Consequence-Based Utility in Decision Models

• Consequence-Based Utility is a formal framework that assigns value to decisions based on the structured evaluation of induced consequences rather than isolated outcomes.
• It integrates diverse methodologies—such as oracle-free evaluation in mathematical proofs, verifiability in decision models, and counterfactual utility—to refine model performance.
• The framework underpins practical applications in reinforcement learning and cost-sensitive resource allocation while expanding traditional consequence spaces in decision theory.

Searching arXiv for the cited papers and closely related work on consequence-based utility. Consequence-based utility denotes a family of formalisms in which evaluation is anchored in consequences rather than in syntactic form, isolated judgments, or realized outcomes alone. In its most classical usage, the primitive objects are consequences $C$, a preference ordering $\succeq$ over $C$, and a real-valued utility representation $u:C\to\mathbb R$ such that $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$ (Kohli et al., 2023). Recent work uses the same consequence-oriented idea in several technically distinct ways: as an oracle-free evaluator for research-level mathematics that scores candidate proofs by downstream transfer to related verifiable questions (Son et al., 6 Feb 2026); as a decision model in which verifiable events mediate between expected utility and maximin (Rommeswinkel, 27 Aug 2025); as utility over cumulative reward in reinforcement learning (2402.02665); and as utility defined on extended potential-outcome spaces that include counterfactual consequences (Koch et al., 6 May 2026). A plausible implication is that “consequence-based utility” is now best understood as a research program rather than a single formalism: the common thread is that the utility-bearing object is some consequence-bearing structure, while the precise consequence space varies by domain.

1. Consequence spaces, preference orderings, and measurement

A standard starting point is a consequence space together with a complete and transitive preference relation. In the differential-privacy treatment, $C$ is the set of all outcomes a participant might incur, and each participant has a binary relation $\succeq$ on $C$ satisfying completeness and transitivity; $u:C\to\mathbb R$ represents $\succeq$ exactly when $\succeq$0 (Kohli et al., 2023). Because only the ordering matters, utility is determined up to a positive-affine transformation, so $\succeq$1 is an interval scale.

The same paper emphasizes that expected utility is evaluated after a mechanism and post-processor induce a distribution over consequences. For an $\succeq$2-differentially private mechanism $\succeq$3 and post-processor $\succeq$4, the expected utility at input $\succeq$5 is

$\succeq$6

It then compares privacy guarantees by Euclidean distance between vectors of expected utilities across neighboring-input pairs, with the argument that positive-affine changes in $\succeq$7 preserve ordering conclusions about these differences (Kohli et al., 2023).

This usage preserves the classic consequence-based intuition: decisions are compared through the utility of the consequences they induce. At the same time, later work broadens both the consequence space and the aggregation rule. In the verifiability model, consequences are aggregated pessimistically within verifiable events and probabilistically across them (Rommeswinkel, 27 Aug 2025). In counterfactual utility, the relevant “consequence” is no longer merely the realized outcome but the full vector of potential outcomes under all available decisions (Koch et al., 6 May 2026). This suggests that current literature treats the consequence space itself as a modeling choice rather than a fixed primitive.

2. Consequence-Based Utility as oracle-free evaluation in research-level mathematics

In "Judging What We Cannot Solve: A Consequence-Based Approach for Oracle-Free Evaluation of Research-Level Math" (Son et al., 6 Feb 2026), Consequence-Based Utility (CBU) is a specific evaluator for candidate mathematical solutions when no oracle is available for the target problem. The paper’s motivation is that research-level mathematics produced by LLMs is often high-variance, while verification typically requires a domain expert or an automated theorem prover. Existing oracle-free validators inspect each solution in isolation and suffer from biases such as verbosity bias and authority bias. CBU replaces direct inspection with “support by consequences”: a correct or near-correct proof sketch should contain enough method-level detail to help solve related, simpler variants of the original question.

Formally, the target problem is $\succeq$8, a candidate solution is $\succeq$9, $C$0 is a set of neighborhood questions $C$1 whose answers are verifiable, $C$2 is a fast verifier, and $C$3 is the same solver model conditioned on the original problem, the candidate exemplar, and the variant question. The paper defines

$C$4

with the practical estimator

$C$5

Algorithmically, each candidate is scored by constructing or retrieving a small neighborhood, running $C$6 independent rollouts per variant, verifying the resulting answers, and ranking candidates by descending $C$7 (Son et al., 6 Feb 2026).

