When Fairness Evolves: Value Orientation and Evolution in Multi-Agent Resource Allocation¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/gaosong1215/when_fairness_evolves
Area: LLM Reasoning
Keywords: Multi-Agent Simulation, Resource Allocation, Fairness Evolution, Negotiation Mechanism, Value Orientation
TL;DR¶
This paper constructs a closed-loop survival-production multi-agent simulation using LLMs to demonstrate that deliberative negotiation endogenously resolves the fairness-efficiency tradeoff, uncovering systematic misalignments between subjective satisfaction and objective resource distribution while tracing the semantic evolution of fairness positions toward hybrid normative principles.
Background & Motivation¶
Resource allocation across scarce goods represents a foundational governance challenge spanning organizational compensation, inheritance partitioning, and national welfare architecture. Social science has long established that distributive justice is not a singular, static ideal but rather a constellation of competing principles including equality, individual need, merit-based desert, and procedural legitimacy. In real-world societies, individuals hold heterogeneous normative orientations shaped by cultural identities and moral convictions. This pluralism induces a core systemic tension: prioritizing merit and productivity incentives often jeopardizes the baseline security of vulnerable groups, whereas strictly enforcing needs-based guarantees can dampen reinvestment incentives and overall community output, locking static allocation mechanisms into an intractable fairness-efficiency tradeoff.
Compounding this challenge is the pervasive divergence between subjective evaluation and objective distribution properties. Individuals' satisfaction with an allocation frequently misaligns with quantifiable metrics such as material sufficiency and proportionality, distorted by motivated reasoning, social comparison, and prior expectations. Consequently, social groups systematically exhibit cognitive biases such as false satisfaction (where disadvantaged households express high contentment) or excessive criticism (where well-rewarded actors express severe dissatisfaction). Traditional game-theoretic formulations and static behavioral economics experiments remain largely incapable of tracing how diverse normative convictions adapt, compromise, and resolve this dual gap over long-term repeated interactions.
The emergence of large language model (LLM) agents equipped with rich background knowledge, persona enactment, and natural language deliberation offers an unprecedented computational vehicle for simulating micro-level social interactions. Grounded in this methodological opportunity, this paper instantiates an artificial society of 30 heterogeneous LLM agents participating in a 20-round survival-production cycle across equal, needs-based, contribution-based, and negotiation-based allocation mechanisms. Core idea: by closing the loop between resource circulation and satisfaction-modulated production, multi-stage language deliberation endogenously reconciles the fairness-efficiency tradeoff without pre-programmed global optimization, while semantic tracking reveals an evolutionary trajectory from isolated moral commitments toward hybrid, pluralistic fairness justifications.
Method¶
Overall Architecture¶
The simulation environment models an artificial community composed of 30 agent families, each characterized by household composition \((L_i, N_i)\) (laborers and non-laborers), a strict survival threshold \(C_i\), and an assigned normative persona \(v_i\). Across discrete rounds \(t\), the community resource pool \(R_t\) circulates through four sequential phases: allocation under a chosen mechanism, survival consumption, satisfaction-modulated production reinvestment, and pool replenishment. Crucially, each family's subjective satisfaction score—derived from an LLM-generated reflection on the perceived fairness of its allocation—linearly scales its productive conversion efficiency, ensuring that perceived injustice imposes direct material consequences on macro-level economic growth.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
A["Community Resource Pool R_t<br/>(Scarce / Moderate / Abundant)"] --> B["Stage 1: Proposal Generation<br/>(Households generate complete allocation plans based on personas)"]
