Fair Like Us? Auditing LLM Alignment in Resource Allocation¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Safety
Keywords: resource allocation fairness, LLM alignment, envy-freeness, experimental game theory, first-person auditing
TL;DR¶
This paper introduces the first strictly controlled audit of large language models' first-person fairness judgments in indivisible resource allocation, revealing that while LLMs mirror human axiomatic hierarchies between envy-freeness and proportionality, they enforce drastically more rigid fairness thresholds, exhibit zero altruism in bundle swaps, drift criteria under framing changes, and merely suffer modal collapse rather than true normative alignment under supervised fine-tuning.
Background & Motivation¶
Fair division of scarce, indivisible resources forms the cornerstone of numerous multi-agent systems and societal allocation mechanisms, spanning public housing assignments, shared cluster compute scheduling, and estate division. As large language models (LLMs) are increasingly explored as autonomous decision-makers, mediator agents, or proxies for human preferences, their natural language reasoning capabilities make them an appealing bridge between unstructured stakeholder inputs and structured economic outcomes. However, whereas classical decision-support systems rely on transparent, formally verifiable axiomatic rules, LLM-based systems operate without explicit normative constraints, are known to suffer from contextual framing sensitivities and hallucinations, and leave it fundamentally ambiguous which economic justice criteria they actually embody.
Evaluating distributive justice is intrinsically complex because social choice theory offers multiple competing formal criteria (such as envy-freeness, proportionality, and maximin share), yet in indivisible allocations, perfectly fair allocations rarely exist, forcing human agents to accept realistic approximations. Recent behavioral economics studies demonstrate that when human stakeholders evaluate their assigned bundles from a first-person perspective, they are neither purely selfish nor indifferent to formal guarantees, displaying nuanced preferences for stronger theoretical properties. Crucially, almost all prior literature examining LLMs in economic domains casts them as omniscient "third-party central planners," overlooking how models behave when placed directly in the shoes of an affected participant whose own payoff is on the line.
Understanding whether LLMs can faithfully represent human constituents requires an apples-to-apples, first-person comparative benchmark under identical preference structures and elicitation treatments. The core idea is to audit LLM first-person fairness reasoning against matched human participants across formal axiomatic tiers (EF, PROP, MMS, PROP1, EF1) and three-dimensional elicitation framings, uncovering their internal decision criteria via chain-of-thought judge attribution and evaluating whether supervised fine-tuning genuinely aligns models with human normative distributions.
Method¶
Overall Architecture¶
The paper formulates an experimental auditing framework where an evaluator assesses an indivisible fair division problem instance \(\mathcal{I} = \langle N, M, V \rangle\) with \(n=4\) agents and \(m=10\) items. Agent valuations \(v_i\) are additive, with item values drawn uniformly and independently from \(\{1, 2, \dots, 50\}\). Every generated allocation \(A = (A_1, \dots, A_n)\) is mathematically constrained to satisfy exactly one formal fairness axiom from a theoretical hierarchy (comprising extreme baselines SMALL and LARGE, and canonical properties EF, PROP, MMS, PROP1, and EF1).
The evaluation protocol pairs each LLM with 150 human participants across 10 sequential scenario-treatment questions (1,500 responses per model). Five controlled experimental treatments (T1โT5) systematically vary across three design dimensions: bundle information (Personal vs. Full), valuation transparency (Private vs. Public), and question framing (Threshold-based acceptability vs. Comparison-based bundle swapping). Finally, chain-of-thought traces from reasoning models are classified via a structured LLM judge to isolate dominant decision criteria.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}%%
flowchart TD
A["Resource Allocation Instances<br/>4 agents / 10 indivisible items / additive values"] --> B["Axiomatic Fairness Hierarchy & Constraint Isolation<br/>7 discrete criteria: LARGE / EF / PROP / MMS / EF1 / PROP1 / SMALL"]
B --> C["Three-Dimensional Treatment Matrix<br/>Bundle visibility ร Value transparency ร Framing (T1-T5)"]
C --> D["First-Person Behavioral Alignment Audit<br/>Matched human comparison: Pointwise Accp and Modal Accm"]
D --> E["Swap Dynamics & Moral Character Categorization<br/>Decision tri-partition: Selfish / Pareto / Selfless swaps"]
E --> F["Reasoning Trace Attribution & Criteria Drift<br/>Structured LLM judge: PROP anchoring / EF dominance / Self-payoff"]
F --> G["Supervised Fine-Tuning & Surface Imitation Probe<br/>LoRA evaluation: Modal collapse diagnosis vs. normative alignment"]
Key Designs¶
1. Axiomatic Fairness Hierarchy & Constraint Isolation: Disentangling Entailed Theoretical Guarantees
In theoretical fair division, stronger axioms often naturally entail weaker relaxations (e.g., envy-freeness strictly implies proportionality under identical valuations, which in turn implies PROP1). If an instance satisfied multiple properties simultaneously, an empirical audit could not isolate which guarantee governed the model's judgment. To eliminate confounding overlaps, the authors engineered mutually exclusive candidate instances across seven distinct criteria: - Extrema: SMALL constrains the participant's bundle value strictly below their maximin share \(v_p(A_p) < \text{MMS}_p(M)\) while remaining one item away from reaching it; LARGE guarantees the participant at least half of the total grand bundle value \(v_p(A_p) \ge \frac{1}{2} v_p(M)\); - Canonical Separations: EF satisfies \(v_i(A_i) \ge v_i(A_h), \forall h, i\); PROP guarantees \(v_i(A_i) \ge \frac{1}{n} v_i(M)\); MMS enforces \(v_i(A_i) \ge \text{MMS}_i(M)\) while explicitly violating PROP (\(v_1(A_1) < \frac{1}{n} v_1(M)\)); PROP1 satisfies proportionality up to one item without satisfying full PROP; and EF1 satisfies envy-freeness up to one item while retaining an envy relation against at least one peer. This isolation isolates the exact boundary where an agent's perception transitions from acceptable to unfair.
