
--- PAGE 1 ---
[proposition]
1

--- PAGE 2 ---
Salience as Contextual Budgeted Set Selection Under Uncertain
Utility
Abstract
Many practical systems must choose a finite-cost subset of candidates when value depends
on context, constraints, and interactions among the selected items. This paper formulates that
problem as contextual budgeted set selection under uncertain utility. The true objective is
latent expected downstream value, while the operational system optimizes a surrogate set utility
that combines itemwise terms for novelty, goal-fit, long-run leverage, coherence, and temporal
relevance with penalties for drag and redundancy and bonuses for complementarity. Hard safety,
policy, validity, and compatibility constraints are represented as feasibility conditions rather than
offsetting penalties. This yields a marginal-gain view of selection and clarifies the distinct roles
of candidate generation, reranking, constrained optimization, control, and learning. Simpler
salience-style formulas appear as restricted special cases. Stronger optimality claims require
additional structure, such as modularity or submodularity, and should not be stated without
those assumptions.
1 Introduction
Finite-capacity selection recurs across retrieval, recommendation, memory management, routing,
and attention support. A system faces more candidates than it can carry forward, each candidate
consumes some resource, and the payoff of selecting a set depends both on context and on what
else is selected. Many salience-style formulations compress this into a single scalar item score. That
compression is often useful as a mnemonic, but it is too coarse as a general theory.
The cleaner view is to treat salience as a constrained set-selection problem. A chooser operates under
finite capacity, hard feasibility constraints, uncertain downstream utility, and interaction effects
among selected items. In that setting, the right formal object is not a universal item score, but a
contextual set objective together with a feasible family and a selection procedure.
This reformulation makes four clarifications. First, it separates latent downstream value from the
surrogate utility the system actually estimates and optimizes. Second, it places hard constraints
in feasibility rather than in soft penalties. Third, it treats redundancy and complementarity as set
properties rather than pretending they are intrinsic to isolated items. Fourth, it separates scoring
from optimization, control, and learning. Those separations are structural, not cosmetic.
2

--- PAGE 3 ---
2 Formal Setup
2.1 Context, candidates, and feasibility
Let x denote the exogenous context for a single decision episode. Context may include user state,
task, session history, device, time horizon, risk posture, and any other state that changes what is
valuable now.
Unless stated otherwise,x is held fixed while the current selection problem is solved. If selection
itself changes the relevant state, the model should be applied stagewise or lifted to a sequential
decision framework.
LetI(x)be the candidate universe available in contextx.
Each candidatei∈I(x)has:
•hard item-level eligibilityT i(x)∈ {0,1},
•costc i(x)>0when cost-normalized scoring is used.
Some hard constraints are set-level rather than item-level. Let
C(A, x)∈ {0,1}
denote any additional hard feasibility condition on a setA, such as mutual exclusion, compatibility,
quota, schema, or channel constraints.
Define the eligible universe
I +(x) ={i∈I(x) :T i(x) = 1}
and the feasible family
F(x) =
(
A⊆I +(x) :
X
i∈A
ci(x)≤B(x), C(A, x) = 1
)
.
This formulation keeps item-level gating and set-level gating distinct while allowing both to be
enforced before or during selection.
2.2 Latent downstream value
Let
V ∗(A, x) =E[Y(A, x)|x]
denote the latent expected downstream value of selectingA in context x, where the expectation is
over whatever uncertainty remains after conditioning on context.
The ideal chooser solves
A∗(x)∈arg max
A∈F(x)
V ∗(A, x).
3

