Preference

Prime #

None

Origin domain

Decision Theory Economics

Also from

Psychology & Behavioral Sciences, Philosophy, Political Science

Aliases

Ordering, Ordered Valuation

Core Idea

Preference is an agent's ordering over a choice set on some evaluative dimension — a disposition that, when consulted, says which alternatives are favored, disfavored, indifferent, or incomparable, the framing Mas-Colell, Whinston, and Green (1995) take as the canonical primitive of modern microeconomic theory. ^[1] The four roles travel together everywhere the prime is invoked: a choice set (the alternatives being ordered), an evaluator (the agent, criterion, system, or model whose preference this is), an ordering relation over the set (which may be complete or partial, transitive or inconsistent, strict or weak), and a context in which the ordering does work — guiding selection, predicting behavior, or supplying the objective that optimization will then maximize, the decomposition Mas-Colell, Whinston, and Green (1995) develop in their chapter contrasting the preference-based and choice-based approaches. ^[1]

The abstraction is substrate-neutral. Utility functions, rankings, revealed choices, policy priorities, qualitative value orderings, and learned reward signals are all implementations of the same ordering relation; the relation is the prime, the implementation is local technology — a substrate range that runs from Debreu's (1954) representation theorem for continuous preference orderings to Christiano et al.'s (2017) deep-RL reward models fit from pairwise human comparisons. ^[2][3] What preference does not include is also load-bearing for the prime. It is not the act of selecting (that belongs to decision); it is not the condition that comparisons share a common metric (value_commensuration); it is not the search for a best feasible option under an objective (optimization, which presupposes preference); it is not the case of multiple evaluators disagreeing (preference_heterogeneity_and_conflict, a child case). The prime names the bare commitment of comparability under an ordering from an evaluator's standpoint — nothing more, but nothing less either, as Sen (1970) insists in separating the ordering relation from the choice procedure and from interpersonal aggregation. ^[4]

How would you explain it like I'm…

Liking One More

If someone offers you chocolate or vanilla ice cream, you probably want one more than the other — that wanting-more is your preference. It just means you have a way of ranking the choices: this one first, that one second, maybe these two are tied.

Ranking Your Choices

A preference is how someone (or something) ranks their choices. If you'd rather have pizza than salad, and salad than cereal, that's a preference ordering — pizza is first, salad second, cereal last. Preferences don't have to come with numbers; they just say what beats what. Almost every decision rule, from picking lunch to a computer choosing a move in a game, starts from some kind of preference over the options.

Ordering over alternatives

A preference is an agent's ordering over a choice set on some evaluative dimension — a disposition that, when consulted, says which options are favored, disfavored, equally good, or unable to be compared. Four things travel with the prime: a choice set (what is being ordered), an evaluator (whose preference it is — a person, a criterion, an algorithm), an ordering relation over the set (which can be complete or partial, consistent or messy), and a context where the ordering does work (guiding selection, predicting behavior, supplying an objective for optimization). Preferences can be expressed as utility numbers, ranked lists, observed choices, or learned reward signals — these are all implementations of the same abstract ordering.

 

Preference is an agent's ordering over a choice set on some evaluative dimension — a disposition that, when consulted, says which alternatives are favored, disfavored, indifferent, or incomparable. Four roles travel with the prime: a choice set (the alternatives being ordered), an evaluator (the agent, criterion, system, or model whose preference this is), an ordering relation over the set (which may be complete or partial, transitive or inconsistent, strict or weak), and a context in which the ordering does work — guiding selection, predicting behavior, or supplying the objective that optimization will then maximize. The abstraction is substrate-neutral: utility functions, rankings, revealed choices (preferences inferred from observed decisions), policy priorities, qualitative value orderings, and learned reward signals are all implementations of the same ordering relation. Preference is not the act of selecting (that is decision), not the assumption that comparisons share a common metric (that is value commensuration), not the search for a best feasible option (that is optimization, which presupposes preference), and not the case of multiple evaluators disagreeing (that is preference heterogeneity and conflict, a child case).

Structural Signature

Preference encodes a structural pattern: choice set → evaluator's disposition → ordering relation → contextual consumption. It separates two regimes (an undifferentiated set of alternatives and a set carrying an evaluator-relative ordering) and names the relation between them. The relation is the minimum addition needed to turn "here are some options" into "here are some options under an ordering," the minimalist axiomatic core Mas-Colell, Whinston, and Green (1995) develop before any utility representation is introduced. ^[1]

Recurring features:

• Ordering relation over a choice set from an evaluator's standpoint
• Disposition that picks out favored alternatives when consulted
• Comparative valuation independent of action commitment
• Substrate-neutral ranking, revealed or stated, complete or partial
• Objective supplied to downstream selection or optimization
• Evaluator-relative rather than view-from-nowhere

The signature applies whether the evaluator is a consumer ranking grocery bundles, a voter ordering candidates, a critic ranking performances (the taste-and-distinction apparatus Bourdieu (1984) analyzes for cultural judgment), a foraging bee preferring flower types, an RLHF labeler picking better completions (the setup Christiano et al. (2017) formalized for deep reinforcement learning), or a recommender system inferring an ordering from clickstream traces. What varies is which slot fills with what; what is conserved is the four-role relation. ^[5][3]

What It Is Not

Preference is not the same as choice. A choice is a discrete act of selecting one option from a set at a moment in time; a preference is the standing ordering that may inform many choices, no choices, or choices that contradict the ordering (under akrasia, constraint, or noise). I can prefer X to Y for years without ever selecting between them; I can choose Y over X once and still prefer X. Conflating preference with choice is the standard error of behaviorist-leaning theories, which assume the act exhaustively reveals the disposition. It does not — choice is a sample from a preference disposition mixed with constraint, noise, framing, and strategic considerations.