2. Three-Dimensional Treatment Matrix: Deconstructing Visibility and Elicitation Framing
Fairness perceptions are heavily molded by the informational environment and task framing. Five experimental treatments (T1โT5) span three orthogonal axes: - T1 (Personal + Private + Threshold): The participant observes only their assigned bundle and their own valuations, with remaining goods pooled into a single composite set, answering: "Do you find your share acceptable? (Yes/No)"; - T2 (Full + Private + Threshold): Discloses the full allocation across all four participants while keeping valuations private, retaining the acceptability question; - T3 (Full + Private + Comparison): Uses identical information to T2 but alters the prompt to a comparative choice: "Would you like to keep the treasure assigned to you or swap with one of the other pirates?" (declining to swap is interpreted as finding the allocation fair); - T4 (Full + Public + Threshold): Discloses full bundles and reveals all agents' additive valuations for their own and the participant's bundle under threshold evaluation; - T5 (Full + Public + Comparison): Provides full bundle and public valuation transparency under the multi-agent bundle swapping format. This matrix cleanly isolates how shifting from an egocentric perspective to holistic multi-agent visibility alters the model's threshold.
3. Swap Dynamics & Moral Character Categorization: Profiling Selfishness, Pareto, and Altruism
Under treatment T5, where participants possess full information regarding how all agents value every bundle, bundle exchange decisions expose the underlying social utility function of the agent. The authors partition all possible swap choices into three mutually exclusive behavioral categories: - Selfish Swaps: Exchanges where the participant strictly increases their own payoff while the trade partner is left indifferent or worse off; - Pareto Swaps: Mutually beneficial trades where both the participant and the trade partner achieve an equal or higher utility, representing cooperative surplus creation; - Selfless Swaps: Exchanges where the trade partner gains at the direct expense of the participant's own utility loss, manifesting genuine altruism or other-regarding concern. This taxonomy directly tests whether models behave as hyper-rational Homo economicus or exhibit human-like empathy.
4. Reasoning Trace Attribution & Criteria Drift: Diagnosing Chain-of-Thought Decision Rules
To uncover the latent decision rules driving final outputs, 4,500 chain-of-thought traces (<think> blocks) from DeepSeek-R1, DeepSeek-70B, and Phi-4-14B were audited using GPT-OSS-120B as a structured judge (with DeepSeek-R1 providing independent inter-annotator validation, achieving Cohen's \(\kappa = 0.70\)). The judge mapped each trace onto a controlled vocabulary comprising the seven formal criteria plus three heuristic labels: self-payoff (focusing solely on absolute bundle value), equal-split (demanding uniform numerical item count), and none. To prevent colloquial semantic drift, formal mathematical definitions of all criteria were embedded directly into the judge's prompt. The analysis revealed that LLMs do not faithfully evaluate the mathematical properties of an allocation; instead, question framing dictates the reasoning anchorโthreshold prompts steer models toward PROP or EF checks, whereas swap prompts collapse reasoning into raw self-payoff maximization.
Key Experimental Results¶
Main Results¶
The benchmark evaluated seven primary language models across varying parameter scales and reasoning architectures: lightweight models Phi-4-14B and Gemma-3-27B; distilled reasoning models DeepSeek-70B and Gemini-2.5-Flash; and advanced-reasoning models DeepSeek-R1 (671B), Gemini-2.5-Pro, and GPT-5.6-Terra. Baseline human data comprises 1,500 decisions from 150 participants evaluated on pointwise accuracy (\(\text{Acc}_p\)) and modal consensus accuracy (\(\text{Acc}_m\)).