--- PAGE 4 ---
In practice V ∗ is not directly known. The system therefore estimates and optimizes a surrogate
objective.
2.3 Remarks on zero-cost and sequential cases
If genuinely zero-cost items exist, ratio scoring by1/ci(x)is not automatically meaningful. Those
items should either be handled by raw marginal gain, by a separate convention, or by a richer cost
model.
If adding an item changes the context materially, write statesx0, x1, . . . and solve a sequence
of stagewise selection problems or a fuller sequential decision problem. The present formulation
addresses a single decision episode with context treated as fixed during selection.
3 Feature Vocabulary and Surrogate Utility
3.1 Stable feature vocabulary
For each candidatei, define:
•N i(x): novelty or information gain,
•G i(x): goal-fit or immediate relevance to the current task,
•L i(x): long-run leverage, such as reuse value, retention value, compounding value, or future
usefulness,
•H i(x): coherence with the current trajectory, working set, or dialogue state,
•τ i(x): age, staleness, or lag,
•D i(x): drag, meaning friction, interruption cost, cognitive burden, latency burden, or soft risk
burden.
For pairs of candidatesi, j, define:
•K ij(x): complementarity, synergy, or unlock value,
•Sim ij(x): redundancy, overlap, or substitutability.
Pairwise terms are taken to be symmetric unless otherwise stated. They are an approximation
chosen for tractability, not a claim that all real interaction effects are pairwise.
4

--- PAGE 5 ---
3.2 Base item utility
Letθcollect the parameters of the surrogate model. Define the base item utility
ui(x;θ) =w N(x)ϕ N(Ni(x);x) +w G(x)G i(x) +w L(x)L i(x)
+w H(x)H i(x) +w F (x)ϕ F (τi(x);x).
Here w∗(x)are contextual weights, ϕN is a learned or chosen novelty response, andϕF maps age or
lag to temporal relevance.
This is the value of candidatei before interactions and soft burdens are applied. The weights should
depend on context. Research mode, emergency mode, cold start, and long-session expert use should
not all share the same tradeoff profile.
3.3 Set utility
Define the surrogate set utility
ˆU(A, x;θ) =
X
i∈A
ui(x;θ) +
X
i<j
i,j∈A
Kij(x;θ)−λ div(x)
X
i<j
i,j∈A
Simij(x;θ)−
X
i∈A
Di(x;θ).
This objective adds itemwise value, rewards complementarity, penalizes redundancy, and subtracts
soft burdens.
Two distinctions matter. First,ci(x)is hard resource use that appears in feasibility.Di(x)is soft
burden that appears in utility. They should not be conflated. Second, ˆU is a set function for
unordered selection. If order matters, the right object is a sequence function rather than a set
function.
3.4 Operational problem
The deployed system solves
ˆA(x)∈arg max
A∈F(x)
ˆU(A, x;θ),
or an approximation to that problem when exact optimization is too expensive.
This makes the separation explicit:V ∗(A, x)is the latent objective of interest, ˆU(A, x; θ)is the
estimated operational surrogate,F(x)captures hard feasibility, and the selector is the algorithm
used to approximate the argmax.
4 Marginal Gain and Operational Priority
The operational quantity for incremental selection is marginal gain relative to the already selected
setA:
∆i(A, x;θ) = ˆU(A∪ {i}, x;θ)− ˆU(A, x;θ).
5

--- PAGE 6 ---
Under the pairwise surrogate above,
∆i(A, x;θ) =u i(x;θ) +
X
j∈A
Kij(x;θ)−λ div(x)
X
j∈A
Simij(x;θ)−D i(x;θ).
Define the feasible marginal-density score
si(A, x;θ) =1[A∪ {i} ∈ F(x)]· ∆i(A, x;θ)
ci(x) .
This rule means: infeasible additions are excluded; candidates are scored by what they add now; and
normalization by cost is used only when heterogeneous costs make such normalization meaningful.
If any feature depends on the evolving set rather than only on exogenous context, write that explicitly
(for example ui(A, x; θ)), and compute∆i directly rather than relying on the pairwise expansion
above.
5 Structural Propositions
The point of this section is to state exactly what follows from the formulation, and no more.
5.1 Basic set-function notions
A set functionf : 2V →R is modular iff(A) = const +P
i∈A ai. It is submodular if for allA⊆B
and alli /∈B,
f(A∪ {i})−f(A)≥f(B∪ {i})−f(B).
It is monotone iff(A) ≤f (B)whenever A⊆B . Submodularity is the diminishing-returns property.
Proposition 1(Itemwise ranking in the modular equal-cost regime).Suppose that for fixed context
x: (1) ˆU(A, x; θ) =P
i∈A vi(x)is modular, (2) all eligible items have equal cost, and (3) feasibility
is a cardinality constraint|A| ≤k. Then any set containing theklargest valuesv i(x)is optimal.
Proof sketch.Under equal costs and a cardinality bound, feasibility depends only on the number of
selected items. For any feasible set that omits a higher-valued item and includes a lower-valued one,
exchanging the lower-valued item for the higher-valued one weakly increases the objective. Repeating
the exchange yields a set of top-kvalues.
Proposition 2(Density ranking fails under nonuniform costs).Even when ˆU is modular, ranking
by value-per-cost is not generally optimal for0–1knapsack selection with nonuniform costs.
Counterexample.LetB= 50and three eligible items have(value,cost)pairs
a= (60,10), b= (100,20), c= (120,30).
Ratios are6 , 5, 4, so density-first picksa then b, yielding value160. But {b, c} is feasible with value
220>160.
6