Preference is also not utility. Utility is one particular quantification of a preference ordering — a real-valued function that represents the ordering when certain axioms hold (completeness, transitivity, continuity, independence, in the classical von Neumann-Morgenstern framing). Many preferences cannot be represented as utility functions (intransitive, incomplete, or lexicographic orderings resist real-valued representation), and many ordering-equivalent utility functions encode the same preference (monotone transformations preserve the ordering). The prime is the ordering; utility is one mathematically convenient representation that imports stronger assumptions than the ordering itself requires.

Preference is not a value either, in the broader normative sense. A value is a standing commitment to what should matter (justice, beauty, autonomy, efficiency); a preference is an ordering over alternatives that may or may not reflect values, that may include strictly hedonic or purely habitual orderings with no normative content, and that may exist in tension with declared values. A person can value health while preferring a cheeseburger; the value sets standards, the preference orders alternatives.

Preference is not a bias. A bias is a systematic deviation from a posited rational benchmark, typically defined relative to a normative theory of preference (expected utility, Bayesian updating). The benchmark itself is preference-theoretic; biases are diagnosed against that benchmark. Calling a preference "biased" presupposes a normative ordering against which the empirical ordering deviates — so bias is not an alternative to preference but a diagnostic mode within preference theory.

Finally, preference is not optimization. Optimization searches for the best feasible option under an objective; preference is the ordering that supplies that objective. Optimization presupposes preference; preference is the more primitive notion. A system can have preferences without optimizing (it may simply consult the ordering when asked, never running a search), and an optimizer is empty without a preference relation specifying what counts as better — the asymmetry Mas-Colell, Whinston, and Green (1995) build into the order of presentation, defining preferences before any utility-maximization problem is posed. ^[1]

Broad Use

Economics & decision theory: Utility functions, ordinal and cardinal preference orderings, indifference curves, revealed preference theory, expected utility under risk, subjective expected utility under uncertainty, time preference and intertemporal discounting. Preference is the foundational primitive on top of which the entire formal apparatus of consumer theory, choice theory, and welfare economics is built — from von Neumann and Morgenstern's (1944) expected-utility axiomatization to Samuelson's (1938) revealed-preference reconstruction of consumer demand without recourse to utility. ^[6][7]

Behavioral & cognitive psychology: Choice under risk and uncertainty (prospect theory, loss aversion, reference-dependence), framing effects, preference reversals, time-inconsistent preferences (hyperbolic discounting), endowment effects, status-quo bias. Each of these is a documented deviation between empirical orderings and the rational-choice benchmark, and each presupposes the prime structurally even while challenging its axiomatic forms.

Political theory & public choice: Voter preference orderings, Arrow's impossibility theorem on aggregating heterogeneous preferences into a social preference, Condorcet cycles, Sen's liberal paradox, social welfare functions, mechanism design that elicits or aggregates individual orderings into collective outcomes. The aggregation problem is non-trivial precisely because preference is evaluator-relative — there is no view-from-nowhere ordering.

Machine learning: Preference learning from pairwise comparisons (RLHF, DPO, ranking losses), reward modeling, contextual bandits, recommender systems learning user orderings from implicit signals (clicks, dwell time, skips), inverse reinforcement learning recovering reward functions from observed behavior. The "evaluator" is a learned scoring function rather than a conscious agent, but the four-role structure is intact.

Biology & behavioral ecology: Revealed preference in animal foraging (patch selection, prey choice), mate choice, habitat selection, optimal foraging theory's apparatus borrowed directly from microeconomics. The substrate-furthest case for the prime: a pollinator's revealed flower preference has no stated valuation, no introspective access, no utility function written down, and yet the ordering-from-an-evaluator's-standpoint pattern fits cleanly.

Aesthetics & criticism: Ranked judgments over artworks, performances, or styles from a critic's or an audience's standpoint; the formal apparatus of taste sociology (Bourdieu); preference orderings as cultural-capital signals. The ordering is evaluator-relative even when dressed in universalist rhetoric.

Clarity

A core function of "preference" is to separate the ordering from everything around the ordering. Many problems present as confusions about valuation that are really confusions about which slot is doing the work. "Why did she pick that one?" is a question about choice (the act); "Why does she rank that one above the other?" is a question about preference (the ordering); "Why is that one worth more?" is a question about value or utility (the magnitude or commitment); "Which is the best one available?" is a question about optimization (the search under the ordering). When the four are conflated, debates ricochet between domains without progress.

