Utilitarian Objective Overview

• Utilitarian Objective is a family of optimization criteria that aggregates individual utilities to maximize overall performance.
• It underpins diverse applications including social choice, algorithm configuration, reinforcement learning, and AI ethics.
• Critiques focus on its distributional blindness and fairness limitations, prompting trade-offs between efficiency and equitable outcomes.

Searching arXiv for recent and foundational papers on utilitarian objectives across social choice, algorithm configuration, MARL, and AI ethics. Utilitarian objective denotes a family of optimization criteria that evaluate actions, allocations, committees, policies, or configurations by aggregating utility across affected agents, outcomes, or users. In the materials surveyed here, the core formulation is typically additive—maximizing a sum of utilities, welfare, or expected utility—but the term is operationalized differently across domains: approval-based multiwinner voting uses the sum of voters’ approved committee members; algorithm configuration uses expected utility induced by a runtime-to-utility map; reinforcement learning uses maximization of expected cumulative reward; and multi-agent systems often use the sum of agents’ returns. Across these settings, utilitarian objectives are valued for tractability, direct alignment with aggregate performance, and quantitative comparability, while recurring criticisms concern distributional blindness, weak fairness guarantees, and vulnerability to objective misspecification (Brill et al., 2023).

1. Definition and canonical mathematical forms

The most direct form of a utilitarian objective is the maximization of total utility. In approval-based multiwinner voting, utilitarian welfare for a committee $W$ is defined as

$\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$

where $A_i$ is voter $i$’s approval set; the utilitarian guarantee is the ratio of this welfare to the welfare of the Approval Voting optimum (Brill et al., 2023). A closely related formulation appears in earlier work on approval-based rules, where the utilitarian score of a committee is

$\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$

which is equivalent to summing each voter’s number of approved committee members (Lackner et al., 2018).

In algorithm configuration, the utilitarian objective is not runtime minimization per se, but maximization of expected utility induced by a bounded, monotonically decreasing utility function $u : \mathbb{R}_{\ge 0} \to [0,1]$. For configuration $i$, expected utility is

$U_i = \mathbb{E}_{t \sim D_i}[u(t)],$

and capped expected utility is

$$U_i(\kappa) = \mathbb{E}_{t \sim D_i}[u(\min(t,\kappa))].$$

This formulation preserves user preference structure while avoiding the instability of mean-runtime objectives under rare long tails (Graham et al., 2023).

In machine learning and moral philosophy, utilitarian structure is often represented as expected-value maximization. Reinforcement learning is described as maximizing

$$\pi^* = \underset{\pi}{\operatorname{argmax}} \ \mathbb{E}\left[\sum_{t=0}^{\infty} \gamma^t r_t \mid \pi \right],$$

which has been interpreted as formally paralleling act-utilitarian decision procedures (Roff, 2020). In algorithmic fairness, the utilitarian baseline is minimizing aggregate individual loss,

$\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$0

with Rawlsian alternatives replacing the sum by a max operator over individuals (Rigobon, 2023).

2. Social choice and committee selection

In approval-based multiwinner voting, the utilitarian objective is represented by Approval Voting and Multiwinner Approval Voting, both of which select committees maximizing aggregate approval support (Lackner et al., 2018). This objective is technically simple but collides with stronger proportionality requirements. The paper on completing priceable committees studies this tension under demanding axioms such as EJR+ and shows that affordable or priceable committees can be completed to recover strong welfare guarantees while preserving proportional structure (Brill et al., 2023).

A central result is that committee completion can reconcile strong proportionality with asymptotically optimal utilitarian approximation. Completing affordable and EJR+ committees with Approval Voting preserves an $\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$1 utilitarian ratio, and the Greedy Justified Candidate Rule with AV-completion achieves the best possible utilitarian guarantee for proportional rules, namely

$\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$2

The Method of Equal Shares with AV-completion achieves

$\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$3

The same work shows that EJR alone is insufficient: for any $\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$4, there exist exhaustive EJR-committees with utilitarian ratio only $\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$5 (Brill et al., 2023).

The broader approximation landscape was quantified earlier. Approval Voting achieves utilitarian guarantee $\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$6, whereas Chamberlin–Courant, sequential CC, and Monroe have utilitarian guarantee $\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$7, and Minimax Approval Voting can have utilitarian guarantee $\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$8 in the worst case. PAV, seq-PAV, SLAV, $\text{uw}(W) = \sum_{i\in N} |A_i \cap W|,$9-Geometric rules, and sequential Phragmén occupy intermediate positions, often with guarantees of order $A_i$0, thereby formalizing the tradeoff between representation and utilitarian efficiency (Lackner et al., 2018).

