SWF Benchmark Overview

• The Social Welfare Function (SWF) benchmark is a formal framework that aggregates individual utilities into a collective welfare metric for policy and AI evaluations.
• It is applied as a fairness objective, optimization target, strategic evaluation, and leaderboard metric across domains, employing methods like alpha-fair and Nash social welfare functions.
• Recent research implements SWF benchmarks in dynamic simulations, empirical elicitation, and constrained optimization to address both efficiency and normative challenges.

Searching arXiv for recent and foundational papers on SWF benchmarks and related formulations. A Social Welfare Function (SWF) benchmark is a formal evaluative setup in which a social welfare function is the reference object for comparing allocations, policies, rankings, strategic mechanisms, or learned decision rules. In the literature represented here, SWFs appear in at least four benchmark roles: as the primary fairness objective in AI systems; as the optimization target in allocation and policy-learning problems; as the efficiency-and-fairness benchmark against which strategic mechanisms are evaluated; and as an explicit leaderboard metric for LLM allocators (Chen et al., 2021, Brânzei et al., 2016, Pardeshi et al., 2024, Shi et al., 1 Oct 2025).

1. Benchmark concept and scope

The benchmark role of an SWF is not confined to one domain. In fairness-aware AI, social welfare optimization is proposed as “a general paradigm for formalizing fairness in AI systems,” with welfare defined over stakeholders and optimized either directly over decisions or jointly with prediction models (Chen et al., 2021). In group-fairness analysis, an SWF serves as the normative standard against which demographic parity, equalized odds, and predictive rate parity are assessed, rather than merely as another scalar metric (Chen et al., 2024). In strategic allocation, the weighted Nash social welfare is the natural benchmark induced by Fisher markets, and the central question becomes how closely equilibrium behavior approximates that benchmark (Brânzei et al., 2016). In LLM evaluation, the “Social Welfare Function (SWF) Benchmark” is a dynamic simulation where an LLM acts as a sovereign allocator and is scored by a combined fairness-efficiency metric (Shi et al., 1 Oct 2025).

These uses share one structural feature: the benchmark is defined by an explicit mapping from individual-level consequences to a collective ordering. What varies is the object being ordered. Depending on the paper, the argument of the SWF is a utility vector, a treatment rule, a ranking, a stochastic choice rule, a type-partitioned opportunity profile, or the cumulative allocation history of an online system (Chen et al., 2020, Wienand et al., 27 Mar 2026, Echenique et al., 2024, Pardeshi et al., 1 Feb 2026).

Benchmark mode	Benchmark object	Representative papers
Optimization benchmark	Utility vector or decision rule	(Chen et al., 2021, Chen et al., 2024, Chen et al., 2020)
Strategic benchmark	Equilibrium approximation to optimal NSW	(Brânzei et al., 2016)
Learning and inference benchmark	Recoverable or estimable welfare criterion	(Pardeshi et al., 2024, Sasaki et al., 2020, Terschuur, 19 Feb 2025, Pardeshi et al., 1 Feb 2026)
Leaderboard benchmark	Allocator behavior under fairness-efficiency trade-off	(Shi et al., 1 Oct 2025)

2. Canonical functional families

A large part of SWF benchmarking consists in fixing a functional family and then varying its parameters. The most prominent family in the cited literature is the $\alpha$-fair or isoelastic family,

$$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$

which is used both as a fairness family in AI and as a normative benchmark for group-fairness assessment. In this family, $\alpha=0$ gives utilitarian aggregation, $\alpha=1$ gives proportional fairness or Nash-style aggregation, and $\alpha\to\infty$ approaches maximin (Chen et al., 2021, Chen et al., 2024). A closely related empirical benchmark appears in survey elicitation of public views over subjective wellbeing, where the estimated isoelastic parameter is $\alpha=0.48$, implying an SWF approximately equal to the sum of square roots of individual utilities (Layard et al., 11 Jun 2026).

A second canonical family is Nash social welfare. In divisible-goods allocation, the unweighted form is

$$\mathrm{NSW}(x)=\left(\prod_{i=1}^{n}u_i(x_i)\right)^{1/n},$$

and with heterogeneous budgets $B_i$, the weighted form becomes

$$\mathrm{NSW}(x)=\left(\prod_{i=1}^{n}u_i(x_i)^{B_i}\right)^{1/\mathcal B}, \qquad \mathcal B=\sum_i B_i.$$

This weighted geometric-mean objective is the benchmark induced by Fisher equilibria and underlies price-of-anarchy analysis for strategic mechanisms (Brânzei et al., 2016).