Naming the prime as the ordering relation, evaluator-relative, distinct from action and from quantification, lets the analyst locate the live question. Is the choice set ill-defined? Then we're upstream of preference — we need a clearer set of alternatives. Is the evaluator ambiguous? Then we have an attribution problem — whose preference is at stake? Is the ordering partial, intransitive, or framing-dependent? Then the disposition itself is structured differently than a single complete transitive ordering can capture, and the formal apparatus must be weakened or restructured. Is the context shifting between consultation and use? Then we have a stability problem — the same ordering may not be available when needed, the family of distinctions Sen (1970) systematizes when separating ordering, consistency, and aggregation as conceptually independent axes. ^[4]

The prime also clarifies why deviations are diagnostic rather than disqualifying. An intransitive cycle in revealed preference is not "wrong" — it is information about how the underlying disposition is structured (reference-dependent, multidimensional, mood-modulated). Treating deviations as data rather than as errors-to-be-corrected requires having the prime named so that "deviation from what?" has a precise referent.

Manages Complexity

Preference decomposes a valuation situation into four named roles that turn an opaque "what's better here?" into a structured problem. Once the choice set, evaluator, ordering, and context are visible, downstream phenomena fall out by topology rather than by reinvention in each domain.

Multiple evaluators with different orderings becomes preference heterogeneity — a structural sub-case where the evaluator slot is plural and orderings disagree. An ordering that flips with framing or with time becomes temporal inconsistency or preference reversal — the structural sub-case Strotz (1955) formalized as time-inconsistent dynamic utility maximization — where the ordering itself is unstable across consultation contexts. An ordering inferred from behavior rather than from reports becomes revealed preference — the Samuelson (1938) sub-case where the access mode to the ordering is indirect. An ordering aggregated across many evaluators raises the social-choice problem and Arrow's (1951) impossibility theorem — a structural sub-case where the aggregation operation itself faces constraints. ^[8][7][9]

The same role-vocabulary that handles a consumer choosing between bundles handles a voter ranking candidates, an RLHF labeler comparing model outputs, and a foraging animal choosing patches. The analyst can locate the live question — is the choice set ill-defined? Is the evaluator ambiguous? Is the ordering partial or intransitive? Is the context shifting? — rather than re-deriving the structure for each domain.

In organizations and policy, this reframes preference-elicitation problems. Stakeholder consultations often collect raw rankings without specifying the choice set ("rank these priorities" — which priorities? compared to what omitted alternatives?), without disambiguating the evaluator ("the community ranks" — which community, aggregated how?), without checking ordering stability across framings, and without naming the context of consumption. The four-role decomposition makes each of these explicit and surfaces design choices that would otherwise be made implicitly.

Abstract Reasoning

Preference supports a family of substrate-neutral operations once the ordering is named. The analyst can take transitive closure — if X ≻ Y and Y ≻ Z, infer X ≻ Z under suitable conditions — to extend a partial ordering and reason about pairs the evaluator has not directly compared. They can check consistency (transitivity, completeness, independence of irrelevant alternatives) and use violations as diagnostic evidence that the underlying disposition is structured differently than assumed (reference-dependent, multidimensional, intransitive cycles). They can perform aggregation across multiple evaluators (voting rules, social welfare functions, Borda counts, Kemeny medians) and confront the impossibility results that come with it.

They can run counterfactuals: if the choice set were enlarged, would the favored option remain favored (independence of irrelevant alternatives, menu effects)? If the framing were inverted, would the ordering reverse (reference-point sensitivity)? If the evaluator were replaced by a similar evaluator, would the ordering survive (interpersonal stability)? Each counterfactual is a probe into a different structural property of the prime.

These operations let preference supply the objective that downstream procedures — optimization, decision, choice under constraint — consume. Subordinate primes inherit the operations and add curvature or asymmetry on top: time_preference specifies the shape of an evaluator's ordering across temporally separated outcomes (the curvature family Frederick, Loewenstein, and O'Donoghue (2002) survey across exponential and hyperbolic discounting); risk_aversion specifies the shape across probabilistically distributed outcomes; loss_aversion specifies an asymmetric ordering around a reference point that treats losses and gains differently (the reference-dependent shape Kahneman and Tversky (1979) introduced in prospect theory). Each presupposes the bare preference relation and adds structured deviations from the linear-in-magnitudes baseline. ^[10][11]

The reasoning also goes upward. From preference, the analyst can derive expected utility under risk by combining the ordering with a probability measure; can derive indifference curves by tracing iso-preference contours through the choice space; can derive marginal utility by examining how the ordering changes as a single dimension varies. These derived objects are not new primes but projections of the underlying preference relation under added structure.

Knowledge Transfer

The vocabulary travels intact across substrates. An economist studying utility functions, a behavioral psychologist studying prospect theory, a political theorist working on Arrow's impossibility, an ML researcher training a reward model from pairwise comparisons, and an ecologist documenting revealed foraging preference in pollinators are all working with the same four-role structure: choice set, evaluator, ordering, context. The transfer is structural, not metaphorical.

The animal case is especially load-bearing for substrate independence. A bumblebee's revealed preference between flower types has no utility function written down anywhere, no stated valuation, no language at all, and yet the ordering-from-an-evaluator's-standpoint pattern fits cleanly. That rules out the suspicion that preference is a specialty of economics or human cognition; it is a structural relation that any system capable of differential selection can instantiate.