A plausible implication is that, in committee design, utilitarian objectives are rarely rejected outright; rather, they are embedded within mechanisms that control the welfare cost of proportionality, or conversely the representation cost of aggregate-welfare maximization.

3. Fairness, distributive justice, and bounded efficiency loss

Several of the surveyed works treat utilitarian objectives as one pole of a fairness–efficiency continuum. In kidney exchange, the utilitarian objective maximizes total matching utility: $A_i$1 with utility often decomposed as $A_i$2 across prioritized and non-prioritized patient groups. The hybrid-lexicographic rule introduces a tolerance parameter $A_i$3: within a fair region it prioritizes disadvantaged groups lexicographically, but outside that region it reverts to utilitarian aggregation, thereby guaranteeing a bound on the price of fairness (McElfresh et al., 2017).

The corresponding utility function is piecewise: $A_i$4 For two groups, the price of fairness is bounded by $A_i$5, with $A_i$6 the efficient utilitarian utility (McElfresh et al., 2017).

In algorithmic fairness, utilitarian and Rawlsian objectives are connected through the softmax-style continuum

$A_i$7

As $A_i$8, this approaches mean loss, i.e. the utilitarian objective; as $A_i$9, it approaches the maximum loss, i.e. a Rawlsian worst-off criterion. The paper studies this continuum as a relaxation of the veil of ignorance and reports that increasing model complexity can improve both average and worst-case loss simultaneously (Rigobon, 2023).

A different relation is established in social choice via reduction rather than interpolation. A utilitarian black-box solver can be transformed into a distribution over deterministic outcomes that is leximin in expectation, and if the utilitarian solver is only $i$0-approximate then the induced lottery is $i$1-approximate in the leximin sense as well (Hartman et al., 2024). This suggests that, computationally, utilitarian optimization can function as an algorithmic primitive for more demanding fairness notions rather than merely as their rival.

4. Algorithm configuration and optimization under uncertainty

In algorithm configuration, utilitarian objectives are motivated by the claim that minimizing expected runtime poorly represents end-user preferences and is statistically hard to estimate when runtime distributions have rare but very long tails. Utility-based configuration replaces mean-runtime optimization with bounded expected-utility maximization, yielding empirical concentration bounds and adaptive procedures with theoretical guarantees (Graham et al., 2023).

The basic empirical estimator is

$i$2

and Hoeffding-style confidence intervals bound $i$3 and then $i$4. The Utilitarian Procrastination procedure is anytime and adaptive; with probability at least $i$5, it returns an $i$6-optimal algorithm after sample complexity within logarithmic factors of information-theoretic lower bounds (Graham et al., 2023).

This framework was extended to infinite parameter spaces by COUP, which replaces finite-arm enumeration with optimistic exploration over continuous or uncountable configuration spaces. Because a true global optimum may be inaccessible in an infinite space, the target guarantee becomes $i$7-optimality, where

$i$8

COUP returns configurations satisfying $i$9 with high probability at the end of each phase (Graham et al., 2024).

Subsequent work emphasized practice as well as theory. Improvements to COUP include KL-divergence confidence bounds, LUCB sampling, adaptive addition of new configurations, and model-guided candidate generation, while preserving guarantees. A case study then analyzes robustness to uncertainty in the utility function itself, using first-order stochastic dominance and parameter sweeps over canonical utility families (Graham et al., 16 Oct 2025).

5. Multi-agent systems, reinforcement learning, and decentralized cooperation

In multi-agent optimization, utilitarian objectives often appear as welfare sums. In mixed-motive Markov games, utilitarian welfare is

$\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$0

or equivalently the problem of maximizing the sum of all agents’ expected returns under a joint policy (Xu et al., 9 Feb 2026). The cited work argues that this dominant approach can yield efficient but highly inequitable outcomes and contrasts it with proportional fairness,

$\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$1

which empirically produced higher group performance and much lower Gini index in the CleanUp social dilemma (Xu et al., 9 Feb 2026).

In distributed constraint reasoning, utilitarian objectives are used differently: not to aggregate social welfare across final allocations alone, but to combine solution reward and privacy loss incurred during the search process. A Utilitarian Distributed Constrained Problem is defined as

$\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$2

with utility given by reward for reaching agreement minus cumulative privacy costs from revealing domains, assignments, or constraints. Agents estimate expected utility of actions and may halt participation if projected privacy loss outweighs expected reward (Savaux et al., 2017).

