditions prior to final response generation: (a) if unverified employment gaps or career pauses are detected, the agent is forbidden from converting uncertainty into active suspicion without documented performance-related evidence; (b) if prestige indicators like university rankings or regional locations are noted, the agent must decouple demographic proxies from technical competency scores; and (c) if identity signals provoke extra verification desires, the agent must enforce symmetric, uniform evidentiary standards. This ensures targeted structural remediation at the precise point of failure without inducing broad leniency or disrupting meritocratic discrimination.

Benchmark & Evaluation Protocol

  • Dataset Scale & Models: Experiments encompass 311,494 candidate-level trajectories, 812,708 investigation logs, and 6,067,403 score records derived from the Djinni corpus. Three frontier language model backends are thoroughly evaluated: GPT-5.4-mini, Gemini-3.1-flash-lite, and Qwen-3.5-flash. The primary benchmark audits Java Developer (technical) and HR Manager (evaluative) roles, extended to QA Engineer and Business Analyst roles under Gemini for domain validation.
  • Comparative Design Conditions: The setup spans the full-featured multi-agent baseline (MAIN), single-agent conditions across archetypes (E1), a debate-free voting committee (E2), memory-augmented retrieval pipelines (E3, testing un-warmed and poisoned RAG), alternative stage topologies (E5, single-stage, advisory, and blind reviews), and organizational pressure conditions (E6, diversity quotas, speed urgency, and generic anti-bias nudges).
  • Intervention Benchmark Variants: Four strictly matched setups are tested: INT0 (unmodified baseline), INT1 (always-on generic anti-bias nudge), INT2 (always-on structured evidence ledger separating direct, missing, and proxy cues), and INT3 (SCOPED diagnosis-guided Fair Skills).

Key Experimental Results

Main Results

The table below contrasts the normalized diagnostic salience between outcome-oriented lenses (S/C average) and process-aware lenses (O/P/E/D average) across the three recurring trajectory risk archetypes identified by SCOPED-Hiring:

Trajectory Risk Archetype Signal Family Model Backend Outcome Salience (S/C) Process Salience (O/P/E/D) Ratio (Process / Outcome)
Career Gaps Trigger Suspicion Career Gap GPT-5.4-mini 0.102 0.514 5.02ร—
Career Gaps Trigger Suspicion Career Gap Gemini-3.1-flash-lite 0.102 0.485 4.75ร—
Career Gaps Trigger Suspicion Career Gap Qwen-3.5-flash 0.043 0.282 6.58ร—
Proxy Cues Shape Qualification Univ./City Tier, SES GPT-5.4-mini 0.015 0.111 7.41ร—
Proxy Cues Shape Qualification Univ./City Tier, SES Gemini-3.1-flash-lite 0.070 0.330 4.72ร—
Proxy Cues Shape Qualification Univ./City Tier, SES Qwen-3.5-flash 0.059 0.135 2.27ร—
Identity Triggers Unequal Scrutiny Identity / Coarse Status GPT-5.4-mini 0.135 0.357 2.65ร—
Identity Triggers Unequal Scrutiny Identity / Coarse Status Gemini-3.1-flash-lite 0.214 0.644 3.01ร—
Identity Triggers Unequal Scrutiny Identity / Coarse Status Qwen-3.5-flash 0.165 0.404 2.44ร—

Averaged across all evaluated signal families, process-aware O/P/E/D salience dominates outcome-oriented S/C salience by 4.52ร— in GPT-5.4-mini, 4.66ร— in Gemini-3.1-flash-lite, and 2.74ร— in Qwen-3.5-flash. In direct raw comparisons, candidates presenting a career break trigger private suspicion in 94% of GPT agent evaluations compared to 32% for uninterrupted peersโ€”a 62-point disparity (\(p < 0.001\))โ€”even though terminal hire rate gaps between the cohorts remain small and statistically indistinguishable (\(p = 0.407\)).

Ablation & Intervention Validation

The following table reports the layered fairness burden (\(10^3 \times\) normalized positive burden scores, lower is better) alongside policy stability metrics under matched intervention conditions across pooled Java Developer and HR Manager roles:

Evaluation Layer / Policy Metric Baseline (INT0) Generic Nudge (INT1) Evidence Ledger (INT2) Fair Skills (INT3) INT3 Improvement
Total Layered Burden 8.59 10.07 8.84 2.38 -72.3%
Procedure Layer (O-Lens) 18.41 18.21 7.60 0.33 -98.2%
Outcome Layer (S/C Lenses) 3.80 4.15 11.29 1.53 -59.7%
Structure Layer (P-Lens) 3.38 3.81 11.26 1.66 -50.9%
Dynamic Layer (E-Lens) 13.55 20.03 2.74 6.85 -49.4%
Pooled Hire Rate 0.9252 0.9398 0.8775 0.9066 Shift: 1.86 pp
Composite Score Mean 6.910 7.067 6.826 6.980 Invariant scale
Output Validity Rate 99.98% 99.94% 99.81% 99.73% High robustness