Reward modeling in ML pulls the same trick from the other direction. The "evaluator" is a learned scoring function rather than an agent with introspective access; the choice set is a pool of candidate outputs (model completions, ranked items, action sequences); the ordering is fit from pairwise comparison data (the recipe Christiano et al. (2017) introduced for deep reinforcement learning from human preferences); the context is the downstream policy or recommender that will consume the ordering. The structural roles are all present. The substrate-furthest case is implicit revealed-preference learning from clickstream: the "evaluator" is a population of users whose orderings are inferred by the system from behavioral traces (clicks, dwell time, skips, returns) with no introspective access to the orderings themselves. The system never asks "what do you prefer?" — it watches behavior, models the ordering that would best explain it, and consults that model for ranking decisions. This is the prime running entirely outside human reflective practice. ^[3]

Practitioners moving between domains gain real leverage from carrying the prime's vocabulary intact. A behavioral economist recognizes Arrow's impossibility in mechanism-design problems they encounter in ML preference aggregation. An ML researcher recognizes intransitive cycles in their pairwise labeling data as a known structural phenomenon, not as label noise to be cleaned away. A political theorist working on participatory budgeting can borrow elicitation techniques developed for consumer choice. The transfer is not a metaphor that breaks under pressure; it is the same relation showing up in different implementations.

Examples

Formal/abstract

Consumer theory (microeconomics): Consider the canonical setup as developed in Mas-Colell, Whinston, and Green (1995). The choice set is a budget-constrained subset of a commodity space (bundles of goods at given prices and income); the evaluator is a consumer with a preference relation ≻; the ordering relation is assumed complete and transitive over the commodity space; the context is a budget-allocation problem the consumer is presumed to solve by selecting the most-preferred bundle in the budget set. From this minimal setup, an entire formal apparatus follows: indifference curves are iso-preference contours; marginal rates of substitution are local slopes of those contours; demand curves are projections of optimal choices as prices vary; consumer surplus is a welfare measure derived from the ordering plus a willingness-to-pay quantification. Mapped back: The four-role structure is fully exposed. Notice what is not in the setup: no utility function (one can be constructed under continuity assumptions following Debreu (1954), but the ordering is primary), no act of choosing (the formalism describes the structure of preference, not the dynamics of decision), no claim that the ordering is correct or rational (only that it satisfies the stated axioms). Stripping all of that out leaves the prime: an ordering, evaluator-relative, defined over a choice set, consumed by a context. ^[1][2]

RLHF preference learning (ML): Consider an RLHF training run, of the kind Christiano et al. (2017) introduced and Ouyang et al. (2022) scaled up in InstructGPT, that consumes a labeler's pairwise comparisons over model completions. The choice set is the pool of generated completions on a given prompt; the evaluator is the human labeler (or a population of labelers, aggregated); the ordering relation is the labeler's revealed pairwise judgments ("completion A is better than completion B"); the context is the training pipeline that will use the ordering to fit a reward model and shape policy. The four roles are immediately legible. The ordering is partial (the labeler did not compare every pair), possibly inconsistent across labelers (heterogeneity), possibly intransitive within a single labeler under fatigue or framing, and the trained reward model is an attempt to extrapolate the labeler's preference disposition to unseen pairs. Mapped back: The same four-role pattern fits a consumer choosing between grocery bundles, a voter ranking candidates, and a foraging bee picking flowers — what changes is which slot fills with what, not the structure of the relation itself. The ML case also makes the inference visible: the reward model is a learned approximation of the underlying preference relation, and the quality of the approximation depends on how the ordering itself is structured (transitivity helps; intransitive cycles do not). ^[3][12]

Applied/industry

Recommender systems learning from clickstream: A streaming service builds a recommender from implicit user signals. The choice set is the catalog of available items (films, songs, articles); the evaluator is each user (or a learned representation of each user); the ordering relation is inferred from behavioral traces — clicks, completion rate, dwell time, returns, skips — and represented as a learned scoring function that ranks items per user; the context is the front-end that surfaces ranked recommendations and the feedback loop that updates the model from continued interaction. No user is ever asked "what do you prefer?" The system constructs the ordering by Bayesian inference over the trace, treating behavior as a noisy sample from a stable underlying preference relation it never sees directly. Mapped back: This is the prime at its substrate-furthest case: no introspective access, no stated valuation, no language, no agent with conscious deliberation about the ordering — just behavior, inference, and a learned ordering consumed by ranking. The fact that the four-role structure still applies cleanly is the strong-form evidence that preference is a substrate-independent relation, not a feature of conscious deliberation. The same case also surfaces well-known structural pathologies: feedback loops where the system's recommendations shape the trace that the system then re-infers from, producing apparent preference shifts that are actually artifacts of the inference loop. Naming the prime separates these artifacts from genuine preference change.