The earlier UDCOP formulation makes the same point for distributed optimization: privacy costs are paid when proposals are revealed, not just in the final accepted solution, so they cannot be treated as an ordinary objective term over terminal outcomes. Instead, agents modify heuristics to compare expected next-state cost including privacy exposure, and privacy-aware variants of DSA and DBO substantially reduce privacy loss with little degradation in solution quality (Savaux et al., 2016).

These examples show that utilitarian objectives in decentralized systems are frequently process-sensitive. The objective is not always “maximize total reward at the end,” but may instead aggregate the utility consequences of interaction, revelation, and coordination throughout the search.

6. Ethical AI, moral reasoning, and empirical interpretations

A recurring theme in AI ethics is that utilitarian objectives are often embedded in technical systems even when not explicitly described as ethical commitments. One argument is that reinforcement learning operationalizes a form of act utilitarianism because it selects policies that maximize expected cumulative reward, making the reward function a morally loaded object rather than a neutral engineering artifact (Roff, 2020). In healthcare AI, utilitarian ethics is proposed as a theoretical framework aimed at the “greatest good for the greatest number,” but the implementation is layered: privacy at data access, beneficence, bias, justice and solidarity, and interpretability at the algorithmic level, accountability, reliability, and risks at the systems level, and regulatory oversight at the organizational level (Emdad et al., 2023).

In explainable AI for consumer decision-making, a utilitarian module is instantiated by multi-criteria aggregation over normalized attributes. Each option receives score

$\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$3

where attributes are min–max normalized within round, weighted, and signed according to whether larger values are desirable or undesirable. A meta-explainer then compares this score with Kantian rule violations and switches to a deontically clean option if the utility gap is within a regret bound of $\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$4 (Atf et al., 4 Oct 2025).

Empirical work on moral judgment complicates any simple inference from “utilitarian answer” to “utilitarian disposition.” A replication study of LLM trolley and footbridge judgments found that apparent non-utilitarian behavior could arise from safety refusals triggered by advisory prompt framing; with neutral framing, GPT-4o produced 99% utilitarian responses on the trolley problem, while the footbridge result remained more heterogeneous across models and prompts (Himmelreich, 24 Mar 2026). A separate study of multi-agent LLM deliberation found a “utilitarian boost” in groups: all tested models were more willing in group settings to endorse norm-violating, utility-maximizing actions in personal dilemmas, though the mechanism differed from human group reasoning (Keshmirian et al., 1 Jul 2025).

A plausible implication is that utilitarian outputs in AI systems may reflect a mixture of objective design, refusal policy, prompting, and group protocol rather than a stable underlying ethical theory.

7. Critiques, tradeoffs, and domain-specific reinterpretations

The strongest recurring criticism of utilitarian objectives in the surveyed literature is distributional blindness. In social choice, mean-based welfare can miss harms concentrated on minorities, motivating quantile-based distributional welfare measures derived from utilitarian social choice functions rather than social welfare functions. The proposed framework characterizes utilitarian social choice as a convex combination of individuals’ stochastic choice rules,

$\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$5

and uses this to justify welfare evaluation by quantiles such as median compensating variation, not just averages (Echenique et al., 2024).

In applied network games, decentralized venue-selection and coloring problems use a utilitarian-looking welfare

$\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$6

with $\mathrm{sc}_{AV}(A, W) = \sum_{c \in W} |N(c)|,$7 equal to the agent’s color preference if no clash occurs and zero otherwise. Greedy policies converge on a restricted class of identical-preference games, while a Metropolis–Hastings-based policy is proven optimal for general preferences and can be extended with an expected-loss term to optimize robust welfare (Chen, 2022). In retrosynthesis evaluation, the term “utilitarian” is reinterpreted operationally: URSA defines utilitarian retrosynthesis assessment not merely by route termination in commercially available building blocks, but by chemical plausibility at the Solv-2 level, as captured by ChemCensor-based route scoring (Zagribelnyy et al., 6 Jul 2026).

Across domains, then, the utilitarian objective is not a single formula but a design pattern: specify a utility representation, aggregate it across affected entities or scenarios, and optimize the aggregate. The scientific disputes concern what counts as utility, how aggregation interacts with uncertainty and fairness, and whether auxiliary constraints, completions, or alternative welfare functionals are required to prevent efficient yet normatively unacceptable outcomes.