Table 1 summarizes overall alignment metrics against human judgments and the percentage of allocations perceived as fair across canonical fairness properties (compiled from Table 2 and Table 4 of the paper):
| Model / Evaluator | Pointwise Acc \(\text{Acc}_p\) (%) | Modal Acc \(\text{Acc}_m\) (%) | EF Fair Rate (%) | PROP Fair Rate (%) | MMS Fair Rate (%) | PROP1 Fair Rate (%) | EF1 Fair Rate (%) |
|---|---|---|---|---|---|---|---|
| Humans (Ground Truth) | โ | โ | 80.83 | 66.33 | 46.00 | 48.00 | 39.17 |
| Phi-4-14B | 67.40 [65.0, 69.7] | 81.87 [79.8, 83.7] | 97.08 โ | 64.00 | 19.33 โ | 21.33 โ | 21.67 โ |
| Gemma-3-27B | 55.67 [53.1, 58.2] | 68.20 [65.8, 70.5] | 74.17 | 40.67 โ | 12.33 โ | 15.33 โ | 19.58 โ |
| Gemini-2.5-Flash | 62.27 [59.8, 64.7] | 73.47 [71.2, 75.6] | 94.17 โ | 55.33 โ | 7.00 โ | 8.67 โ | 14.58 โ |
| DeepSeek-70B | 62.80 [60.3, 65.2] | 81.67 [79.6, 83.5] | 81.67 | 58.33 | 42.00 | 42.00 | 34.17 |
| DeepSeek-R1 (671B) | 55.13 [52.6, 57.6] | 60.00 [57.5, 62.5] | 95.00 โ | 53.33 โ | 11.00 โ | 12.67 โ | 12.08 โ |
| Gemini-2.5-Pro | 60.20 [57.7, 62.7] | 69.27 [66.9, 71.6] | 96.25 โ | 48.00 โ | 0.33 โ | 0.33 โ | 4.58 โ |
| GPT-5.6-Terra | 60.88 [58.4, 63.3] | 68.29 [65.9, 70.6] | 100.00 โ | 47.32 โ | 0.67 โ | 0.67 โ | 4.17 โ |
Note: Arrows indicate statistically significant increases (โ) or decreases (โ) relative to humans via two-sided Fisher's exact test (\(p < 0.05\)).
In swap dynamics under treatment T5, models and humans exhibited stark behavioral divergence (paper Table 5): | Evaluator / Model | Selfish Swaps (%) | Pareto Swaps (%) | Selfless Swaps (%) | |---|---|---|---| | Humans | 71.66 | 47.50 | 17.07 | | Phi-4-14B | 99.47 โ | 85.00 โ | 0.00 โ | | Gemma-3-27B | 67.38 | 80.00 โ | 2.44 | | Gemini-2.5-Flash | 97.33 โ | 82.50 โ | 0.00 โ | | DeepSeek-70B | 79.14 | 85.00 โ | 0.00 โ | | DeepSeek-R1 | 34.76 โ | 85.00 โ | 0.00 โ | | Gemini-2.5-Pro | 99.47 โ | 85.00 โ | 0.00 โ | | GPT-5.6-Terra | 99.47 โ | 85.00 โ | 0.00 โ |
Ablation Study & Fine-Tuning Analysis¶
To investigate whether supervised fine-tuning bridges the human-AI alignment divide, the authors trained Phi-4-14B and Gemma-3-27B with LoRA (\(r=32, \alpha=32\)) on 1,190 human responses across 298 unique instances, evaluating on a disjoint test set of 310 human responses across 74 scenarios (paper Table 14):
| Model | Stage | T1 (Solo/Private) | T2 (Full/Private/Thresh) | T3 (Full/Private/Swap) | T4 (Full/Public/Thresh) | T5 (Full/Public/Swap) | Combined Accp (%) |
|---|---|---|---|---|---|---|---|
| Gemma-3-27B | Before | 32.8 | 79.5 | 70.1 | 37.7 | 58.2 | 55.65 |
| Gemma-3-27B | After | 70.5 | 79.5 | 16.41 | 64.2 | 17.9 | 49.70 |
| Phi-4-14B | Before | 55.7 | 27.4 | 80.6 | 5.66 | 60.71 | 46.01 |
| Phi-4-14B | After | 78.7 | 79.5 | 88.1 | 64.2 | 53.6 | 72.82 |
| Human Modal Baseline | โ | 88.5 | 84.9 | 88.1 | 79.2 | 66.1 | 81.36 |
Robustness checks across narrative domains (pirates, siblings, competitors, refugees) and resources (treasure, competition awards, disaster supplies) demonstrated that the axiomatic hierarchy and human-AI divergence remained qualitatively identical across all contexts (paper Tables 6โ7). Injecting generous or selfish persona prompts produced localized, notion-specific modulations rather than uniform shifts in perceived fairness (paper Tables 8โ9).