Participatory budgeting (public policy): A city allocates a fraction of its capital budget through a participatory process. The choice set is a slate of capital projects (parks, transit improvements, library renovations); the evaluators are residents of the affected districts; the ordering relation is elicited via approval voting, ranked-choice voting, or quadratic voting depending on the design; the context is a budget allocation procedure that aggregates individual orderings into a collective decision under budget constraint. The structural design choices are visible at every step: how is the choice set selected (and by whom, screening out which alternatives)? How is the evaluator set defined (residents only? Property owners? Visitors? Weighted by stake?)? Which aggregation rule is used (and what impossibility results, in the line opened by Arrow (1951), does it skirt or accept)? How is the elicited ordering interpreted in the budget-allocation context, given that the elicitation framing itself can reverse the ordering as Tversky and Kahneman (1981) demonstrated for choices over public outcomes (lexicographic? Threshold? Welfare-weighted?)? Mapped back: Participatory budgeting is preference elicitation and aggregation made deliberately visible as a political process. The same four-role decomposition that handles a consumer at a checkout handles a city of a million making collective capital decisions; only the mechanism design at each slot is more elaborate. The structural unity is the point: the same prime supports both micro-level consumer theory and macro-level democratic mechanism design, and the design choices that look like political philosophy reduce, structurally, to choices about how to populate each slot of the four-role relation. ^[9][13]

Structural Tensions

T1: The ordering and the act diverge under constraint, noise, or akrasia. Preference is the standing ordering; choice is an act sampled from it under constraints, distractions, framing, and momentary state. A reveal-preference researcher treats the act as the canonical data on the ordering — yet the act is always preference-plus-noise-plus-constraint, and disentangling the components requires modeling assumptions that go beyond the prime itself. Treating choice as transparent to preference (the strong behaviorist move) is theoretically convenient and empirically false; treating choice as opaque to preference (the strong cognitivist move) gives up the only external access channel. The working compromise is always a structured inference under stated assumptions, and the assumptions do real work.

T2: Preference and value pull apart, especially under reflective scrutiny. An evaluator can prefer X to Y without endorsing the preference on reflection (the cheeseburger-vs-salad case, the addiction case, the impulse-purchase case), and can value Y over X while continuing to prefer X. Some traditions resolve the tension by privileging reflective endorsement (Frankfurt's higher-order desires, idealized-preference theories), others by privileging revealed orderings (behavioral welfare economics). The tension cannot be resolved at the level of the prime — preference names the ordering, and the ordering can disagree with what the evaluator says they value or what they would value under idealized conditions. Building any normative apparatus on top of preference requires deciding how to handle this gap.

T3: Substrate-furthest cases stress-test what counts as "an evaluator." A bumblebee, a clickstream-fed recommender, a thermostat, a noisy sensor, a learned reward model — at what point does the "evaluator" slot become so attenuated that calling it preference is overreach rather than substrate independence? The prime gains its power from working at arbitrary substrate distance, but the same power makes it easy to over-apply. Does a thermostat prefer temperatures near setpoint, or is that anthropomorphism dressing up a control loop? The structural answer (yes, if you populate the four roles cleanly) and the cautious answer (no, lest the prime dissolve into "any system with differential response") are in genuine tension, and the choice has consequences for downstream theorizing.

T4: Aggregation across evaluators faces structural impossibility results. Arrow's theorem shows that no aggregation rule can satisfy a modest set of fairness conditions (unrestricted domain, Pareto, independence of irrelevant alternatives, non-dictatorship) once we have three or more alternatives and two or more evaluators. Sen's liberal paradox and the Gibbard-Satterthwaite theorem extend the impossibility. The tension is structural rather than technical: any move to aggregate preference orderings into a collective ordering must sacrifice at least one intuitively desirable property. There is no neutral aggregation. This shapes how social choice, welfare economics, and mechanism design must proceed — by stating which properties they preserve and which they sacrifice, never by hoping for a clean aggregation.

T5: Stated and revealed preference systematically disagree, and neither is canonical. Stated preferences (what an evaluator says they prefer) and revealed preferences (what their behavior implies) often diverge — and not just because of measurement noise. Stated preferences are filtered through self-presentation, social desirability, and incomplete introspective access; revealed preferences are filtered through constraint, framing, habit, and choice-architecture. Each captures something the other misses. Privileging either as "the real preference" embeds a methodological commitment that the prime itself does not authorize. Practitioners must decide case-by-case which access mode to trust for which question, and the decision is rarely neutral.

T6: The choice set is rarely given; it is constructed, and the construction shapes the ordering. Preference is defined over a choice set, but in practice the set is selected — by a designer, by attention, by what is presented as alternatives. Menu effects (the presence of irrelevant alternatives changes the ordering over relevant ones), decoy effects (a dominated alternative shifts the ordering toward a target), and choice-architecture effects all show that the choice set is not innocent. The prime treats the set as given; the empirical reality is that whoever populates the set has substantial influence over what the consulted ordering looks like. This is not a defect of the prime — the prime is correctly modular at this seam — but it means that any preference-elicitation exercise is also, unavoidably, a choice-set construction exercise, and the two cannot be neatly separated in practice.

Structural–Framed Character

Preference sits at the structural end of the structural–framed spectrum, with one small framed-side caveat from its presupposition of an evaluator. Strip that to its formal core and what remains is the structure of an ordering relation over a choice set on some evaluative dimension — a pattern statable equivalently in microeconomics, decision theory, social choice, ML reward modeling, and any other substrate that supports orderings.