subgraph DELIB["Deliberation & Consensus Pipeline"]
direction TB
B --> C["Stage 2: Discussion, Defense & Multi-Stage Voting<br/>(Iterative peer critiques with bottom-50% pruning per round)"]
C --> D["Stage 3: Weighted Consensus Formation<br/>(Weighted aggregation of remaining top-rated proposals)"]
end
D --> E["Stage 4: Subjective Satisfaction & Reflection<br/>(LLM evaluates allocation, outputs 1-5 score mapped to multiplier s_i)"]
E --> F["Survival Consumption & Production Reinvestment<br/>(Check C_i; net surplus processed under capacity constraints)"]
F --> G["Resource Pool Update R_{t+1}<br/>(Produced output and carry-over surplus replenish pool)"]
G --> A
Key Designs¶
1. Heterogeneous Value Personas with Closed-Loop Production-Survival Constraints
To avoid agent homogeneity and ensure realistic economic friction, the population incorporates five idealized normative orientations: Altruistic, Egalitarian, Merit-based, Needs-based, and Pragmatic, evenly divided into six families per profile. Each family faces a hard survival consumption constraint \(C_i = 2L_i + N_i\), where laborers require twice the sustenance of dependents. If an allocation fails to meet this threshold (\(a_i(t) < C_i\)), the household suffers an existential deficit and the simulation terminates immediately, embedding catastrophic existential risk into the system. Surviving households reinvest their net surplus \(\text{surplus}_i(t) = a_i(t) - C_i\) into production. Household processing capacity is bounded by its labor force and community-wide processing ceiling \(M = S/L\), yielding total output:
where \(E\) denotes base technological efficiency, and \(s_i(t) \in [0.9, 1.1]\) represents the productivity multiplier linearly mapped from the LLM's subjective satisfaction score \(S_i(t) \in [1, 5]\). This formulation directly bridges internal normative perceptions and objective systemic replenishment.
2. Four-Stage Procedural Deliberation and Consensus Pipeline
To transcend rigid distributional formulas, the negotiation-based mechanism operationalizes procedural justice through a structured four-stage deliberative protocol: 1. Proposal Generation: Each family proposes a complete 30-household allocation plan reflecting its normative commitments, current pool abundance, and past production history; 2. Discussion and Voting: Agents engage in multi-turn dialogues, defending their proposals, offering constructive peer critiques, and identifying acceptable compromise margins. In each iteration, candidate plans are rated from 1 to 5, and only the top 50% advance, continuing until fewer than 10% of plans remain unresolved; 3. Consensus Formation: Rather than adopting an authoritarian single-winner outcome or unweighted voting, the final distribution is synthesized as a weighted average across the surviving high-ranking proposals; 4. Final Evaluation: All households inspect the finalized allocation and generate reflective verbal explanations alongside numerical satisfaction scores. This protocol enables hybrid fairness standards to emerge organically from peer interaction rather than algorithmic fiat.
3. Three-Dimensional Embedding Space and Cognitive Misalignment Metrics
To track long-term semantic drift in expressed fairness principles without relying on superficial keyword frequencies, the methodology maps agent reflection texts into a three-dimensional vector space spanning Equality, Merit, and Needs. Using the BAAI/bge-large-en-v1.5 encoder, agent reflections after each round are compared via cosine similarity against prototypical anchor sentences representing each dimension's centroid, producing the coordinate triple \((s_{\text{eq}}, s_{\text{merit}}, s_{\text{needs}})\). Semantic change is quantified across magnitude (Euclidean displacement), directional drift, and dimensional emergence. Concurrently, by comparing round-level percentile ranks of objective resource share \(\tilde{O}_{i,t}\) and subjective satisfaction \(\tilde{S}_{i,t}\), the framework formalizes two canonical cognitive deviations: - False Satisfaction: \(\tilde{S}_{i,t} \in [P_{75}, P_{100}] \land \tilde{O}_{i,t} \in [P_0, P_{25}]\), characterizing underprivileged agents exhibiting anomalous contentment; - Excessive Criticism: \(\tilde{S}_{i,t} \in [P_0, P_{25}] \land \tilde{O}_{i,t} \in [P_{75}, P_{100}]\), characterizing resource-rich agents expressing severe discontent.