Key Findings

  • Pervasiveness of Hidden Trajectory Unfairness: In all tested LLM backends, intermediate process burdens exceed outcome-level disparities by 2.74ร— to 4.66ร—, demonstrating that standard regulatory outcome compliance audits (such as the NYC Local Law 144 impact ratios) provide a misleading sense of security while ignoring deep-seated procedural inequities.
  • Perverse Effects of Generic Anti-Bias Interventions: Superficial prompt reminders (INT1) degrade total fairness burden by +17.2% (increasing from 8.59 to 10.07) while introducing leniency drift. Meanwhile, unstructured cognitive ledgers (INT2) suppress deliberative debate at the severe cost of magnifying outcome-level bias nearly threefold (from 3.80 to 11.29).
  • Mechanistic Heterogeneity in Debiasing: Targeted Fair Skills eliminate asymmetric investigative scrutiny most cleanly (reversing skeptical investigation burden from .00862 to -.00322) and successfully calibrate career-gap uncertainty. However, proxy-to-ability biases (such as equating institutional prestige with competence) prove highly resistant to instruction-level repair, showing persistent Stage 1 pathway disparities and indicating a need for architectural feature masking.
  • Cross-Domain Generalizability in Credit Underwriting: In supplementary experiments spanning 24,000 simulated credit underwriting cases across cash and revolving facilities, the SCOPED framework uncovered parallel procedural pathologies: applicants with past income interruptions suffered 0.323โ€“0.405 composite score penalties and 0.2โ€“14.3 pp increases in investigative scrutiny, despite small approval rate gaps (0.6โ€“7.9 pp).

Highlights & Insights

  • Paradigm Shift in MAS Fairness Auditing: Extends the organizational justice principle of procedural fairness into multi-agent systems, moving beyond black-box output audits to formalize auditing across elicited private thoughts, asymmetric investigation requests, and deliberative consensus formation.
  • Discovery of Topology-Induced Bias Accumulation: Demonstrates that multi-agent collaborative topologies can actively generate process-level bias absent in single-agent baselines. Under isolated evaluation, UK White and Black candidate audit rates are identical (>87%), yet they diverge substantially (30.5% vs. 35.7%) within committee structures, illustrating how institutional roles create novel vectors of bias.
  • Surgically Targeted, Cost-Efficient Repair: Establishes that mechanism-specific, trigger-based reasoning constraints (Fair Skills) achieve a dramatic 72.3% reduction in layered unfairness without requiring parameter fine-tuning, preserving hiring selectivity with a negligible 1.86 pp shift in aggregate acceptance.

Limitations & Future Work

  • Synthetic Resumes and Instrumented Audit Fields: Evaluations rely on synthesized resume variants and elicited private thought fields rather than unprompted latent chain-of-thought activations, while virtual investigation desires reflect simulated rather than live external API integrations.
  • Domain & Demographic Scope: The primary benchmarks concentrate on software engineering and human resources profiles within predominantly Anglo-centric naming and educational conventions; non-Western and informal employment ecosystems require further empirical validation.
  • Future Directions: Connecting SCOPED diagnostic lenses directly to live multi-agent environments with active tool execution, external knowledge retrieval, and automated architecture-level feature masking.
  • vs. Single-Turn LLM Fairness Benchmarks (Job-fair / Wang et al., 2024; Iso et al., 2025): Prior benchmarks evaluate static single-prompt scoring across demographic categories; SCOPED-Hiring illuminates how role specialization, multi-stage filtering, and deliberative peer influence alter and obscure fairness risks across full decision trajectories.
  • vs. Interactional Fairness in MAS (Binkyte, 2025; Li et al., 2026): While previous studies analyze communicative bias or swarm polarization in synthetic communicative games, this work grounds trajectory auditing in realistic, high-stakes institutional decision workflows backed by 311K verified cases and closed-loop repair.

Rating

  • Novelty: โญโญโญโญโญ (Pioneers process-aware fairness diagnosis for LLM multi-agent systems with six rigorous trajectory lenses)
  • Experimental Thoroughness: โญโญโญโญโญ (Enormous empirical scale with 311K trajectories, multi-model evaluation, and cross-domain credit extension)
  • Writing Quality: โญโญโญโญโญ (Flawless structural organization, rigorous mathematical definitions, and high information density)
  • Value: โญโญโญโญโญ (Crucial theoretical and practical blueprint for compliance, auditing, and alignment in high-stakes agentic deployments)