No domain vocabulary needs to come along: utility functions, rankings, revealed-choice tables, policy priorities, and learned reward signals are all implementations of the same underlying relation. The prime carries no evaluative weight — a preference ordering is descriptive of a comparative disposition, not normatively loaded. Institutional origin reads zero: orderings can be defined over any choice set, with no institution required. The half-step toward framed comes from human-practice-bound: every instance requires some evaluator (the agent, criterion, or system whose preference this is), and most paradigm cases are human or animal agents, though an ML reward model or a fitness function instantiates the relation just as cleanly. Import-vs-recognize is recognition: when a roboticist defines a reward function, they are exercising the ordering structure already required by the optimization problem, not importing decision-theoretic framing. On the spectrum, the verdict is structural with a mild evaluator-binding tint.

Substrate Independence

Preference is highly substrate-independent — composite 4 / 5 on the substrate-independence scale. The pattern is one substrate-neutral relation: an ordering or comparative evaluation over possible outcomes, actions, states, or bundles, from the standpoint of an evaluator. Domain breadth is at the ceiling because the same four-role structure (choice set, evaluator, ordering relation, context) recurs across economic utility, psychological preference, political and social orderings of priorities, machine-learning reward modeling, biological behavioral choice (revealed preference in animals), and aesthetic judgment. Structural abstraction is also at the top: the prime is defined purely as an ordering relation over a choice set, a fully relational signature with no home vocabulary required. Transfer evidence is high without being maximal because, while utility theory, revealed-preference analysis, and reward modeling have been ported across economics, psychology, biology, and AI with much success, the working vocabularies still cluster around agent-laden domains. The verdict is that preference is near the top of the scale, a structurally clean prime recognized wherever an evaluator orders alternatives, with its rating held just below the ceiling by the agent-centric clustering of its strongest evidence.

• Composite substrate independence — 4 / 5
• Domain breadth — 5 / 5
• Structural abstraction — 5 / 5
• Transfer evidence — 4 / 5

Relationships to Other Primes

Foundational — no parent edges in the catalog.

Children (10) — more specific cases that build on this

• Approach-Avoidance Conflict is a kind of Preference

  Approach-avoidance conflict is a specialization of preference. The general pattern is an agent's ordering over a choice set on some evaluative dimension, with a choice set, evaluator, ordering relation, and context. Approach-avoidance instantiates this with a single goal that simultaneously holds positive and negative valence, producing approach and avoidance gradients that are functions of proximity. The ordering relation is internally conflicted: at the crossover point neither dominates, yielding oscillation rather than stable selection. It is preference with the specific structural feature that valences are coupled rather than aligned on a single dimension.

• Loss Aversion is a kind of Preference

  Loss aversion is a kind of preference specialized by a reference-dependent, asymmetric value function: outcomes below the reference point are coded as losses and weighted more heavily than equivalent gains above it. It inherits preference's general commitment to an ordering over a choice set on an evaluative dimension, and supplies the specific case where the evaluative dimension is gain/loss relative to a reference point and the ordering relation systematically privileges loss avoidance — producing reproducible deviations from expected-utility predictions in framing, endowment pricing, and status-quo bias.

• Prioritization is a kind of Preference

  Prioritization actively orders competing claims on finite resources by some criterion of value, urgency, dependency, or feasibility and commits to honoring that ranking. The ranking step is exactly an instance of preference: an evaluator imposes an ordering relation over a choice set on some evaluative dimension. Prioritization specializes preference by binding the ranking to scarce resources and to a downstream commitment about which item is served first, turning a disposition into an enforceable sequence.

▸ Show 7 more

Neighborhood in Abstraction Space

Preference sits among the more crowded primes in the catalog (20^th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.

Family — Preferences, Trade-offs & Commensuration (9 primes)

Nearest neighbors

• Decision — 0.84
• Modal Reasoning — 0.82
• Preference Heterogeneity and Conflict — 0.81
• Comparison — 0.81
• Self Control — 0.81

Computed from structural-signature embeddings · 2026-05-29

Not to Be Confused With

Preference must be distinguished from utility. Utility is one particular quantification of a preference ordering — a real-valued function that represents the ordering when certain axioms hold. The classical von Neumann-Morgenstern theorem shows that under completeness, transitivity, continuity, and independence, a preference relation over lotteries can be represented by a utility function unique up to positive affine transformation; the Debreu representation theorem extends the result to richer settings. The crucial direction of dependence: the ordering is primary, the utility function is derived. Many ordering-equivalent utility functions encode the same preference (monotone transformations preserve the ordering), and many preferences cannot be represented as utility functions at all (lexicographic preferences resist real-valued representation, intransitive orderings violate the basic axioms). Calling preference "utility" conflates the prime with one of its representations and imports stronger assumptions than the prime itself requires.