The experimental setting uses ExpertMath: 192 expert-written research-level problems spanning algebraic combinatorics, geometry, homotopy theory, and related areas; for each original $C$8, the authors hand-craft up to 2 neighborhood questions, with total variants $C$9; and for each original $u:C\to\mathbb R$0, they sample 9 LLM-generated solutions, consisting of 4 verified correct and 5 verified incorrect, from a mixture of GPT-OSS-120B, GPT-OSS-20B, GPT-5 (and Pro versions), Gemini-3-Pro, and Gemini DeepThink (Son et al., 6 Feb 2026).

The reported ranking gains are large. For GPT-OSS-120B, Acc@1 improves from 67.21 to 76.27 and AUC from 71.42 to 79.63; for GPT-OSS-20B, AUC improves from 69.03 to 79.18. Across four backbones, CBU improves Acc@1 by $u:C\to\mathbb R$1 to $u:C\to\mathbb R$2 and AUC by $u:C\to\mathbb R$3 to $u:C\to\mathbb R$4, while reward models and generative reward models lag far behind, with AUC approximately $u:C\to\mathbb R$5–$u:C\to\mathbb R$6 (Son et al., 6 Feb 2026). The paper also reports a larger solver–evaluator gap than LLM-Judges: as question difficulty increases, LLM-Judge separability collapses, whereas CBU maintains a robust gap even on near-unsolved instances. Among 112 cases where CBU downranks a solution that the judge over-ranks, common failure modes include incorrect reasoning (68.8%), unjustified compression or missing critical steps (71.4%), and external citations without derivations (31.3%).

Within this literature, CBU is consequence-based in a precise operational sense: the value of a candidate proof is identified by the consequences of conditioning future problem-solving on that proof. A plausible implication is that this reframes verification as a transfer problem: rather than asking whether a proof “sounds correct,” the evaluator asks whether the proof causes better performance on a neighborhood of questions that can actually be checked.

3. Verifiability, maximin structure, and the recovery of subjective consequence cells

"Preference for Verifiability" extends the usual consequence-based approach to decision under uncertainty by introducing subjective verifiable events (Rommeswinkel, 27 Aug 2025). The primitives are a finite state space $u:C\to\mathbb R$7, a consequence space $u:C\to\mathbb R$8, acts $u:C\to\mathbb R$9, and an algebra of events $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$0. An event $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$1 is called verifiable if, whenever the true state $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$2, the decision maker expects to be able to prove that $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$3 occurred. The family of such events, $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$4, is required to be a $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$5-system: closed under finite intersections and containing at least the whole $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$6.

The representation theorem states that a preference relation satisfying biseparability, comonotonic independence, supermodularity, and critical-event modularity is equivalent to the existence of a $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$7-system of verifiable events $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$8, a probability measure $c_1\succeq c_2 \Leftrightarrow u(c_1)\ge u(c_2)$9 on $C$0 that is grounded, normalized, and additive on $C$1, and a von Neumann–Morgenstern utility $C$2, such that

$C$3

where $C$4 is the Möbius inverse of $C$5 restricted to $C$6. In many applications the paper writes the simpler form

$C$7

If $C$8 partitions $C$9 into singletons, the model recovers Subjective Expected Utility; if $\succeq$0 only, it becomes classical Gilboa–Schmeidler maximin (Rommeswinkel, 27 Aug 2025).

The carbon-emission technology example makes the structure concrete. With $\succeq$1, acts Trees, RECs, and Efficiency, and $\succeq$2, a verification seeker evaluates

$\succeq$3

If $\succeq$4 is small, Trees is preferred because $\succeq$5 beats $\succeq$6; if $\succeq$7 is large, the decision maker switches to Efficiency to avoid the very low worst-case of 10 (Rommeswinkel, 27 Aug 2025).

This framework is still consequence-based, but the relevant consequence is not merely state-contingent payoff. What matters is what can later be verified about the consequence. A plausible implication is that consequence-based utility here internalizes an epistemic asymmetry: the decision maker values consequences partly through the future evidential position from which those consequences can be certified.

