Utility-based Selection

• Utility-based Selection is a framework that defines decision-making by maximizing a task-specific utility function rather than relying on generic metrics.
• It is applied across diverse domains—algorithm configuration, road-network jobs, multimodal evidence selection, and training-data curation—tailoring the selected output to the task at hand.
• The approach encompasses scalar and multi-utility measures, Pareto front analysis, and heuristic search methods that balance performance, fairness, and robustness.

Utility-based selection denotes a class of decision and optimization problems in which the selected object is the one that maximizes a task-specific utility functional, rather than a generic proxy such as expected runtime, topical relevance, predictive accuracy, or uncertainty alone. In recent arXiv work, this formulation appears in utilitarian algorithm configuration, personalized recommendation from pairwise comparisons, retrieval-augmented generation, road-network job selection, and set-valued choice modeling. Across these settings, utility encodes user preferences over runtimes, menus, evidence usefulness, travel-feasible reward, or latent multi-criteria tradeoffs, and the selected output may be a single arm, a subset, a trajectory, or a Pareto frontier rather than a single top-ranked item (Graham et al., 16 Oct 2025, Boroomand et al., 12 Aug 2025, Zhang et al., 25 Jul 2025, Singhal et al., 2022, Pfannschmidt et al., 2020).

1. Formal problem classes

A recurring formulation casts utility-based selection as direct maximization of an expected or constrained utility objective. In utilitarian algorithm configuration, if $t$ is runtime and $u(t)\in[0,1]$ is a user-specified utility function, the goal is to choose a configuration with maximum expected utility,

$\max_i \ \mathbb{E}[u(t_i)].$

That framework also uses the capped mean utility

$U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$

and the completion probability

$F_i=\Pr_{t_i}(t_i\le\kappa),$

so that selection can be driven by confidence bounds over utility rather than by expected runtime alone (Graham et al., 16 Oct 2025).

In constrained routing and scheduling settings, the same idea appears as subset selection under feasibility constraints. The road-network job-selection problem chooses a subset $\mathcal I\subseteq\mathcal J$ maximizing earned utility

$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$

subject to temporal feasibility between consecutive jobs and a travel-budget constraint

$\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$

Here utility is not a score attached to isolated items; it is realized only by a feasible sequence (Singhal et al., 2022).

In multimodal evidence selection, utility is defined information-theoretically. For a query $q$ and candidate evidence $c$, visual evidence utility is

$u(t)\in[0,1]$0

that is, the KL divergence between the model’s output distribution with and without the evidence. Because direct answer-space optimization is expensive, the same work introduces a latent binary helpfulness variable $u(t)\in[0,1]$1 and ranks candidates by $u(t)\in[0,1]$2, with theoretical conditions under which the latent ranking preserves the answer-space ranking (Luo et al., 13 May 2026).

A related optimization-grounded formulation appears in multimodal data curation. In One-Step-Train, a sample $u(t)\in[0,1]$3 is scored by the validation-loss reduction caused by a simulated one-step update,

$u(t)\in[0,1]$4

with the first-order approximation

$u(t)\in[0,1]$5

This turns sample selection into ranking by marginal optimization utility (Jing et al., 8 May 2026).

Setting	Utility formulation	Selected object
Algorithm configuration	$u(t)\in[0,1]$6	Configuration
Road-network jobs	$u(t)\in[0,1]$7 under budget and temporal constraints	Feasible job subset
Visual evidence selection	$u(t)\in[0,1]$8	Evidence item or top-$u(t)\in[0,1]$9 set
One-Step-Train	$\max_i \ \mathbb{E}[u(t_i)].$0	Training sample subset

This diversity of formulations suggests that “utility” is best understood as a task-aligned objective functional, not as a fixed mathematical template.