Preference must be distinguished from choice. Choice is a discrete act of selecting one option from a set at a moment in time; preference is the standing ordering that may inform many choices, no choices, or choices that contradict the ordering. The two pull apart in multiple directions. Under constraint, the chosen option may not be the most-preferred (the consumer prefers the luxury car but chooses the affordable one). Under akrasia, the chosen option may be known to be worse on reflection (the addict chooses the cigarette while preferring health). Under framing or impulse, the chosen option may not even reflect the ordering the chooser would endorse moments later. Conversely, an ordering can exist with no choice attached: a person can prefer Beethoven to Bach without ever choosing between them at any moment in life. Behaviorist programs that try to define preference as the disposition to choose conflate the two; reflective-endorsement programs that try to define choice as preference-plus-deliberation conflate them in the other direction. The prime is the ordering; the act is downstream and noisy.

Preference must be distinguished from value. Value is the broader normative concept — a standing commitment to what should matter, often elaborated in moral, political, or aesthetic theories. Preferences may or may not reflect values, and the gap is theoretically important. Hedonic preferences (I prefer chocolate to vanilla) carry no obvious normative content; habitual preferences may reflect inertia rather than commitment; reflective preferences under idealized conditions may diverge sharply from immediate preferences. Values, by contrast, are typically articulated as endorsable in principle and defensible to others. A person can value justice while preferring convenience; can value health while preferring the cheeseburger; can value autonomy while preferring deference in particular moments. Building a normative theory on preference alone (preference utilitarianism, for example) commits to treating any ordering as authoritative; building a normative theory on value alone risks ignoring what evaluators actually rank. The two prime-level distinctions matter for downstream theorizing: preference is the ordering relation, value is the normative commitment, and conflating them obscures cases where they pull apart.

Preference must be distinguished from bias. A bias is a systematic deviation from a posited rational benchmark, and the benchmark itself is preference-theoretic. Expected utility theory specifies what a rational preference under risk should look like (transitive, complete, satisfying independence); empirical preferences that deviate systematically (loss aversion, framing effects, reference-dependence, the Allais paradox) are called biases relative to that benchmark. The prime structure is preserved at both ends — the benchmark is a preference relation with strong properties, the empirical observation is a preference relation with weaker or different properties. Bias is a diagnostic mode within preference theory, not an alternative to it. Treating bias as outside preference (as if biased agents have no preferences, only errors) misunderstands the relation: biased agents have preferences that systematically deviate from a normative target, and the deviation is itself informative about how the underlying disposition is structured. The prime survives the bias literature; the rational-agent caricature does not.

Preference must finally be distinguished by access mode: revealed preference vs. stated preference. Revealed preference is the ordering inferred from observed behavior; stated preference is the ordering an evaluator reports verbally or in self-report. The two are not the same prime — they are the same prime accessed differently. Revealed preference assumes that behavior is a reliable proxy for the ordering (subject to constraint, framing, and noise corrections); stated preference assumes that introspective access is reliable (subject to self-presentation, hypothetical-bias, and articulation corrections). Empirically they often disagree, sometimes substantially. The methodological dispute is about which access mode is more trustworthy for which question; the structural point is that both are inferences about the underlying ordering, neither is direct access. The prime is the ordering; revealed and stated are methodologies for surfacing it, and each comes with its own structural distortions. Treating either as canonical embeds a methodological commitment the prime itself does not authorize.

Solution Archetypes

No catalogued solution archetypes reference this prime yet.

Notes

ChatGPT Pro's R16 pass surfaced this as a clear gap, observing that the catalog already had a CONFLICT-flavored node (preference_heterogeneity_and_conflict) but not the underlying ordering prime. The case proved strong on review: the risk/utility/discounting/loss-aversion cluster of behavioral economics and decision theory was genuinely orphaned, and "preference" is the canonical vocabulary across all those fields. The candidate children re-home cleanly under the accepted prime; preference_heterogeneity_and_conflict becomes a sub-case (preferences differing across evaluators) rather than the umbrella.

The R17a/b/c wiring decisions are worth recording explicitly. The three "shape" children — time_preference, risk_aversion, loss_aversion — are all subsumption-type children of preference: each is a preference relation with added structure (intertemporal curvature, probability-weighting curvature, reference-point asymmetry). R17a confirmed subsumption for all three on the first pass; R17c re-confirmed loss_aversion as subsumption; R17b retyped risk_aversion to composition/presupposes after closer review of cases where risk preferences combine probability and utility components that can be modeled separately. The "derived object" children — expected_utility, marginal_utility, indifference_curves, preference_heterogeneity — are presupposes-type rather than subsumption-type: each uses a preference relation as input and adds structure on top (a probability measure, a differentiation operation, an iso-preference contour map, a multi-evaluator aggregation). The distinction between subsumption (the child is a structured instance of the parent) and presupposes (the child requires the parent as a component) is the load-bearing typology for the project-06b hierarchy work.

The clickstream-driven recommender case is worth flagging as the substrate-furthest concrete example currently catalogued. Unlike RLHF (where a human labeler is explicitly providing pairwise judgments), implicit revealed-preference learning operates entirely outside human reflective practice: no agent ever consults their ordering and reports it, no labeler ever provides a comparison, the system simply infers the ordering from behavioral traces. This is the substrate-distance regime where the prime is doing the most work — defending the structural unity of preference against the temptation to confine it to conscious deliberation. If the four-role decomposition still applies cleanly here (and it does), the substrate-independence claim is solid.