--- PAGE 7 ---
Proposition 3(Redundancy-only penalties imply submodularity).Suppose Kij(x; θ) = 0for all
pairs andSim ij(x;θ)≥0. Define
bi(x;θ) =u i(x;θ)−D i(x;θ)
and
f(A) =
X
i∈A
bi(x;θ)−λ div(x)
X
i<j
i,j∈A
Simij(x;θ).
Thenfis submodular.
Proof sketch.ForA⊆Bandp /∈B:
f(A∪ {p})−f(A) =b p −λ div
X
j∈A
Simpj,
f(B∪ {p})−f(B) =b p −λ div
X
j∈B
Simpj.
Since A⊆B and similarities are nonnegative, the second subtraction is at least as large, so
diminishing returns holds.
Corollary 3.1(Monotonicity requires an additional condition).Under the assumptions of Proposi-
tion 3, the surrogate is monotone if and only if all feasible marginal gains are nonnegative.
A convenient sufficient condition is
bi(x;θ)≥λ div(x)
X
j∈A
Simij(x;θ)
for every feasible addition ofitoA.
Proposition 4(Positive complementarity can break submodularity).Suppose there exist distinct
items a, b such that Kab(x; θ) > 0. Then the pairwise complementarity term is supermodular, and
the full surrogate may fail to be submodular.
Proof sketch.Consider
f(A) =
X
i∈A
bi +K ab 1[{a, b} ⊆A].
With A = ∅, B = {b}, adding a gives marginal gainsba versus ba + Kab. Since Kab > 0, marginal
gain increases after selectingb, violating diminishing returns.
Proposition 5(Finite penalties cannot uniformly encode hard exclusion).Suppose ineligibility is
represented by subtracting a finite penaltyM >0instead of using a hard gate:
˜U(A, x) =U(A, x)−M
X
i∈A
(1−T i(x)).
Then for any fixedM, there exist instances in which an ineligible item is still selected.
Proof sketch.Pick an ineligible itemq with Tq(x) = 0that fits within budget. If∆ q(∅, x) > M,
then penalized marginal gain remains positive:∆q(∅, x)−M >0.
7

--- PAGE 8 ---
6 Algorithmic Implications
The propositions above pin down what can and cannot be claimed.
In the modular, equal-cost, cardinality-constrained regime, top-k ranking is exact. That is the
cleanest setting in which a single itemwise score really does solve the problem.
With modular utility and nonuniform costs, the problem becomes0–1knapsack. Even then, a
density rule is only a heuristic, not a general optimum rule.
With redundancy-only pairwise penalties and nonnegative similarities, the surrogate becomes
submodular. In that regime, diminishing returns appears naturally, and submodular-style algorithms
become mathematically natural.
Once positive complementarity enters, the objective can become neither modular nor submodular.
Then greedy procedures lose any blanket claim to optimality, and richer search or optimization
procedures become appropriate. Depending on scale and latency budget, those may include beam
search, local search, reranking loops, mixed-integer optimization, or learned approximate selectors.
The safe statement is therefore narrow:
Algorithmic guarantees depend on the structure of the surrogate objective and the feasible
family. They do not follow from a salience score alone.
7 Architectural Decomposition
A practical system should separate six layers.
7.1 Eligibility layer
Apply hard feasibility checks. Item-level disqualifiers go intoTi(x). Set-level disqualifiers go into
C(A, x).
Examples include safety rules, policy rules, schema validity, source validity thresholds, compatibility
checks, mutual exclusion, and quota rules.
7.2 Candidate generation
Recall a broad, cheap pool from the candidate universe. This stage optimizes coverage and speed,
not final precision.
8