2. Scalar utility, multi-utility representations, and set-valued choice

A central distinction in the literature concerns whether a scalar utility is sufficient. For top-1 decisions, scalar utilities are natural. For subset choice, they are often restrictive. The Pareto-embedding approach explicitly argues that a scalar latent utility $\max_i \ \mathbb{E}[u(t_i)].$1 induces a total order and therefore fits discrete top-1 choice, whereas subset choice is better modeled by embedding objects into a higher-dimensional utility space $\max_i \ \mathbb{E}[u(t_i)].$2 and defining the selected subset as the Pareto-optimal set $\max_i \ \mathbb{E}[u(t_i)].$3 of an input task $\max_i \ \mathbb{E}[u(t_i)].$4. The learned choice set is then the non-dominated frontier in utility space rather than a thresholded scalar ranking (Pfannschmidt et al., 2020).

The Gaussian-process choice-function model develops the same idea in probabilistic form. It assumes a latent vector-valued utility

$\max_i \ \mathbb{E}[u(t_i)].$5

and interprets observed choice sets as Pareto-undominated subsets. Each latent utility receives an independent GP prior,

$\max_i \ \mathbb{E}[u(t_i)].$6

and the model uses variational inference together with PSIS-LOO to learn both the latent utility values and the number $\max_i \ \mathbb{E}[u(t_i)].$7 of utility dimensions (Benavoli et al., 2023).

Personalized recommendation adds a different form of non-scalar utility. Recommendations are menus $\max_i \ \mathbb{E}[u(t_i)].$8 of size $\max_i \ \mathbb{E}[u(t_i)].$9, scored by a task-specific utility $U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$0 under a latent user ranking $U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$1. The recommender chooses

$U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$2

The objective is therefore utility of the final menu, not prediction of ratings or pairwise preferences as ends in themselves (Boroomand et al., 12 Aug 2025).

Biomechanical movement selection offers another explicitly multi-goal formulation. There the selected movement is the maximizer of

$U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$3

and a key special case adds metabolic energy as a negative term,

$U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$4

Utility thus aggregates heterogeneous goals such as minimizing metabolic energy while also respecting speed and step-length preferences (Hagler, 2016).

These models show that utility-based selection need not imply a single hidden score. In several domains, the selected object is defined by non-domination, menu value, or weighted multi-goal optimization.

3. Selection mechanisms and search procedures

Utility-based selection has produced a broad algorithmic repertoire. In utilitarian algorithm configuration, COUP maintains for each configuration an upper confidence bound $U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$5, a lower confidence bound $U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$6, empirical capped utility statistics, empirical completion rates, and a captime $U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$7. It samples configurations optimistically by the highest $U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$8, uses the largest $U_i=\mathbb{E}_{t_i}[u(\min(t_i,\kappa))]$9 as the current best-proven configuration, and refines the guarantee until

$F_i=\Pr_{t_i}(t_i\le\kappa),$0

is sufficiently small. The improved version replaces Hoeffding-style bounds with KL-based bounds, switches from vanilla UCB to LUCB, adds an adaptive rule for introducing new configurations, and uses an XGBoost-guided search for promising configurations while preserving the $F_i=\Pr_{t_i}(t_i\le\kappa),$1 guarantee structure (Graham et al., 16 Oct 2025).

In pairwise preference elicitation, active selection can itself be utility-based. The recommendation framework selects the query pair $F_i=\Pr_{t_i}(t_i\le\kappa),$2 that maximizes the expected improvement in downstream recommendation utility,

$F_i=\Pr_{t_i}(t_i\le\kappa),$3

This explicitly contrasts with entropy- or variance-based query choice: the next observation is selected because it is expected to improve the final recommendation, not because it is intrinsically uncertain (Boroomand et al., 12 Aug 2025).

