Decision-Evaluation Paradox

Updated 12 July 2026

Decision-Evaluation Paradox is a phenomenon where the criteria used to assess decisions diverge from the substantive properties that truly matter, resulting in evaluation mismatches.
The literature illustrates this paradox across fields such as natural language generation, clinical decision-making, and fairness auditing, with examples ranging from proxy metric failures to self-referential inconsistencies.
Methodological responses recommend multi-metric evaluations, human-in-the-loop verification, and tailored loss functions to ensure that evaluation aligns with genuine decision quality.

The Decision-Evaluation Paradox denotes a recurring pattern in which the criterion used to evaluate a system, output, or decision procedure diverges from the substantive property that matters for the decision itself. Across recent work, this divergence appears in several forms: intrinsic text metrics that weakly correlate with downstream decision efficacy, calibrated models whose decision rules rely on shortcuts, reinforcement-aligned clinical models that improve automatic scores while losing clinician trust, fairness criteria for classifiers that cannot be reconciled with fair rational downstream use, axioms that can drive a decision but fail when used to audit that same decision, and decision-theoretic losses that yield degenerate or otherwise pathological optima (Huang et al., 2 Jul 2025, Oh et al., 2024, Shukla et al., 20 May 2026, Liu et al., 22 Nov 2025, Morgan et al., 2017, Abramowitz et al., 26 Sep 2025, Kabaila, 2012). This suggests a broader family of paradoxes in which proxy metrics, average-case objectives, or self-referential evaluative schemes fail to preserve the property they were intended to secure.

1. Core formulation

In one prominent formulation from natural language generation, model outputs $S$ are scored by intrinsic metrics such as BLEU, ROUGE, or perplexity, while the actual value of $S$ lies in the quality of decisions $D$ that it informs. The paradox is that these evaluations can diverge sharply: an intrinsic score $I(S)$ can be high while decision efficacy $E(S)$ shows no improvement, with empirical correlation $\rho(I(S),E(S)) \approx 0$ or even negative in some settings (Huang et al., 2 Jul 2025). Here the evaluation target is surface alignment, but the operational target is downstream action.

A second formulation is evaluator incompetence. In question answering, a model may perform strongly on the SOLVE task—generating an answer—yet perform materially worse on the EVALUATE task—judging whether an answer is correct or incorrect. The paradox arises when a model correctly answers a question and then mislabels that answer as wrong, or produces a wrong answer and later judges it correctly as wrong; competence at producing a solution does not reliably translate into competence at assessing solutions (Oh et al., 2024).

A third formulation is self-referential inconsistency. In axiomatic choice, a decision is the triple $(x,f,y)$ , with profile $x$ , rule $f$ , and outcome $y=f(x)$ , and an axiom is a Boolean predicate $S$ 0. Under certain forcing axioms, the axiom determines a unique outcome for each profile, yet the implied rule $S$ 1 can yield a decision $S$ 2 that violates the very axiom used to derive it (Abramowitz et al., 26 Sep 2025). In this version, the paradox is not a proxy failure but a failure of reusing the same normative object for both forward choice and backward evaluation.

2. Decision-theoretic antecedents

A classical antecedent appears in statistical decision theory. In Recht’s decision-theoretic reconstruction of Meehl’s prediction problem, a decision rule $S$ 3 is evaluated by average loss

$S$ 4

The Bayes rule $S$ 5 minimizes conditional expected loss pointwise in $S$ 6, and therefore no other rule can achieve strictly lower risk under the chosen loss. Recht calls this “metrical determinism”: once a loss or scoring rule is fixed, the metric itself determines the optimal procedure (Recht, 4 Sep 2025). In this setting, the paradox is that evaluating clinical judgment only on average performance under a fixed metric effectively rigs the contest in favor of the statistical rule.

