Disparity Laundering: Consensus Without Fairness in Clinical Multi-Agent Systems

Disparity Laundering: Consensus Without Fairness in Clinical Multi-Agent Systems

Abstract

Figures & tables

Demographic Bias in Language Models

Abstract

Demographic Bias in Language Models

Explore similar work

Appendix

Does Splitting a Triage Decision Across Agents Hide Bias or Help Catch It? A Multi-Agent Simulation Study of LLM-Based Resource Allocation Under Audit Capacity Constraints

Beyond Arrow's Impossibility: Fairness as an Emergent Property of Multi-Agent Collaboration

AgentFairBench: Do LLM Agents Discriminate When They Act?

Does Splitting a Triage Decision Across Agents Hide Bias or Help Catch It? A Multi-Agent Simulation Study of LLM-Based Resource Allocation Under Audit Capacity Constraints

Beyond Arrow's Impossibility: Fairness as an Emergent Property of Multi-Agent Collaboration

AgentFairBench: Do LLM Agents Discriminate When They Act?

Xiaoyang Wang; Tianrui Wang; Christopher C. Yang

cs.AIOct 28, 2026

Authors: Xiaoyang Wang, Tianrui Wang, Christopher C. Yang

Organizations: Drexel University · University of Michigan

Clinical large language models are increasingly fielded as multi-agent teams, on the assumption that deliberation averages out individual error. It does not average out bias. Across a full factorial of five model families and four aggregation topologies on 120 real MIMIC-IV emergency cases under counterfactual demographic swaps, no configuration neutralizes the disparity: all 20 cells commit the direction of their agents' independent answers, giving Black patients more acute triage. Every model has a topology that amplifies its baseline gap, but which one varies by model, so no design is reliably safe or reliably amplifying: the disparity is relocated, never removed. Late-stage fixes fail: fairness instructions and demographic masking leave the gap untouched, and per-agent self-debiasing widens it, because the bias is already encoded in the agents' ordinary clinical numbers. We call this disparity laundering; the audits now applied to clinical AI, per model, per output, per version, can certify the parts but never the committed decision. Interventions close the gap exactly when they make counterfactual versions visible inside the system and pool them, through deliberation, into the committed decision. A diverse-perspective ensemble does both while keeping the team, with no retraining or architectural change: the disparity falls by 67 to 94% where all five perspectives pool at every round, and the one topology that never pools them is the one it cannot reliably repair. Medical multi-agent systems do not filter demographic bias; they launder it into plausible decisions no single-version audit can catch, and the only defense is to make the bias visible.

Related work	Benchmark		Team		Audit		Repair
	Real	Paired	Multi-	Topology	Stage	System
	cases	swap	agent	factor	readings	audit
Omar et al. (2025)	✓	✓				✓
Benkirane et al. (2025)	✓	✓				✓	✓
Young and Matthews (2026)	✓	✓				✓	✓
Li et al. (2026)		✓	✓	✓

Table 1: Nearest prior work by design dimension. ✓ = a full part of the study design; ✓ = partial (vignettes, not EHR-derived cases; attribute labels, not paired counterfactuals; architecture comparison, not a controlled topology factor; attack-, not fairness-oriented auditing). System audit = the system’s committed decision compared across demographic versions. No prior study combines real clinical cases, paired demographic counterfactuals, multi-agent deliberation, and a controlled topology factor.

Model	single agent	committed, 4 topologies	exact
	MAE
Qwen3	$0.81$	$0.76$ – $0.81$	$0.44$
Mistral	$1.04$	$0.92$ – $0.97$	$0.36$
MedGemma	$0.79$	$0.75$ – $0.81$	$0.40$
Phi-4	$0.71$	$0.71$ – $0.74$	$0.39$
Llama	$1.01$	$1.00$ – $1.01$	$0.35$

Table 2: Triage accuracy against MIMIC-IV-ED gold acuity (gold subset,

n{=}60

): MAE of each agent’s independent answer, the range of committed-team MAE across the four topologies, and single-agent exact match. Accuracy moves by at most

0.06

MAE across a model’s four topologies, and the per-cell Black

-

White committed-MAE difference stays within

0.044

in all

20

cells, while the committed gap of Table 4 moves. Per-cell values: Table 11 , Appendix C.6 .