Benchmark & Evaluation Protocol¶
The empirical investigation addresses three primary research questions (RQ1: fairness-efficiency synergy; RQ2: subjective-objective misalignment; RQ3: fairness language evolution). Experiments contrast Equal, Needs-based, Contribution-based, and Negotiation mechanisms. Primary simulations employ GPT-4o-2024-11-20 across 5 independent runs per condition up to 20 rounds, while subjective satisfaction robustness is validated across seven additional models (GPT-5-Mini, Claude-3.5-Sonnet, Claude-Sonnet-4, DeepSeek-V3.2, Gemini-2.5-Flash, Llama-2-70B, and Qwen-2.5-72B). System state is quantified via objective metrics (Productive Return Ratio PRR, Gini coefficient) and subjective metrics (Satisfaction Score SS, score standard deviation Std). Additionally, the 16 health-care resource allocation scenarios from Ahlert & Schwettmann (2017) are reconstructed to benchmark GPT-4o's structural sensitivity against empirical human choice distributions.
Key Experimental Results¶
Main Results¶
Under moderate resource conditions, the macroeconomic and normative performance of the 30-agent artificial society across allocation mechanisms is summarized below (quoted from Table 1 and Table 2 in the paper):
| Allocation Mechanism | PRR (%) ↑ | Gini ↓ | Satisfaction Score (SS) ↑ | Score Std ↓ | Survival Horizon |
|---|---|---|---|---|---|
| Equal | 75.3 ± 1.3 | 0.000 ± 0.000 | 2.37 ± 0.03 | 0.68 ± 0.06 | 20 rounds |
| Needs-based | 69.8 ± 1.7 | 0.126 ± 0.001 | 2.52 ± 0.09 | 0.78 ± 0.06 | 20 rounds |
| Contribution-based | 86.6 ± 2.3 | 0.113 ± 0.002 | 2.82 ± 0.08 | 0.76 ± 0.04 | 20 rounds |
| Negotiation | 91.1 ± 1.9 | 0.064 ± 0.002 | 3.03 ± 0.04 | 0.60 ± 0.01 | 20 rounds |
Across all eight tested LLMs, negotiation consistently produces the highest subjective satisfaction (GPT-4o: 3.03, Claude-3.5-Sonnet: 3.12, Claude-Sonnet-4: 3.16, DeepSeek-V3.2: 3.04, Gemini-2.5-Flash: 3.05, Qwen-2.5-72B: 3.02), confirming that the deliberative consensus advantage is universal across contemporary frontier architectures.
Ablation Study¶
To dissect whether negotiation gains stem from substantive procedural consensus rather than a mere cathartic "voice effect", the authors evaluated an Equal+Discussion baseline alongside component-level ablations of the negotiation pipeline (quoted from Table 5):
| Experimental Variant | PRR (%) ↑ | Gini ↓ | Satisfaction Score (SS) ↑ | Score Std ↓ | Persistence (Rounds) |
|---|---|---|---|---|---|
| Full Negotiation | 84.2 ± 1.1 | 0.065 ± 0.001 | 3.28 ± 0.04 | 0.63 ± 0.03 | 20–20 |
| w/o Iterative Filtering | 83.3 ± 1.8 | 0.064 ± 0.001 | 2.72 ± 0.03 | 0.69 ± 0.01 | 8–9 |
| w/o Weighted Averaging | 84.6 ± 3.1 | 0.061 ± 0.006 | 2.65 ± 0.06 | 0.71 ± 0.02 | 8–9 |
| Equal + Discussion | 79.0 ± 3.2 | 0.000 ± 0.000 | 2.65 ± 0.03 | 0.66 ± 0.03 | 6–7 |
| Static Voting | 81.1 ± 2.9 | 0.069 ± 0.003 | 2.50 ± 0.06 | 0.67 ± 0.03 | 7–8 |
Furthermore, semantic displacement analysis across the 20 rounds (Table 4) reveals pronounced shifts toward hybrid normative reasoning: Merit-based agents exhibit a -2.70% decline in Merit proximity alongside a +5.48% surge in Needs proximity; Egalitarian agents increase Needs proximity by +2.77%; Pragmatic agents increase Needs proximity by +2.14%.
Key Findings¶
- Deliberation Navigates the Fairness-Efficiency Tradeoff: Full negotiation attains the highest productive return ratio (91.1%) and the highest subjective satisfaction (3.03) with minimal Gini dispersion (0.064), demonstrating that procedural dialogue can balance meritocratic incentives with vulnerability protection without an externally prescribed optimization target.