Key Findings¶
- Advanced reasoning exacerbates normative divergence: Scaling model size and test-time reasoning tokens does not improve human alignment. Advanced-reasoning models (DeepSeek-R1, Gemini-2.5-Pro, GPT-5.6-Terra) achieve lower pointwise alignment (\(\sim 55\% - 60\%\)) than mid-sized models (Phi-4 at 67.4%). Reasoning models enforce an extreme, uncompromising fairness standard: they accept 95%โ100% of envy-free (EF) allocations but reject virtually all relaxed allocations (MMS, PROP1, and EF1 rates plummet below 5%, with Gemini-2.5-P and GPT-5.6-Terra accepting MMS/PROP1 at just 0.33%โ0.67%), whereas humans accept such approximations 40%โ48% of the time.
- Cold economic rationality and absence of altruism: LLMs consistently exploit mutually beneficial trading opportunities, selecting Pareto swaps in \(>80\%\) of eligible scenarios compared to only 47.5% for humans. Conversely, while 17.07% of human choices demonstrate selfless generosity by absorbing personal utility losses to help a peer, LLMs (with the lone exception of Gemma at 2.44%) exhibit an absolute 0.00% selfless swap rate.
- Fine-tuning causes modal collapse rather than moral alignment: Although LoRA fine-tuning boosted Phi-4's overall accuracy from 46.0% to 72.8%, error analysis reveals this gain is entirely illusory: the model simply collapsed to predicting "Yes" across all threshold prompts to emulate the majority class, without distinguishing fair from unfair bundles. On comparison-based swap tasks, fine-tuning caused Gemma's accuracy to plunge by \(>40\%\), proving that existing human fair division datasets are insufficient for authentic distributional alignment.
Highlights & Insights¶
- First-Person Diagnostic Paradigm: Moving beyond the conventional third-party social planner setup exposes the intense self-interest, rigid thresholds, and lack of benevolence inherent in current frontier LLMs when their own outcomes are at stake.
- The Paradox of Algorithmic Rigidity: In realistic indivisible allocation problems where perfect envy-freeness is mathematically impossible, human societies rely on compromises like MMS and EF1. Advanced LLMs' refusal to accept these valid relaxations presents a severe risk of algorithmic deadlock in automated mediation or resource triage.
- Framing Overrules Objective Guarantees: An allocation's mathematical fairness properties do not govern model behavior; rather, the syntactic framing of the query (threshold acceptability vs. comparative swap) determines whether the model activates a fairness checking heuristic or raw greedy utility maximization.
Limitations & Future Work¶
- Geographic and Cultural Homogeneity: The human participant benchmark is drawn primarily from North American crowdsourced pools, limiting cultural generalization across divergent concepts of distributive justice.
- Restricted Economic Assumptions: The study focuses strictly on additive valuations across four symmetric agents without incorporating real-world equity considerations like need, merit, entitlement, or historical disadvantage.
- Future Agenda: Research must move beyond scalar point-prediction fine-tuning toward distributional alignment algorithms that preserve human moral pluralism and legitimate disagreement.
Related Work & Insights¶
- vs Hosseini et al. (2025a): Established the human empirical benchmark for first-person fair division; this work bridges that foundation into AI evaluation, executing direct apples-to-apples human-AI auditing under matched treatments.
- vs Hosseini & Khanna (2025); Hua et al. (2024): Examined LLMs as third-party arbiters or game-theoretic negotiators; this paper demonstrates that adopting a first-person stakeholder persona drastically amplifies self-interested utility optimization and axiomatic rigidity.
Rating¶
- Novelty: โญโญโญโญโญ Pioneering first-person auditing benchmark grounding LLM decision-making in formal social choice axioms.
- Experimental Thoroughness: โญโญโญโญโญ Comprehensive coverage of 11 LLMs, 4,500 CoT traces, 5 experimental treatments, LoRA SFT, and cross-persona robustness checks.
- Writing Quality: โญโญโญโญโญ Flawless conceptual clarity bridging theoretical economics, prompt semantics, and empirical alignment diagnostics.
- Value: โญโญโญโญโญ Critical cautionary guidance for deploying LLMs in autonomous negotiation, public resource distribution, and algorithmic triage.