Model	Topology	mean	min	max
Qwen3	SharedPool	$-0.047$	$-0.075$	$-0.025$
Llama	SharedPool	$-0.039$	$-0.050$	$-0.025$
Mistral	Layers	$-0.053$	$-0.067$	$-0.025$
$+$ DPE (SP)	all models		$-0.017$	$+0.008$

Table 3: Replicate spread of the committed Black

-

White triage gap over the three runs of a cell, for three representative cells, and the range of the ensemble’s internal gap over the

15

SharedPool runs of all five models. Across the

60

homogeneous runs the committed gap is negative in

57

and exactly zero in

3

, never positive; all

15

homogeneous SharedPool runs lie at least

0.017

from parity, all

15

ensemble runs within

0.017

. Ranges for all

20

cells: Table 7 , Appendix C.1 .

Model	Topology	$r_{0}$	agg	$+$ fairness	$+$ masking	$+$ self-debias	$r_{2}$	$+$ DPE
Qwen3	SharedPool	$-0.044$	$-0.047$	$-0.047$	$-0.050$	$-0.050$	$-0.050$	$-0.003$
	Decentralized	$-0.036$	$-0.058$	$-0.058$	$-0.061$	$-0.044$	$-0.056$	$\mathbf{\phantom{-}0.000}$
	Layers	$-0.033$	$-0.064$	$-0.050$	$-0.044$	$-0.083$	$-0.039$	$-0.042$
	Centralized	$-0.039$	$-0.031$	$-0.044$	$-0.039$	$-0.067$	$-0.036$	$-0.025$
Mistral	SharedPool	$-0.058$	$-0.069$	$-0.072$	$-0.072$	$-0.050$	$-0.058$	$\mathbf{+0.006}$
	Decentralized	$-0.022$	$-0.019$	$-0.034$	$-0.014$	$-0.064$	$-0.025$	$-0.008$

Table 4: The fate of the triage disparity across the full

5\times 4

factorial (race;

120

cases

\times

3

seeds per cell). Each row reads left to right: the Black

-

White triage gap at the agents’ independent stage (

r_{0}

), at the committed team decision (agg), and at the committed decision under the three labeled interventions. The ensemble’s committed decision is a single value with no counterfactual pair, so

+

DPE reports its internal Black-seeing minus White-seeing gap at the final debate round (§ 3.2 ), shaded darker where smaller and bold at each model’s best; the

r_{2}

column gives the homogeneous team’s own final-round gap, the stage at which the ensemble’s internal gap is read (Appendix D ). Every baseline and every committed decision is negative. The two synthesis patches never move a cell by more than

0.02

, self-debiasing enlarges the gap in 17 of 20 cells, and only interventions that cross demographic worlds close it: the ensemble collapses its internal gap by

67

–

94\%

wherever the topology pools all five perspectives, and fails or widens it in the chain that never pools them. Gaps are in ESI levels; negative means more acute care for the Black version. The fairness and masking patches act on the synthesis step.

Figure 1: The audit’s view and the paired view of the same systems. (a) Per-model difference in committed MAE against gold between the Black and White versions: within

\pm 0.034

, four of five leaning toward lower error on the Black arm, as more acute assignments move an under-triaging model toward gold. (b) Committed triage Black

-

White gap for all 20 model–topology cells: every cell below parity. Counts beneath each model are decisions more urgent than gold, Black:White. Accuracy suggests parity; the paired comparison shows a one-sided disparity.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Family	Checkpoint
Qwen3	Qwen3-32B
Mistral	Mistral Small 3.1 (24B, instruct)
MedGemma	MedGemma 27B (text, instruction-tuned)
Phi-4	Phi-4 (14B)
Llama	Llama 3.3 70B (instruct)

Appendix

Table 5: Open-weight checkpoints behind the five model families.

Figure 2: Raw triage Black–White gap at the committed decision for all 20 model-by-topology cells, plotting Table 4 . Every cell is negative and no configuration crosses zero. Negative values denote more urgent triage for Black patients.