4. Extended consequence spaces: counterfactual utilities and model-based utilities

"An Axiomatic Foundation for Decisions with Counterfactual Utility" enlarges the domain further by defining utility on the full potential-outcome space rather than on realized outcomes alone (Koch et al., 6 May 2026). The decision set is $\succeq$8; each decision $\succeq$9 has a potential outcome $C$0 taking values in a finite set $C$1; the joint potential-outcome space is $C$2; and $C$3 is the full potential-outcome space, possibly including covariates. A counterfactual utility is then

$C$4

and the expected value of a policy $C$5 in state $C$6 is

$C$7

The central theorem is a direct vNM representation on $C$8: completeness, transitivity, independence, and continuity on lotteries over the extended space are equivalent to the existence of $C$9 such that preferences are represented by expected counterfactual utility. Additional axioms then recover narrower forms. Axiom 5, irrelevance of untaken decisions, yields a standard realized-outcome utility $u:C\to\mathbb R$0; Axiom 6, irrelevance of counterfactual correlation, yields additive counterfactual utility

$u:C\to\mathbb R$1

which the paper identifies as exactly the class whose expectation depends only on the marginals of $u:C\to\mathbb R$2 (Koch et al., 6 May 2026).

The Russian roulette and Allais-paradox examples illustrate how apparently nonstandard choice patterns can be represented coherently on this enlarged space. In the Russian roulette example, a physician may prefer a treatment with slightly lower average survival if it reduces the chance of directly causing harm to a patient who would otherwise survive. In the Allais-paradox example, regret-type counterfactual utility

$u:C\to\mathbb R$3

reproduces the empirical pattern $u:C\to\mathbb R$4 yet $u:C\to\mathbb R$5 for sufficiently large $u:C\to\mathbb R$6, without violating vNM on the extended domain (Koch et al., 6 May 2026).

A distinct but related expansion appears in "Model-based Utility Functions" (Hibbard, 2011). There, utility is not written directly on the observed interaction history but is computed in two steps: first infer an environment model $u:C\to\mathbb R$7 from history $u:C\to\mathbb R$8 by Bayes’s rule, then compute utility as a function of the learned model. With $u:C\to\mathbb R$9, internal-state histories $\succeq$0 compatible with $\succeq$1, and $\succeq$2, the induced utility is

$\succeq$3

The paper argues that this avoids self-delusion pathologies associated with history-based utility functions, because actions that cut off genuine observation increase uncertainty about the hidden variables on which utility depends (Hibbard, 2011).

Taken together, these works show two distinct methods for enlarging consequence-based utility. Counterfactual utility enlarges the consequence space to include unrealized alternatives; model-based utility interposes a learned world model between observations and utility. In both cases, the consequence-bearing object is richer than an observed payoff stream.

5. Utility-based reinforcement learning and sequential consequence aggregation

"Utility-Based Reinforcement Learning: Unifying Single-objective and Multi-objective Reinforcement Learning" defines a utility-based paradigm in which both environmental rewards and a user-specified utility function determine the optimization target (2402.02665). The setting has a finite state space $\succeq$4, a finite action space $\succeq$5, transition kernel $\succeq$6, discount factor $\succeq$7, start-state distribution $\succeq$8, and immediate reward vector $\succeq$9. A utility function

$\succeq$00

assigns scalar utility to cumulative vector outcomes, and the objective is

$\succeq$01

The paper distinguishes Scalarised Expected Return,

$\succeq$02

from Expected Scalarised Return,

$\succeq$03

When $\succeq$04 is nonlinear, standard Bellman recursion no longer applies directly in the original state space. The proposed remedy is an augmented state $\succeq$05, where $\succeq$06 accumulates partial return. The augmented value function satisfies

$\succeq$07

with a corresponding action-value function $\succeq$08 (2402.02665).

The paper states that single-objective MDPs are recovered by taking $\succeq$09 and $\succeq$10, while arbitrary multi-objective problems are also handled by the same objective; hence UBRL strictly generalises both single-objective and multi-objective RL. It then sketches augmented-state value iteration, augmented-space Q-learning, and a conditioned network $\succeq$11 for multi-policy learning across parameterised utilities $\succeq$12. The listed applications include uncertain objectives, CVaR-based risk sensitivity, multi-policy discounting, and satisficing or safe RL (2402.02665).

In this sequential setting, consequence-based utility means that the decision criterion is applied to cumulative outcomes rather than to one-step rewards alone. A plausible implication is that the framework turns many “nonstandard” RL desiderata into ordinary dynamic programming over an augmented consequence accumulator.