Kabaila’s note on decision-theoretic interval estimation gives a sharper pathology. For $S$ 7 with both $S$ 8 and $S$ 9 unknown, confidence intervals are assessed by expected length and coverage probability. A combined loss of the form

$D$ 0

with the standard noninformative prior

$D$ 1

produces the paradox identified by Casella, Hwang and Robert: for every fixed $D$ 2 and for any $D$ 3 up to a moderate size, the generalized Bayes minimizer collapses to the degenerate interval $D$ 4, with length $D$ 5 and coverage $D$ 6 (Kabaila, 2012). The evaluation criterion rewards zero length so aggressively that the object returned is not a usable confidence interval.

Kabaila removes the pathology by modifying the loss to

$D$ 7

Under the same prior, the generalized Bayes rule becomes exactly

$D$ 8

the usual $D$ 9 $I(S)$ 0-interval for $I(S)$ 1 (Kabaila, 2012). The significance is precise: a small change in how length enters the loss removes the degenerate optimum and restores the classical interval. This is a canonical example of an evaluation functional whose naive form destroys the purpose of the decision object it scores.

3. Text generation, decision quality, and evaluator competence

In decision-oriented text evaluation, Huang et al. replace reference-overlap metrics with direct measurement of how generated market digests affect decisions by human investors and LLM agents. Their framework evaluates morning briefs for intraday return prediction and closing-bell reports for next-day opening return, using thresholded correctness with $I(S)$ 2 for a correct buy and $I(S)$ 3 for a correct sell (Huang et al., 2 Jul 2025). Two asset-selection pipelines are used upstream of generation: Performance-Based, which selects top $I(S)$ 4 stocks by volatility, volume, or institutional flows, and Professional-Insight, which selects equities explicitly mentioned by human experts in the transcript.

The empirical pattern is the paradox in its most direct proxy-gap form. Intrinsic BLEU/ROUGE scores on LLM versus human briefs weakly correlate with decision accuracy, with $I(S)$ 5. Using journalist-authored morning briefs, neither humans nor LLMs beat random, with $I(S)$ 6. GPT-4o-generated briefs improve LLM accuracy from $I(S)$ 7 to $I(S)$ 8 under the Performance-Based pipeline or $I(S)$ 9 under Professional-Insight, while human investors improve from $E(S)$ 0 to $E(S)$ 1 or $E(S)$ 2. For closing-bell analyses, LLM agents perform best on journalist-written analyses at $E(S)$ 3, while human investors perform best on GPT-4o commentaries at $E(S)$ 4; under Professional-Insight filtering, combined human-LLM accuracy can exceed $E(S)$ 5 (Huang et al., 2 Jul 2025). Richer analytical commentaries matter more than surface similarity to a reference.

Oh et al. document a related but distinct paradox in evaluator competence using 905 TriviaQA pairs and four models—GPT-3.5, GPT-4, PaLM-2, and Vicuna-13B—run at temperature $E(S)$ 6 (Oh et al., 2024). Generation accuracy and evaluation accuracy diverge materially:

Model	Generation Accuracy	Evaluation Accuracy
GPT-3.5	0.79	0.62
GPT-4	0.88	0.80
PaLM-2	0.66	0.70
Vicuna-13B	0.46	0.33

GPT-3.5 and GPT-4 each show a statistically significant drop in evaluation accuracy versus generation accuracy, and the Pearson correlation between per-question generation correctness and evaluation correctness is modest at $E(S)$ 7 (Oh et al., 2024). Hallmark failure modes are “unfaithful evaluations”: approximately $E(S)$ 8– $E(S)$ 9 of self-generated correct answers are mislabeled as wrong, $\rho(I(S),E(S)) \approx 0$ 0– $\rho(I(S),E(S)) \approx 0$ 1 of self-generated wrong answers are labeled correct, and the “I don’t know” label is used in fewer than $\rho(I(S),E(S)) \approx 0$ 2 of cases. The result is not merely that evaluation lags generation; it can become detached from the model’s own demonstrated competence.