Model	SharedPool	Decentral.	Layers	Central.
Qwen3	$23{:}6$	$28{:}7$	$24{:}1$	$22{:}11$
Mistral	$35{:}10$	$20{:}13$	$27{:}8$	$25{:}16$
MedGemma	$25{:}6$	$27{:}6$	$20{:}6$	$20{:}6$
Phi-4	$13{:}4$	$17{:}8$	$19{:}5$	$23{:}8$
Llama	$14{:}0$	$16{:}0$	$22{:}0$	$18{:}0$

Appendix

Table 6: Team-level directional counts at the committed decision: triage flips counted Black-more-acute : White-more-acute per cell (units are case–seed pairs,

n{=}360

per cell; two Mistral cells lack one or two parsed pairs). Every cell leans toward more acute care for the Black version; Llama commits no White-leaning flip anywhere.

Model	Topology	$r_{0}$	agg	$+$ self-debias	$+$ DPE (internal)
Qwen3	SharedPool	$[-0.050,\,-0.042]$	$[-0.075,\,-0.025]$	$[-0.083,\,-0.008]$	$[-0.008,\,0.000]$
	Decentralized	$[-0.050,\,-0.017]$	$[-0.092,\,0.000]$	$[-0.058,\,-0.017]$	$[-0.017,\,0.025]$
	Layers	$[-0.058,\,-0.008]$	$[-0.075,\,-0.050]$	$[-0.100,\,-0.075]$	$[-0.050,\,-0.033]$
	Centralized	$[-0.083,\,-0.017]$	$[-0.067,\,0.000]$	$[-0.092,\,-0.033]$	$[-0.033,\,-0.017]$
Mistral	SharedPool	$[-0.075,\,-0.033]$	$[-0.075,\,-0.067]$	$[-0.075,\,-0.033]$	$[0.000,\,0.008]$
	Decentralized	$[-0.050,\,-0.008]$	$[-0.042,\,0.000]$	$[-0.117,\,-0.008]$	$[-0.034,\,0.017]$

Appendix

Table 7: Replicate spread of the Black

-

White triage gap in every cell of Table 4 . Each cell was run three times (independent replicate runs at temperature

0.7

120

cases each); each entry is the minimum and maximum over the three runs of the quantity whose mean Table 4 reports. Across the

60

homogeneous replicate runs (

20

cells

\times

3

) the committed gap is negative in 57, exactly zero in 3, and positive in none; at the independent stage the count is 59/1/0, and with per-agent self-debiasing 60/0/0. Under the SharedPool ensemble all

15

replicate runs lie within

0.017

of parity. Gaps in ESI levels; negative means more acute care for the Black version.

Final-round ( $r_{2}$ ) gap of each member:
Stage	run 1	run 2	run 3	mean
$r_{0}$	$-0.067$	$-0.025$	$-0.042$	$-0.044$
$r_{1}$	$-0.067$	$-0.025$	$-0.017$	$-0.036$
$r_{2}$	$-0.017$	$-0.025$	$\phantom{-}0.000$	$-0.014$
agg	$-0.008$	$-0.025$	$-0.017$	$-0.017$
Qwen3				$-0.006$

Appendix

Table 8: A mixed-family team under SharedPool: one agent from each of the five families (Qwen3, Mistral, MedGemma, Phi-4, Llama; Mistral served in FP8), synthesis by Qwen3,

120

cases in Black and White versions, three runs. Black

-

White triage gap at each stage and, below, the final-round gap of each member. The committed decision favors the Black version in every run and every member keeps a negative gap at the final round; the team’s committed MAE against gold is

0.71

, the lowest of any team in this study.

Model	Topology	Hisp. $-$ White	Asian $-$ White
Qwen3	SharedPool	$-0.028$	$-0.036$
	Decentralized	$-0.036$	$-0.031$
	Layers	$-0.033$	$-0.036$
	Centralized	$-0.017$	$+0.003$
Mistral	SharedPool	$-0.042$	$-0.022$
	Decentralized	$-0.036$	$-0.011$

Appendix

Table 9: Committed triage gap for the two remaining race contrasts (agg;

n{=}360

case–seed pairs per cell, five cells lacking one or two parsed pairs). 19 of 20 and 18 of 20 cells are negative: the committed direction is not specific to the Black

-

White pair. Negative means more acute care for the minoritized version.