--- PAGE 9 ---
7.3 Reranker
Estimate surrogate utility terms and interactions. This stage computes itemwise value signals, drag
estimates, redundancy estimates, and complementarity estimates.
7.4 Selector
Choose the final feasible set underF(x). This stage is where greedy construction, beam search, local
search, ILP, or other constrained optimization methods belong.
7.5 Controller
Adjust exploration, aggressiveness, budget tightening, temperature, channel multipliers, or other
policy knobs that govern behavior across contexts. Exploration belongs here, not inside a single
salience score.
7.6 Learner
Update θ, thresholds, transforms, interaction estimates, and controller policies from observed
outcomes. Learning belongs here, not inside the same scalar equation used for online ranking.
This decomposition is useful because it prevents hard constraints, soft utility, exploration, optimiza-
tion, and learning from being crammed into one overworked formula.
8 Learning and Evaluation
The separation betweenV ∗ and ˆUhas direct learning consequences.
Observed feedback typically arrives only for the selected set, not for all counterfactual feasible sets.
That means learning the surrogate is partly a counterfactual estimation problem. A system may
estimate component models for itemwise terms and pairwise terms, fit a direct predictor of set-level
outcomes, or combine both approaches.
A generic learning target is to estimate parametersθ so that higher surrogate utility corresponds to
better observed outcomes under deployment-relevant conditions. In symbolic form, one may fitθ by
minimizing a loss of the form X
t
ℓ
 
ψ(At, xt;θ), y t

+ Ω(θ).
where At is the selected set in episodet, yt is the observed outcome,ψ is the model output used for
training, andΩis any regularizer or structural penalty.
Two evaluation cautions follow.
9

--- PAGE 10 ---
First, when interactions matter, itemwise ranking metrics can become a misleading proxy for decision
quality. A system that ranks individual items well can still assemble poor sets.
Second, evaluation should track the level at which the system acts. If the selector chooses sets, then
at least part of the evaluation should be set-level or policy-level rather than purely item-level.
This is one more reason the latent-versus-surrogate split is not just notation cleanup. It affects the
learning target and the evaluation protocol.
9 Domain Instantiations
9.1 Retrieval and RAG
In retrieval settings:
•Gcaptures query relevance,
•Lcaptures downstream usefulness for reasoning or future reuse,
•Hcaptures coherence with the current dialogue or working set,
•τcaptures source age or lag,
•Simcaptures chunk overlap,
•Kcaptures multi-hop complementarity across evidence pieces,
•Dcaptures token cost, latency, or distraction burden,
•TandCcapture trust, source validity, safety, and policy constraints.
9.2 Training data selection
In training data selection:
•Ncaptures undercoverage or surprise,
•Gcaptures fit to the current training objective,
•Lcaptures long-run capability gain,
•Hcaptures curriculum coherence,
•τcaptures distribution lag where relevant,
•Simcaptures duplication or oversampling overlap,
•Kcaptures complementary samples that unlock generalization,
•Dcaptures labeling cost, noise burden, or instability risk,
•TandCcapture license, validity, and policy constraints.
10

--- PAGE 11 ---
9.3 Product and engagement systems
In product flows:
•Gcaptures immediate task completion value,
•Lcaptures retention or future utility,
•Hcaptures journey coherence,
•τcaptures timeliness,
•Simcaptures recommendation redundancy,
•Kcaptures next-step unlocks or useful bundles,
•Dcaptures friction and interruption burden,
•TandCcapture trust, safety, and experience constraints.
9.4 Human operator mode
For human reasoning, a compact salience heuristic can still be useful. But it should be treated as an
operator shorthand over this substrate, not as the substrate itself.
10 Scope of Claims
The model supports strong structural claims and weaker universal claims. Keeping those separate
prevents puffed-up nonsense.
Safe claims:
• Many practical selection systems can be modeled as contextual feasible-set selection under
uncertain utility.
•Itemwise salience formulas are restricted special cases of that broader problem.
•Hard exclusions are more robustly modeled as feasibility conditions than as finite penalties.
•Redundancy can produce diminishing returns, while complementarity can destroy them.
Unsafe claims unless extra assumptions are stated:
•One scalar item score explains all selection behavior.
•Greedy selection is optimal in general.
•Value-per-cost ranking solves the nonuniform-cost case.
•Hard constraints can always be replaced by sufficiently large soft penalties.
11