4. Reliability and alignment gaps in LLM-based decision systems

In machine unlearning, Shukla and Modi show that probabilistic reliability and decision-rule reliability can diverge sharply. They evaluate decoder-only LLMs on the TOFU benchmark in a multiple-choice setting using Expected Calibration Error, Maximum Calibration Error, and Brier score, while probing shortcut reliance by intersecting top- $\rho(I(S),E(S)) \approx 0$ 3 Integrated Gradients tokens with label-specific high-LMI correlation tokens and reporting shortcut-cued prediction rate $\rho(I(S),E(S)) \approx 0$ 4 (Shukla et al., 20 May 2026). Calibration improves dramatically after fine-tuning and remains low after unlearning—on the Retain split, ECE falls from $\rho(I(S),E(S)) \approx 0$ 5 in pretrained models to approximately $\rho(I(S),E(S)) \approx 0$ 6, and unlearned models retain ECE in the $\rho(I(S),E(S)) \approx 0$ 7– $\rho(I(S),E(S)) \approx 0$ 8 range.

The paradox is that this improved evaluation profile coexists with shortcut-driven decisions. In the reported $\rho(I(S),E(S)) \approx 0$ 9 forget / $(x,f,y)$ 0 retain setting, $(x,f,y)$ 1 is already high in fine-tuned models and remains above $(x,f,y)$ 2 after unlearning; the paper states that shortcut reliance often increases slightly, by more than $(x,f,y)$ 3 percentage points, even as calibration stays good (Shukla et al., 20 May 2026). Good calibration therefore does not certify that the model’s decision rule is meaningful or generalizable.

A clinically grounded variant appears in infertility care. Liu et al. compare Supervised Fine-Tuning, Direct Preference Optimization, Group Relative Policy Optimization, and In-Context Learning on an 8 billion-parameter OpenBioLLM backbone using more than 8,000 infertility treatment records. Their dual-layer evaluation combines automatic metrics—accuracy, macro- $(x,f,y)$ 4, semantic entailment match rates, and gonadotropin starting dose MAE—with blinded doctor-in-the-loop assessments across Diagnosis Accuracy, Reasoning Clarity, Clinical Feasibility, and Hallucination Rate (Liu et al., 22 Nov 2025).

Automatic benchmarks and clinician judgments diverge. GRPO achieves the highest overall average accuracy at $(x,f,y)$ 5 and macro- $(x,f,y)$ 6 at $(x,f,y)$ 7, with field-level gains such as $(x,f,y)$ 8 percentage points in ART strategy accuracy and $(x,f,y)$ 9 percentage points in exact initial diagnosis matches versus SFT. Yet clinicians rate SFT higher on Reasoning Clarity, $x$ 0 versus $x$ 1 with $x$ 2, and on Clinical Feasibility, $x$ 3 versus $x$ 4 with $x$ 5; in blind pairwise comparisons, SFT is preferred $x$ 6 of the time, compared with $x$ 7 for GRPO and $x$ 8 for the physician’s original decision (Liu et al., 22 Nov 2025). The paper attributes the divergence to reward-centric GRPO optimization of final-answer correctness, which can erode chain-of-thought integrity, whereas SFT preserves human-style stepwise justification.

5. Fairness, axioms, and the structure of evaluative inconsistency

In computer-aided decision making, Morgan and Pass formalize a classification context $x$ 9, classifier fairness as $f$ 0-fair treatment, and decision-maker fairness as $f$ 1-rational fairness. Their main theorem states that for non-trivial CADM, either the classifier must be discriminatory, or a rational decision-maker using the output of the classifier is forced to be discriminatory (Morgan et al., 2017). The only settings in which both can be achieved are $f$ 2-trivial contexts: those admitting a partition of the class space into blocks with nearly equalized base rates within each block and perfect separability across blocks.