For large candidate spaces, utility can guide candidate generation before downstream optimization begins. Sequential query recommendation with countably many arms defines a utility of arm selection through a preference probability conditioned on the current arm and chooses a candidate set $F_i=\Pr_{t_i}(t_i\le\kappa),$4 of size at most $F_i=\Pr_{t_i}(t_i\le\kappa),$5 by maximizing

$F_i=\Pr_{t_i}(t_i\le\kappa),$6

Because the objective is submodular under the paper’s construction, greedy maximization yields the standard $F_i=\Pr_{t_i}(t_i\le\kappa),$7 approximation structure. This paper’s claim is not merely that utility should rank the final arm, but that it should prune the candidate set itself (Parambath et al., 2021).

In road networks, the same logic yields lightweight heuristics. The Best First Search approach chooses a feasible next job by maximizing utility minus travel cost,

$F_i=\Pr_{t_i}(t_i\le\kappa),$8

while the Nearest Neighbor Search approach uses the priority

$F_i=\Pr_{t_i}(t_i\le\kappa),$9

Both are heuristic rather than optimal, but both instantiate the same principle: local selection is utility-sensitive and feasibility-aware (Singhal et al., 2022).

In RAG, utility-based selection often replaces fixed top-$\mathcal I\subseteq\mathcal J$0 ranking with dynamic set construction. The passage-selection distillation work uses a front-to-back sliding window, carries selected passages forward in a preselected queue, and treats the output as an adaptively selected subset rather than a full permutation. This is a set-selection process driven by passage usefulness for answer generation rather than relevance ranking alone (Zhang et al., 25 Jul 2025).

4. Retrieval-augmented generation and evidence usefulness

In retrieval-augmented generation, utility-based selection is largely motivated by the mismatch between semantic relevance and downstream answer quality. The passage-selection work argues that the useful passages are the ones that help generate an accurate and complete answer, not necessarily the ones with the highest topical similarity. It distills utility judgments from Qwen3-32B into a 1.7B student, trains the student to imitate both pseudo-answer generation and utility judgments, and shows that utility-based selection is especially effective on complex questions such as HotpotQA. The same study reports that utility selection is more robust than fixed top-$\mathcal I\subseteq\mathcal J$1 ranking and uses about 30% less inference time than relevance ranking in its setup (Zhang et al., 25 Jul 2025).

The visual-evidence framework pushes the same shift into multimodal RAG. Instead of optimizing similarity between image and query, it optimizes posterior change in the model’s output distribution. Operationally, it uses a binary auxiliary probe with target vocabulary $\mathcal I\subseteq\mathcal J$2 and scores candidates by the positive logit,

$\mathcal I\subseteq\mathcal J$3

where $\mathcal I\subseteq\mathcal J$4 is a template combining the query, the candidate image, and the auxiliary question. The top-$\mathcal I\subseteq\mathcal J$5 images by $\mathcal I\subseteq\mathcal J$6 are then passed to the main model. The paper reports that this latent-helpfulness method consistently outperforms strong baselines on MRAG-Bench and Visual-RAG, and that its decoding FLOPs are over $\mathcal I\subseteq\mathcal J$7 lower than answer-level uncertainty estimation (Luo et al., 13 May 2026).

A common misconception is that utility-based retrieval is simply another name for reranking. The cited RAG work does not support that interpretation. In both text and visual settings, the key distinction is that utility is defined with respect to generation behavior: evidence is selected because it changes the model’s answer distribution or improves the quality of the final answer, not because it is merely similar to the query.

5. Training-data selection, unlearning, and utility-preserving curation

A major recent use of utility-based selection concerns training data. In balanced unlearning, UPCORE treats forget-set selection as a trade-off between deletion efficacy and collateral damage, with the objective

$\mathcal I\subseteq\mathcal J$8

Its central empirical observation is that hidden state variance on the forget set correlates with retained model utility, with a reported Pearson correlation of $\mathcal I\subseteq\mathcal J$9. The method therefore prunes high-anomaly forget points in hidden-state space using Isolation Forest and keeps

$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$0

This is utility-based selection in the sense of preserving useful model behavior while still achieving forgetting (Patil et al., 20 Feb 2025).