6. Consequence-aware deployment, consequence-sensitive compression, and downstream control

Recent machine-learning work uses consequence-based ideas to regulate inference-time resources rather than only to rank terminal outcomes. "Not All Errors Are Equal: Consequence-Aware Reasoning Compute Allocation" formulates test-time compute allocation as cost-weighted loss minimization under a total-compute budget (Wen et al., 3 Jun 2026). With tasks $\succeq$13, compute tiers $\succeq$14, tier price $\succeq$15, consequence label $\succeq$16, and success probability $\succeq$17, the primary objective is

$\succeq$18

The marginal-utility signal is

$\succeq$19

At the top-25% premium operating point on 300 SWE-bench Lite tasks, the priority-aware strategy with the Qwen predictor reduces cost-weighted loss to $\succeq$20, compared with $\succeq$21 for difficulty-aware routing, a $\succeq$22 improvement relative to difficulty-aware routing; the oracle priority-aware variant attains $\succeq$23, and the predictor-driven version retains 92.9% of the oracle priority gain (Wen et al., 3 Jun 2026).

A related metacognitive formulation appears in "Support Sufficiency as Consequence-Sensitive Compression in Belief Arbitration" (Walsh, 6 Apr 2026). The system maintains a hypothesis support geometry $\succeq$24, a consequence geometry $\succeq$25, and arbitration memory $\succeq$26. Resolution is selected by

$\succeq$27

the support code and selected content are

$\succeq$28

and policy is chosen through

$\succeq$29

The paper defines a compressed support code as sufficient if

$\succeq$30

where $\succeq$31 is the ideal full-geometry policy. It then introduces the bounded objective

$\succeq$32

Across 10 random seeds $\succeq$33 50 000 trials, Adaptive–Agile attains cumulative utility $\succeq$34, Adaptive–Slow $\succeq$35, Fixed–High $\succeq$36, Fixed–Mid $\succeq$37, and Fixed–Low $\succeq$38; Fixed–High achieves the best immediate accuracy, 0.8725, but still trails adaptive controllers in cumulative utility because resource and fragmentation costs offset the gains from richer retention (Walsh, 6 Apr 2026).

These two papers move consequence-based utility away from ex post evaluation of decisions alone and toward regulation of computation, compression, and control. The consequence-bearing object is no longer only the external task outcome; it also includes the downstream cost of routing, verification, abstention, and recovery under limited resources.

7. Limitations, variants, and conceptual boundaries

The main limitations are domain-specific. In research-level mathematics, CBU requires neighborhood questions $\succeq$39, which for research-level problems currently must be hand-crafted or rely on imperfect automatic generation; automatic neighborhood construction is partly model-dependent because variants must be neither too easy nor too hard (Son et al., 6 Feb 2026). In the verifiability model, $\succeq$40 need not be point-identified, although its atoms and the closure of $\succeq$41 under unions are uniquely recovered from preferences (Rommeswinkel, 27 Aug 2025). In counterfactual utility, menu-dependent projections can violate Sen’s $\succeq$42 and need not be transitive, even though the underlying preference on the full potential-outcome space is coherent (Koch et al., 6 May 2026). In utility-based RL, nonlinear utility generally requires an augmented state or related machinery because the standard Bellman recursion no longer applies directly (2402.02665).

The literature also contains closely related but non-identical notions. The privacy paper is explicitly consequence-based in the classical decision-theoretic sense, using a consequence space $\succeq$43, preference ordering, and interval-scale utility representation (Kohli et al., 2023). The verifiability paper modifies the aggregation rule over state-contingent consequences (Rommeswinkel, 27 Aug 2025). The counterfactual paper redefines the consequence domain itself (Koch et al., 6 May 2026). The CBU paper in mathematical reasoning uses “utility” as downstream transferable problem-solving value (Son et al., 6 Feb 2026). A plausible implication is that the phrase now covers at least three technical families: utility over realized consequences, utility over extended or epistemically structured consequence spaces, and utility defined by downstream instrumental consequences.

Future directions are likewise heterogeneous. The math-evaluation paper proposes fully automated, reliable variant generation, extension to other STEM domains such as physics and chemistry, and evaluation on genuinely open conjectures where ground truth and variants are inherently unavailable (Son et al., 6 Feb 2026). Consequence-aware compute allocation points toward deployment recipes that combine a cost predictor, a scheduler, and a matched budget plan on top of off-the-shelf LLMs (Wen et al., 3 Jun 2026). Consequence-sensitive compression suggests treating compression itself as a regulated decision variable under changing stakes and memory (Walsh, 6 Apr 2026). Taken together, these directions indicate that consequence-based utility is increasingly used not only to compare actions, but also to decide what to verify, what to retain, how much compute to spend, and which parts of a proof, belief state, or counterfactual structure matter for downstream control.