Other benchmark families are designed to encode more specialized normative structure. The leximax-utilitarian family introduced for optimization under fairness-efficiency trade-offs uses a parameter $\Delta$ that defines a “fair region” consisting of those within $$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$0 of the worst-off. At $$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$1 the benchmark collapses to utilitarianism; for sufficiently large $$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$2 it yields pure leximax (Chen et al., 2020). Opportunity-sensitive welfare applies expected utility within circumstance-defined types and then aggregates type welfare by an entropic second-stage transform,

$$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$3

so that $$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$4 recovers utilitarianism and $$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$5 yields maximin over types (Wienand et al., 27 Mar 2026). Under uncertainty, Belief-Averaged Relative Utilitarianism aggregates beliefs and utilities separately: social belief is the average of individual beliefs, while social utility is the sum of unit-range-normalized utilities (Brandl, 2020).

3. Optimization and construction of benchmark problems

Many SWF benchmarks are posed as explicit constrained optimization problems. In social welfare optimization for AI, the generic template is

$$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$6

where $$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$7 is the decision vector, $$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$8 is stakeholder $$W_{\alpha}(u) = \begin{cases} \frac{1}{1-\alpha}\sum_i u_i^{1-\alpha}, & \alpha \ge 0,\ \alpha\neq 1 \ \sum_i \log(u_i), & \alpha=1 , \end{cases}$$9’s utility, and $\alpha=0$0 aggregates the utility vector. This template is instantiated both in post-processing, where a predictive model is trained first and welfare is optimized downstream, and in in-processing, where predictive loss and welfare are combined in one objective (Chen et al., 2021).

In binary selection under scarcity, the benchmark takes the form

$\alpha=0$1

with individual utility $\alpha=0$2. The crucial quantity is the welfare differential

$\alpha=0$3

and the optimum is obtained by selecting the $\alpha=0$4 individuals with the largest $\alpha=0$5. In that benchmark, demographic parity and equalized odds are not primitive objectives; they arise only under specific equal-marginal conditions (Chen et al., 2024).

The leximax-utilitarian benchmark uses sequential mixed-integer optimization. The first-stage objective is

$\alpha=0$6

and higher-order objectives $\alpha=0$7 refine the welfare ordering among the disadvantaged part of the utility distribution. The benchmark is therefore not a single-shot scalarization but a sequence of optimization problems that determine the low end of the ordered utility vector before reverting to utilitarian aggregation outside the fair region (Chen et al., 2020).

Mechanism-design benchmarks use a different construction. In strategic allocation, equilibrium outcomes are evaluated by

$\alpha=0$8

where $\alpha=0$9 is the optimal weighted-NSW allocation and $\alpha=1$0 is the equilibrium set. The benchmark is thus the optimal SWF value itself, and the object of study is how much welfare is lost under strategic behavior (Brânzei et al., 2016).

4. Learning, estimation, and online inference

A major recent development is the use of SWFs as learnable or estimable objects. One line of work studies direct recovery of an SWF from observed policy-maker judgments. For the weighted power-mean family

$\alpha=1$1

the paper on learning SWFs considers both regression from observed welfare scores and classification from pairwise comparisons. It proves polynomial sample complexity in both settings, with pseudo-dimension $\alpha=1$2 when weights are known and below $\alpha=1$3 when weights are unknown; for the comparison class, the VC dimension is below $\alpha=1$4 with known weights and below $\alpha=1$5 with unknown weights (Pardeshi et al., 2024).

A second line of work studies welfare identification under econometric endogeneity. In treatment assignment with an instrumental variable, the mean social welfare of policy $\alpha=1$6 is

$\alpha=1$7

and the key representation result is

$\alpha=1$8

Here the marginal treatment effect becomes the operator kernel that identifies welfare and supports plug-in, Bayes, and empirical welfare maximization rules (Sasaki et al., 2020).

More general policy learning with semiparametric SWFs extends this logic beyond mean outcomes. The relevant welfare may depend on nonlinear transforms, predicted outcomes $\alpha=1$9, pairwise absolute differences, or even Kendall-$\alpha\to\infty$0-type dependence terms. The paper on locally robust policy learning derives orthogonal-score estimators for both linear SWFs and U-statistic SWFs, then proves regret bounds over policy classes with controlled VC complexity (Terschuur, 19 Feb 2025).

In online allocation, the SWF itself enters the bandit objective. The framework for SWF-based online resource allocation assumes any monotonic, concave, and Lipschitz-continuous $\alpha\to\infty$1, then lifts individual confidence sequences to anytime-valid confidence sequences for optimal welfare. Its SWF-UCB algorithm optimizes

$\alpha\to\infty$2

at each round and achieves near-optimal $\alpha\to\infty$3 regret. The framework is instantiated for Weighted Power Mean, Kolm, and Gini welfare families (Pardeshi et al., 1 Feb 2026).