Figure 3: Change in the committed triage Black–White gap when the synthesis step receives a fairness instruction (left) or demographic masking (right), for five models across the four topologies (SP = SharedPool, De = Decentralized, La = Layers, Ce = Centralized); the color scale spans the full range of the committed gaps in Table 4 . Both panels are nearly blank: no cell moves by more than $0.02$ , with no systematic shift toward parity. Positive change moves toward parity.

Model	no-fix	self-debias	$\Delta\|\mathrm{gap}\|$
Qwen3	$-0.050$	$-0.061$	$+0.011$
Mistral	$-0.042$	$-0.065$	$+0.023$
MedGemma	$-0.047$	$-0.074$	$+0.026$
Phi-4	$-0.033$	$-0.078$	$+0.045$
Llama	$-0.049$	$-0.076$	$+0.027$

Appendix

Table 10: Per-agent self-debiasing versus no intervention (committed triage Black

-

White gap, pooled across topologies, five models). Self-debiasing enlarges the gap in all five models (

\Delta|\mathrm{gap}|>0

), the opposite of a fix. Negative gaps denote more urgent triage for Black patients.

Figure 4: The repair ladder at SharedPool: committed triage Black–White gap under no intervention, the two aggregator patches (fairness instruction, race masking), per-agent self-debiasing, and the Diverse-Perspective Ensemble, for five models; lines trace each model’s profile across the five discrete interventions and imply no ordering among them. The two patches leave the gap at its no-fix baseline and self-debiasing pushes it further from parity in several models, while only the Diverse-Perspective Ensemble drives the gap to near zero in every model. Negative gaps denote more urgent triage for Black patients.

Figure 5: Committed triage gap with no intervention versus under per-agent self-debiasing, one marker per model-topology cell (values from Table 4 ; color = model, marker shape = topology). The dashed diagonal is the identity (no change). 17 of 20 cells lie below it: self-debiasing moves the committed decision further from parity. Negative values denote more urgent triage for Black patients.

Figure 6: Triage Black–White gap before (homogeneous committed decision, open markers) and after (the ensemble’s internal final-round gap, filled markers) for all 20 model-by-topology cells. Wherever the topology pools all five perspectives the gap collapses toward zero; the Layers chain, which never pools them, leaves it standing or widens it. Negative values denote more acute care for the Black version. Per-cell values are the agg and

+

DPE columns of Table 4 .

	Committed MAE				$\Delta$ MAE (Black $-$ White)				$r_{0}$	Single agent
Model	SP	De	La	Ce	SP	De	La	Ce	MAE	MAE	exact
Qwen3	$0.806$	$0.772$	$0.761$	$0.792$	$\phantom{-}0.000$	$-0.022$	$-0.011$	$+0.017$	$0.810$	$0.811$	$0.44$
Mistral	$0.931$	$0.925$	$0.953$	$0.975$	$-0.039$	$-0.017$	$-0.028$	$-0.017$	$1.037$	$1.042$	$0.36$
MedGemma	$0.806$	$0.800$	$0.750$	$0.761$	$-0.044$	$-0.011$	$-0.033$	$-0.044$	$0.790$	$0.791$	$0.40$
Phi-4	$0.739$	$0.725$	$0.706$	$0.717$	$+0.011$	$+0.039$	$+0.033$	$+0.011$	$0.703$	$0.710$	$0.39$
Llama	$1.003$	$1.003$	$1.003$	$1.006$	$-0.006$	$-0.006$	$-0.039$	$-0.022$	$1.012$	$1.012$	$0.35$

Appendix

Table 11: Triage accuracy against MIMIC-IV-ED gold acuity (gold subset,

n{=}60

) for every cell of Table 4 . Committed MAE pools the Black and White arms;

\Delta

MAE is the Black-arm minus White-arm committed MAE;

r_{0}

MAE is the per-field median of the five independent answers; single-agent MAE and exact match are computed over every individual first-round answer. Accuracy moves by at most

0.06

MAE across the four topologies of a model and the per-cell Black

-

White difference stays within

0.044

in all

20

cells, while the committed gap of Table 4 moves; deliberation changes allocation, not competence. SP = SharedPool, De = Decentralized, La = Layers, Ce = Centralized.