Other work defines sample utility directly through optimization geometry. Grad-Mimic introduces the Mimic Score,

$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$1

which measures whether a sample’s negative gradient points toward a better reference model in weight space. It uses these scores both for reweighted training,

$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$2

and for post-hoc filtering, thereby making utility selection model-based rather than heuristic (Huang et al., 12 Jan 2025).

One-Step-Train ranks synthetic multimodal samples by immediate validation utility under a simulated one-step update and then fine-tunes only on the top-$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$3 subset. The paper reports that the Top-50% subset reduces training costs by 43% and total time consumption by 17 while surpassing the LLM-as-a-Judge baseline by 1.8 points, and that under a fixed compute budget the Top-20% subset achieves a 5.6 point gain over LLM-as-a-Judge and an 8.8 point gain over Full-SFT (Jing et al., 8 May 2026).

A different line argues that utility alone is incomplete without diversity. UDS defines

$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$4

where $$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$5 is the nuclear norm of the logits matrix,

$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$6

and $$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$7 is the average Euclidean distance from a memory buffer of past selected samples. The paper’s position is explicit: utility-only scoring may favor hard but redundant samples, so online batch selection should blend optimization value and diversity (Zou et al., 19 Oct 2025).

The market-based selector develops a multi-signal aggregation view of example utility. Signals such as uncertainty, rarity, and diversity act as traders in an LMSR market with cost

$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$8

Under a token budget, examples are selected by the price-per-token rule

$$\mathcal{B}(\mathcal{I})=\sum_{j\in\mathcal{I}}\mathcal{U}(j)$$9

This yields a maximum-entropy aggregation of heterogeneous utility signals with an explicit length-bias parameter $\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$0 (Jha et al., 2 Oct 2025).

Mathematical domain adaptation uses an even simpler explicit utility score,

$\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$1

combined with an embedding-based diversity term

$\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$2

and optimized under the budgeted objective

$\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$3

Here again, utility-based selection is not synonymous with “pick the hardest examples”; representativeness is part of the formal objective (Kotecha et al., 2 May 2025).

6. Robustness, fairness, and theoretical boundaries

Recent work also emphasizes that utility-based selection is only as reliable as its utility specification and its observed utility estimates. In utilitarian algorithm configuration, a SAT Competition case study shows that rankings are relatively stable as the PAR penalty factor $\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$4 varies, but can change substantially as the timeout $\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$5 varies. That paper therefore advocates robustness analysis over families of utility functions rather than committing to a single precisely known utility (Graham et al., 16 Oct 2025).

When observed utilities are biased across groups, direct utility maximization can produce unfair allocations. In centralized selection with preferences, disadvantaged-group utilities are modeled as $\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$6 for $\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$7, and the paper proves that the unconstrained stable assignment can yield poor representational and preference-based fairness. It proposes institution-wise proportional constraints and proves that, under its assumptions, the resulting algorithm achieves $\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$8, near-optimal preference fairness, and near-maximal true utility. This work shows that utility-based selection may require distributional or fairness constraints to recover the intended welfare criterion when utility observations are biased (Celis et al., 2024).

Portfolio selection pushes the issue into well-posedness. In one-period utility-risk optimization, the paper shows that market-independent well-posedness holds if and only if either the utility functional or the risk functional is sensitive to large losses. In the special case of expected utility maximization without a risk constraint, the exact criterion is the asymptotic loss-gain ratio

$\sum_{i=1}^{k-1}\mathbb{C}(v_i,v_{i+1})\le B.$9

with universal well-posedness characterized by $q$0. This result sharply separates utility-based selection that is merely formally specified from utility-based selection that is mathematically stable (Baggiani et al., 12 Sep 2025).

A broader implication of these results is that utility-based selection is not a single doctrine. Some works treat utility as the decisive objective; others treat it as one component that must be balanced with diversity, risk, fairness, or robustness. A plausible implication is that the most durable formulations are those that make these trade-offs explicit, either in the utility itself or in the constraints surrounding selection.