5. Empirical benchmark implementations

Empirical SWF benchmarking appears in several distinct forms. One is direct elicitation of a welfare function. Using a representative UK sample of $\alpha\to\infty$4, the public-SWF paper estimates an isoelastic social welfare function over life satisfaction and reports a median inequality-aversion parameter $\alpha\to\infty$5 with standard error $\alpha\to\infty$6. The median respondent values improving the wellbeing of the least satisfied by one unit roughly twice as much as improving the most satisfied by one unit, and the paper offers both the smooth isoelastic benchmark and the nonparametric weights $\alpha\to\infty$7 for life-satisfaction steps $\alpha\to\infty$8 through $\alpha\to\infty$9 (Layard et al., 11 Jun 2026).

Another form is application-specific welfare optimization. In mortgage-loan processing on the German credit dataset, social welfare optimization is combined with logistic regression through both post-processing and in-processing. The experiments use 5 random 80/20 train-test splits, a budget $\alpha=0.48$0, and age-group analysis. Welfare-sensitive objectives improve group parity relative to the utilitarian baseline, but the paper explicitly reports that improving fairness costs efficiency in the post-processing setting and that in-processing reduces test accuracy only slightly (Chen et al., 2021).

Analytical benchmark scenarios are also used. In the group-fairness paper, stylized distributions for protected and nonprotected groups show that demographic parity can arise at a specific $\alpha=0.48$1, that equalized odds may collapse into accuracy when the selection quota matches the qualified share, and that predictive rate parity is of limited usefulness. In Scenario 3, where some protected individuals are harmed by selection, no value of $\alpha=0.48$2 yields demographic parity when $\alpha=0.48$3, because alpha-fairness never endorses selecting individuals with negative welfare differential merely to equalize rates (Chen et al., 2024).

The most explicit benchmark in the narrow sense is the LLM “Social Welfare Function Benchmark.” It is a dynamic simulation with 12 heterogeneous recipient agents, 63 task allocation cases, and 50 tasks per case, where an allocator LLM repeatedly decides who gets the next task. Efficiency is measured by

$\alpha=0.48$4

fairness by $\alpha=0.48$5, and the leaderboard score by

$\alpha=0.48$6

The paper evaluates 20 state-of-the-art LLMs and reports three headline findings: general conversational ability is a poor predictor of allocation skill, most models exhibit a strong default utilitarian orientation, and allocation strategies are highly vulnerable to output-length constraints and social-influence framing (Shi et al., 1 Oct 2025).

6. Normative assumptions, limits, and controversies

SWF benchmarks are powerful precisely because they force normative choices into the open. The literature repeatedly emphasizes that utility specification is a bottleneck. In AI fairness, individual utilities are cardinal, manually engineered, outcome-based, and aggregated under an assumption of interpersonal comparability; the authors explicitly note the need for context-specific utility and social welfare definitions (Chen et al., 2021). In wellbeing-based policy appraisal, the empirical SWF rests on explicit assumptions of cardinality and interpersonal comparability for life-satisfaction scores, and the paper states that it proceeds under the assumption that life satisfaction, like income, has these characteristics (Layard et al., 11 Jun 2026).

A second controversy concerns what kind of “equity” an SWF is supposed to encode. The welfare-economics perspective paper argues that economists often simplify analysis by assuming homogeneous, consequentialist, and self-centered preferences, while actual personal and social preferences may be heterogeneous. It also stresses that “equity” has multiple formal interpretations: an implication of welfare maximization, an independent criterion, or a lexicographic constraint, among others (Manski, 14 Jan 2025). A related critique appears in the social-choice-function paper, which argues that Harsanyi-style utilitarian aggregation may overlook distributional considerations and motivates quantile-based welfare measures derived from social choice functions rather than standard SWFs (Echenique et al., 2024).

A third limit is formal impossibility. In Arrow-style social choice with voter qualifications, the paper on qualified voters proves that if a transitive valued social welfare function satisfies independence of irrelevant alternatives and the Pareto principle, then a dictator who is qualified to evaluate all alternatives exists. It further shows that if no voter is qualified to evaluate all alternatives, then under a transitive valued social welfare function satisfying weak Pareto and independence of irrelevant alternatives, all alternatives are indifferent for any preference profile (Okumura, 2023). This places hard structural constraints on what an SWF benchmark can demand in partially informed collective evaluation.

Finally, benchmark adequacy itself is contested. Group parity metrics are criticized for being normatively thin, mutually incompatible, and insensitive to welfare magnitudes (Chen et al., 2024). Leaderboard-style SWF evaluation of LLMs operationalizes fairness as equality of task counts and efficiency as reward per normalized compute cost, which is precise but narrow (Shi et al., 1 Oct 2025). This suggests that SWF benchmarks are best understood not as value-neutral measurement devices but as explicit normative models whose usefulness depends on the appropriateness of their utility representation, aggregation rule, and deployment context.