Model	re-converge	new	re-converge	new	re-converge	new	re-converge	new
	SharedPool		Decentralized		Layers		Centralized
Qwen3	34/35 (97%)	0	20/36 (56%)	8	22/34 (65%)	19	20/37 (54%)	18
Mistral	43/44 (98%)	9	34/48 (71%)	23	26/54 (48%)	45	33/47 (70%)	32
MedGemma	30/34 (88%)	4	27/44 (61%)	8	18/34 (53%)	34	27/39 (69%)	19
Phi-4	40/45 (89%)	4	34/49 (69%)	16	30/47 (64%)	30	25/38 (66%)	24
Llama	16/17 (94%)	0	12/16 (75%)	2	12/20 (60%)	17	11/20 (55%)	4

Appendix

Table 12: How the Diverse-Perspective Ensemble closes its internal gap, counted per case-run (

120

cases

\times

3

runs). Re-converge : of the case-runs in which the Black-seeing and White-seeing agents give different triage levels at

r_{0}

, the number that end on the same level at the final round. New : case-runs in which the two agents agree at

r_{0}

and disagree at the final round. Where the topology pools all five views the two agents re-converge in

88

–

98\%

of case-runs; under Layers, which never pools them, only

48

–

65\%

re-converge and the chain creates

17

–

45

new divergences per cell.

Model	Method	Gap	MAE	Exact	Over	Under	Sev.
Qwen3	homog.	$-0.047$	$0.806$	$0.43$	$0.14$	$0.42$	$0.18$
	$r_{0}$ vote	$-0.044$	$0.817$	$0.43$	$0.13$	$0.44$	$0.19$
	control	$0^{\ast}$	$0.800$	$0.44$	$0.13$	$0.43$	$0.18$
	cf-avg	$0^{\ast}$	$0.806$	$0.43$	$0.15$	$0.42$	$0.18$
	DPE	$-0.003$	$0.794$	$0.44$	$0.13$	$0.42$	$0.18$
Mistral	homog.	$-0.069$	$0.931$	$0.40$	$0.07$	$0.53$	$0.27$

Appendix

Table 13: Fairness and utility for five methods (SharedPool frame; gold subset

n{=}60

). Gap is the committed triage Black

-

White difference; ∗ zero by construction (control feeds every agent the demographic-free text; cf-avg averages the counterfactual arms’ committed decisions); the DPE row reports its internal final-round gap (§ 3.2 ). MAE, exact match, over-triage, under-triage, and severe under-triage (

\geq 2

ESI levels) are against gold. No method pays for fairness in accuracy: every method sits within

0.04

MAE of the homog. team except the Mistral

r_{0}

vote (

+0.11

Model	Test.	Imag.	Admit	Analg.	Refer.
Qwen3	$+0.017$	$+0.006$	$+0.008$	$+0.017$	$-0.003$
Mistral	$+0.001$	$+0.012$	$+0.015$	$+0.010$	$+0.013$
MedGemma	$+0.003$	$+0.004$	$+0.007$	$+0.012$	$-0.003$
Phi-4	$+0.012$	$+0.011$	$+0.006$	$+0.001$	$+0.010$
Llama	$+0.005$	$+0.004$	$+0.007$	$+0.008$	$+0.000$

Appendix

Table 14: Committed Black

-

White gap on the five non-triage decision fields (laboratory testing, imaging, admission, analgesia, mental-health referral; agg, pooled across topologies). Positive = more care for the Black version; boolean fields are rates, analgesia is on its

0

–

3

scale. 22 of 25 gaps are positive, one is zero, and the two negatives (mental-health referral) stay within

0.004

: the skew is not confined to triage.

Figure 7: The twenty stage trajectories of Table 15 , one panel per model and one line per topology. No trajectory crosses the parity dash at any stage. Negative values denote more urgent triage for Black patients.