--- PAGE 12 ---
11 Conclusion
The useful general principle is not that one salience equation explains everything. The useful
principle is that finite-capacity selection is a contextual feasible-set problem under uncertain utility.
Candidates have costs. Some candidates are ineligible. Value depends on context. Redundancy
and complementarity are properties of sets, not just of isolated items. The deployed chooser acts
on an estimated surrogate through feasible marginal gains, often normalized by cost when that
normalization is meaningful.
That framing preserves what was good in the original salience idea while removing notation drift,
layer confusion, and fake universality.
A Compact Canonical Form
Given contextx, define
F(x) =
(
A⊆I +(x) :
X
i∈A
ci(x)≤B(x), C(A, x) = 1
)
.
Latent downstream value:
V ∗(A, x) =E[Y(A, x)|x].
Surrogate utility:
ˆU(A, x;θ) =
X
i∈A
ui(x;θ) +
X
i<j
i,j∈A
Kij(x;θ)−λ div(x)
X
i<j
i,j∈A
Simij(x;θ)−
X
i∈A
Di(x;θ).
Ideal and operational problems:
A∗(x)∈arg max
A∈F(x)
V ∗(A, x), ˆA(x)∈arg max
A∈F(x)
ˆU(A, x;θ).
Incremental priority:
∆i(A, x;θ) = ˆU(A∪ {i}, x;θ)− ˆU(A, x;θ),
si(A, x;θ) =1[A∪ {i} ∈ F(x)]· ∆i(A, x;θ)
ci(x) .
B Legacy Mapping ofS ′, UEA, and SRC
Because exact legacy formulas are not reproduced in the visible thread, this mapping is structural
rather than symbol-for-symbol.
12

--- PAGE 13 ---
B.1S ′
If S′ was an itemwise scalar for ranking, it maps either toui(x; θ) −D i(x; θ)(raw item value) or to
si(A, x;θ)(marginal-density priority). Under restrictive assumptions,
S′
i(x)≈ ui(x;θ)−D i(x;θ)
ci(x)
whenA=∅,K ij = 0,Sim ij = 0,T i(x) = 1, and set-level feasibility is trivial.
B.2 UEA
If UEA denoted true downstream performance, it maps toV ∗(A, x). If it denoted operational scoring,
it maps to ˆU(A, x;θ)or one of its components. Those two roles should not share one symbol.
B.3 SRC
If SRC governed exploration, aggressiveness, routing, or channel choice, it belongs to the controller
layer. If it performed constrained feasible-set choice, it belongs to the selector layer. It should not
be folded as another additive utility term.
B.4 Trust/safety modifiers
Binary disqualifiers should move intoTi(x)or C(A, x). Graded source-quality effects can remain
soft terms but should be represented explicitly as such.
B.5 Freshness/novelty multipliers
Fixed multipliers map to contextual weightswN(x), wF (x)or response functions ϕN(·; x), ϕF (·; x),
separating importance weighting from response shape.
C Drop-in Claim Language
1. Structural framing.Many retrieval, recommendation, routing, and memory systems can be
modeled as contextual feasible-set selection under uncertain utility.
2. Operational framing.A useful strategy is to estimate feasible marginal utility, often normalized
by cost, using a surrogate with drag, redundancy, and complementarity.
3. Special-case framing.Itemwise salience scores are exact only in restricted regimes where the
objective is effectively modular and the feasible family is simple enough to collapse set selection
into ranking.
13

--- PAGE 14 ---
4. Constraint framing.Hard policy, safety, and validity constraints are better represented as
feasibility conditions than as tradeable penalties.
5. Guarantee framing.Approximation or optimality claims require explicit assumptions such as
modularity, monotonicity, or submodularity.
14