Open question worth carrying forward: the relationship to a possible value prime, if one is added to the catalog. Preference might decompose to ordering-over-value, but the simpler near-root positioning is probably right for v1 — preference is a structural relation that does not require an articulated value space to be operative, and the bumblebee/clickstream cases would be hard to handle under a value-decomposed positioning.

Preference is sometimes confused with optimization. The relation is asymmetric: optimization presupposes preference (it needs an objective, supplied by an ordering), but preference does not presuppose optimization (an ordering can exist and be consulted without any search-for-best occurring). The asymmetry is load-bearing for the catalog hierarchy: optimization is downstream of preference, not parallel to it.

References

[1] Mas-Colell, A., Whinston, M. D., & Green, J. R. (1995). Microeconomic Theory. Oxford University Press. Canonical graduate microeconomics textbook: develops the preference-based and choice-based approaches in parallel, takes the binary preference relation (with completeness and transitivity) as the primitive of consumer theory before introducing utility, and frames optimization as derived from a primitive preference ordering. ↩

[2] Debreu, G. (1954). Representation of a preference ordering by a numerical function. In R. M. Thrall, C. H. Coombs, & R. L. Davis (Eds.), Decision Processes (pp. 159–165). John Wiley & Sons. Foundational representation theorem: a complete, transitive, continuous preference relation on a suitable space admits a continuous real-valued utility representation; establishes that utility is one (non-unique) representation of an underlying ordering, not the primitive itself. ↩

[3] Christiano, P. F., Leike, J., Brown, T. B., Martic, M., Legg, S., & Amodei, D. (2017). Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems (NeurIPS 2017). Introduces the now-standard pipeline of fitting a reward model from human pairwise preference comparisons over agent trajectories and optimizing a policy against it; canonical reference for preference learning as a substrate for deep RL. ↩

[4] Sen, A. K. (1970). Collective Choice and Social Welfare. Holden-Day. Foundational treatment of preference aggregation: rigorously distinguishes structural preference incompatibility from coordination or information problems, developing the formal pattern of incompatible objectives producing collective decision impasse. ↩

[5] Bourdieu, P. (1984). Distinction: A Social Critique of the Judgement of Taste. Harvard University Press. foundational study of class boundaries organized through consumption patterns and cultural tastes, showing that class categories are sustained through continuous boundary-marking and that individuals deploy aesthetic judgment as boundary work. ↩

[6] von Neumann, J., & Morgenstern, O. (1944). Theory of Games and Economic Behavior. Princeton University Press. Axiomatic foundation of expected utility theory: shows that a preference relation over lotteries satisfying completeness, transitivity, continuity, and independence admits a utility representation unique up to positive affine transformation. ↩

[7] Samuelson, P. A. (1938). A note on the pure theory of consumer's behaviour. Economica, 5(17), 61–71. Original revealed-preference paper: proposes that consumer preference orderings be reconstructed from observed choices under varying prices and income, rather than from postulated utility, founding the revealed-preference research program. ↩

[8] Strotz, R. H. (1955). Myopia and inconsistency in dynamic utility maximization. The Review of Economic Studies, 23(3), 165–180. Foundational formalization of dynamic inconsistency: an agent's optimal plan over future consumption is not in general the plan that the agent's later self will choose to follow, generating preference reversals as the decision horizon approaches. ↩

[9] Arrow, K. J. (1951). Social Choice and Individual Values. Wiley. Foundational social-choice text containing the impossibility theorem: no aggregation rule over heterogeneous individual preferences can simultaneously satisfy unrestricted domain, Pareto efficiency, independence of irrelevant alternatives, and non-dictatorship—so any commensuration metric inevitably privileges some values over others. ↩

[10] Frederick, S., Loewenstein, G., & O'Donoghue, T. (2002). Time discounting and time preference: A critical review. Journal of Economic Literature, 40(2), 351–401. Comprehensive critical review of intertemporal-choice models: surveys the discounted utility model, its empirical anomalies, and alternative formulations (hyperbolic, quasi-hyperbolic, dual-self), with extensive evidence on the shape and stability of time preference. ↩

[11] Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–291. Foundational behavioral-economics result: outcomes are evaluated as gains and losses relative to a reference point rather than in absolute terms, with diminishing sensitivity and loss aversion — making the choice of baseline (and the contrast it creates with the treatment) constitutive of perceived value and decision behavior. ↩

[12] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., & Lowe, R. (2022). Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems (NeurIPS 2022). Scales the RLHF preference-learning pipeline to large language models (InstructGPT): collects pairwise human comparisons over model completions, fits a reward model, and optimizes the policy via PPO; canonical industrial application of preference learning at scale. ↩

[13] Tversky, A., & Kahneman, D. (1981). The framing of decisions and the psychology of choice. Science, 211(4481), 453–458. Seminal demonstration that the same problem framed differently produces predictable shifts of preference, explicitly likening frames to perceptual perspectives; supports the transfer of figure-ground reversibility and perceptual set to framing effects, where an audience may organize a message around a different figure than the one intended. ↩

[14] (definition not found) ↩