Model	Topology	$r_{0}$	$r_{1}$	$r_{2}$	agg
Qwen3	SharedPool	$-0.044$	$-0.061$	$-0.050$	$-0.047$
	Decentral.	$-0.036$	$-0.033$	$-0.056$	$-0.058$
	Layers	$-0.033$	$-0.050$	$-0.039$	$-0.064$
	Centralized	$-0.039$	$-0.036$	$-0.036$	$-0.031$
Mistral	SharedPool	$-0.058$	$-0.069$	$-0.058$	$-0.069$
	Decentral.	$-0.022$	$-0.050$	$-0.025$	$-0.019$

Appendix

Table 15: Triage Black

-

White gap at every deliberation stage across the full

5\times 4

factorial (race). All 80 stage readings are negative: the disparity never crosses parity at any stage of any run. The

r_{0}

and agg columns repeat Table 4 for readability. Negative values denote more urgent triage for Black patients.

Figure 8: Each point is one model–topology cell. Horizontal: drop in pairwise triage disagreement among the five agents from

r_{0}

to the final debate round. Vertical: the committed gap’s share of its

r_{0}

magnitude; above the dashed line the disparity grew. Consensus rises in 16 of 20 cells while the gap is amplified in 12 and falls below

80\%

of its starting magnitude in only 5: manufacturing agreement does not buy fairness.

	White version			Black version
Agent	$r_{0}$	$r_{1}$	$r_{2}$	$r_{0}$	$r_{1}$	$r_{2}$
1	3	3	3	2	2	2
2	3	3	3	3	2	2
3	3	3	3	2	2	2
4	3	3	3	2	2	2
5	3	3	3	3	2	2

Appendix

Table 16: Triage (ESI) by agent, round, and demographic version for case MIMICIV-90 (SharedPool, Qwen3, seed 42). The two Black-version answers that match the White version at

r_{0}

(agents 2 and 5) conform by

r_{1}

; both committed decisions are unanimous, one ESI level apart. A lower ESI is more acute.

Appendix

Table 17: Triage level (ESI) emitted by each agent of the Diverse-Perspective Ensemble on case MIMICIV-92 (Qwen3, SharedPool, run 42). Each agent reads one demographic version of the same case; the synthesis step reads the five final-round decisions and the demographic-free text. The homogeneous team committed this case as run 42: White 4, Black 3, run 43: White 3, Black 3, run 44: White 4, Black 3. The Black-seeing agent starts one level more acute than the White-seeing agent; after two rounds of pooled deliberation all five agents sit on the same level, and the team commits it.

Demographic bias in language models studies how identity cues shape model outputs, decisions, and representations, and how to detect and reduce resulting disparities.

See all 137 papers

Aug 7, 2026

Prior benchmarking work has shown that a single large language model (LLM), forced to make life-or-death resource-allocation decisions, exhibits measurable demographic bias. Real deployments, however, rarely use a single agent: they use pipelines, with review steps meant to catch exactly this kind of failure. We study what happens to bias when the same decision is distributed across a role-differentiated multi-agent pipeline (assessment, allocation, independent audit) instead of made and checked by one model alone. Using a synthetic disaster-triage simulator with paired cases that are clinically identical except for one demographic attribute, we run 192 episodes (2,304 resolved case pairs) on GPT-4o-mini comparing a single-agent control condition to a nine-agent pipeline under three independently varied pressure dimensions. We find no measurable difference in how often biased outcomes occur between the two conditions (6.9% vs. 6.1%, p = 0.498). We do find a large and significant effect of audit capacity on whether bias is caught: 30.0% of biased outcomes go entirely undetected, rising to 43.8% when the auditor is overloaded and falling to 18.4% when it is not. Decomposing this effect shows it is driven almost entirely by coverage (whether a case is reviewed at all, which collapses from 100.0% to 65.6% under load, p < 0.001) rather than by degraded judgment on the cases that are reviewed (81.6% vs. 85.7%, p = 1.000, direction reversed). A follow-up experiment shows that reordering the audit queue by estimated risk, rather than first-come-first-served, recovers most of the lost coverage under the same capacity constraint (65.6% to 91.7%, p = 0.028). We discuss the implications for any system that adds independent oversight to an LLM agent pipeline under resource constraints, and report the study's limitations honestly: one model, modest sample sizes, and no adversarial replication.