- Genuine Procedural Influence Trumps Mere Voice: Equal+Discussion marginally improves satisfaction over static equal division (2.37 to 2.65), but fails to match full negotiation (3.28) and collapses early (6–7 rounds), proving that deliberation must causally govern the final allocation rule to yield enduring legitimacy.
- Cognitive Misalignments Are Conditioned by Values and Time: Merit-based agents manifest the highest excessive criticism due to acute sensitivity to effort-reward parity, while altruistic agents exhibit pronounced false satisfaction by prioritizing communal resilience. Over repeated rounds, excessive criticism declines as resource constraints become salient, while false satisfaction gradually rises.
- Evolution Toward Emergent Hybrid Norms: Rather than collapsing into uniform consensus, agents retain their assigned persona commitments while expanding their lexical and semantic reasoning to incorporate multi-criteria justifications combining survival guarantees, effort incentives, and system sustainability.
Highlights & Insights¶
- Bi-Directional Feedback Between Social Cognition and Production: The framework innovatively models subjective satisfaction not merely as a passive readout, but as an active economic driver that scales real output through a conservative productivity multiplier, operationalizing psychological well-being as a macro-level production factor.
- Rigorous Embedding-Based Normative Vector Space: By utilizing sentence transformers and prototypical semantic anchors, the study quantifies moral and linguistic drift in continuous coordinates, establishing a repeatable methodological blueprint for measuring ideological adaptation in agentic societies.
- Structural Benchmarking Against Human Empirical Data: Benchmarking GPT-4o on 16 published healthcare allocation scenarios reveals positive rank correspondence (mean Spearman \(\rho = 0.665\), modal match in 9/16 cases), while highlighting a critical behavioral discrepancy: LLM choices exhibit extreme normative concentration (modal share 78.2% vs. 46.0% in humans), establishing essential epistemic guardrails for social simulation.
Limitations & Future Work¶
- Constrained Tri-Dimensional Normative Representation: Expressed fairness is mapped along three classic axes (equality, merit, need), omitting essential sociopolitical dimensions such as relational reciprocity, historical entitlements, or retributive justice.
- Heuristic Productivity Multiplier Formulation: While grounded in labor economics meta-analyses, the linear satisfaction-productivity mapping \([0.9, 1.1]\) represents an engineered modeling assumption rather than an empirically validated behavioral law.
- Absence of Institutional Power Asymmetries: All agents interact under egalitarian communication bandwidth and symmetric voting weights, lacking the structural power imbalances, strategic rent-seeking, and formal enforcement apparatus characteristic of real-world institutions.
Related Work & Insights¶
- vs. SRAP-Agent (Ji et al., 2024): Whereas SRAP-Agent optimizes resource distribution via rigid external heuristics such as queuing and priority sorting, this paper demonstrates that decentralized multi-turn linguistic negotiation can endogenously discover Pareto-superior, consensus-driven allocation policies.
- vs. Mathematical Distributive Optimization (Joe-Wong et al., 2013; Moulin, 2004): Traditional fair division frameworks solve for static mathematical optima under fixed social welfare functions; this work models fairness as an evolutionary, communicative equilibrium that dynamically adapts under subjective feedback.
- vs. LLM Social Simulation Validity Critiques (Wang et al., 2025; Zhou et al., 2025): Addressing widespread concerns regarding sycophantic convergence and persona drift, this study imposes physical survival termination constraints, sliding-window memory buffers, and multi-model robustness audits, positioning the simulator as a rigorous hypothesis generator.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Pioneering computational exploration of emergent fairness norms and subjective-objective misalignments in closed-loop multi-agent LLM societies.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation covering 4 allocation mechanisms, 3 resource regimes, 8 frontier LLMs, internal ablations, and human benchmark calibration.
- Writing Quality: ⭐⭐⭐⭐⭐ Masterful synthesis of political philosophy, social psychology, microeconomic theory, and natural language processing.
- Value: ⭐⭐⭐⭐⭐ Provides an open-source, reproducible foundation for computational social science, algorithmic governance, and multi-agent mechanism design.