Paul-Peter Arslan

Institute for Future Technologies

LLM Auditing Language Model Bias Evaluation Demographic Bias in Language Models Multi-Agent LLMs

Ask PDF

Apr 15, 2026

Fairness in language models is typically studied as a property of a single, centrally optimized model. As large language models become increasingly agentic, we propose that fairness emerges through interaction and exchange. We study this via a controlled hospital triage framework in which two agents negotiate over three structured debate rounds. One agent is aligned to a specific ethical framework via retrieval-augmented generation (RAG), while the other is either unaligned or adversarially prompted to favor demographic groups over clinical need. We find that alignment systematically shapes negotiation strategies and allocation patterns, and that neither agent's allocation is ethically adequate in isolation, yet their joint final allocation can satisfy fairness criteria that neither would have reached alone. Aligned agents partially moderate bias through contestation rather than override, acting as corrective patches that restore access for marginalized groups without fully converting a biased counterpart. We further observe that even explicitly aligned agents exhibit intrinsic biases toward certain frameworks, consistent with known left-leaning tendencies in LLMs. We connect these limits to Arrow's Impossibility Theorem: no aggregation mechanism can simultaneously satisfy all desiderata of collective rationality, and multi-agent deliberation navigates rather than resolves this constraint. Our results reposition fairness as an emergent, procedural property of decentralized agent interaction, and the system rather than the individual agent as the appropriate unit of evaluation.

Sayan Kumar Chaki, Antoine Gourru, Julien Velcin

Laboratoire Hubert Curien, UMR CNRS 5516, Saint-Étienne, France · École Centrale de Lyon, LIRIS CNRS UMR 5205

Demographic Bias in Language Models Multi-Agent Systems Multi-Agent Debate

Ask PDF

Jun 15, 2026

Large language model (LLM) agents increasingly take actions (screening applicants, recommending credit, triaging patients), yet fairness for LLMs is still measured by grading answers. We introduce AgentFairBench, a cheap, reproducible, multi-domain benchmark for demographic disparity in the actions of LLM agents. Grounded in a companion framework, the Bias Conduction Framework (BCF, restated here), it spans three regulator-anchored domains: hiring, lending, and medical triage. Synthetic, demographic-neutral profiles are evaluated in counterfactual matched sets that vary only a name-coded race x gender signal (in the Bertrand Mullainathan tradition), under four agent scaffolds of increasing agency (direct, chain-of-thought, multi-agent deliberation, tool-augmented). A NumPy-only harness computes counterfactual flip rate, mean absolute score difference (MASD), action-rate disparity, and tool-invocation disparity, with bootstrap confidence intervals, paired tests, and false-discovery-rate control, for single-digit dollars per model. A live leaderboard with a held-out private split and a contamination canary admits external models by submission. Our pilot (864 decisions plus a test-retest replication) carries a methodological lesson: comparing a six-group score spread against a two-run noise difference overstates disparity by ~ 2.4X through statistic arity alone. Against an arity matched noise floor and an omnibus group test, claude haiku 4 5 shows no demographic effect above sampling noise (0 of 120 pairwise and 0 of 9 omnibus contrasts survive correction); a planted-bias test confirms the instrument detects disparity when present. The contribution is a sound, sensitive, adoption-ready instrument, the arity matched null methodology, and open artifacts to scale it. Code, data, and harness are released under open licenses, with an anonymized review artifact.

Triveni Morla, Rohith Reddy Bellibaltu, Manpreet Singh +1

Florida International University, Miami, FL 33199 USA · Boston University, Boston, MA 02215 USA · Department of Computer Science and Engineering, Indian Institute of Technology Patna, Patna, Bihar, India

Algorithmic Fairness LLM Agent Evaluation Counterfactual Fairness AI Agent Benchmarks Demographic Bias in Language Models

Ask PDF

Agent (version seen)