This is a Decision-Evaluation Paradox because fairness at the level of the classifier does not survive the evaluative rationality imposed on the downstream human actor. A classifier may satisfy fair treatment, but once the decision-maker is asked to maximize expected utility using group and classifier output, rational fairness fails except in trivial settings (Morgan et al., 2017). The paradox therefore lies between two evaluation frames: fairness of the predictive instrument and fairness of its rational use.

Abramowitz and Mattei push the issue into explicit axiomatics. They define an axiom $f$ 3 as a predicate on decisions $f$ 4 and classify axioms as positively trivial, negatively trivial, structural, procedural, consequentialist, blackbox, caudal, or exigent. Forcing blackbox and forcing procedural axioms imply rules that satisfy their own tests, but there exist forcing caudal or exigent axioms such that the implied rule $f$ 5 yields

$f$ 6

even though $f$ 7 is the unique allowed outcome for that profile (Abramowitz et al., 26 Sep 2025). They further characterize the paradox as arising precisely when an axiom is forcing and also an Arrovian impossibility theorem. In these cases, the same axiom cannot stably serve as both decision driver and decision auditor.

Taken together, these works show two structurally different failures. In the fairness literature, evaluation at one layer does not compose with rational action at another. In axiomatic choice, evaluation of the realized decision triple can contradict the forward use of the very axiom that produced the decision. This suggests that “evaluation” is not a neutral post hoc activity; it changes which objects are being judged and can alter the feasibility of the normative program itself.

Cui identifies a paradox in individualized decision-making under partial identification. With binary treatment $f$ 8, covariates $f$ 9, and instrumental-variable bounds $y=f(x)$ 0 on the conditional average treatment effect, the deterministic minimax-regret rule can misclassify an entire subgroup even when there is no actual unmeasured confounding. In the toy model, the true optimal rule is $y=f(x)$ 1 for both $y=f(x)$ 2 and $y=f(x)$ 3, yet the minimax-regret rule selects $y=f(x)$ 4 and $y=f(x)$ 5 (Cui, 2021). Cui summarizes the effect as “1+1<2”: individualized decision-making plus IV partial identification can be worse than either individualized decision-making under unconfoundedness or population-level IV-bounds reasoning alone.

A related class of reversals appears when additional information or additional options worsen decisions rather than improving them. In the analysis of Simpson’s reversal and the disjunction effect, additional or differently aggregated data can flip threshold-based decisions; the paper argues that classical threshold or causal-calculus approaches do not capture the way the act of observing or not observing an intermediate result changes subsequent probabilities, whereas a quantum-vector model builds in non-commutativity, collapse, and interference (Kak, 2017). In entropy-driven collective decision dynamics, the “paradox of choice” appears as a non-monotonic synchronization time $y=f(x)$ 6: too few neighbors yields weak coupling, too many neighbors yields information overload, and there is an optimal sensing radius $y=f(x)$ 7; biased sensing or memory removes the paradox (Gupta et al., 2024). These are not identical to the Decision-Evaluation Paradox, but they exhibit the same general form in which seemingly stronger informational or evaluative conditions do not monotonically improve decision quality.

Across the cited literature, several methodological responses recur. Decision-oriented evaluation recommends measuring thresholded accuracy, financial returns, clinical outcomes, or analogous domain-specific endpoints alongside intrinsic scores (Huang et al., 2 Jul 2025). Reliability assessment in unlearning recommends jointly measuring calibration and decision-rule reliability rather than relying on ECE, MCE, or Brier alone (Shukla et al., 20 May 2026). Clinical alignment work recommends explicitly rewarding coherent, clinically faithful reasoning chains rather than optimizing only final-answer correctness (Liu et al., 22 Nov 2025). Classical decision theory and CADM work recommend multi-metric evaluation, human-in-the-loop overrides or verification, periodic auditing and retraining, and careful scrutiny of whether the loss, fairness criterion, or axiom being optimized is narrow enough to predetermine an undesirable or inconsistent solution (Recht, 4 Sep 2025, Oh et al., 2024, Abramowitz et al., 26 Sep 2025). The common implication is that evaluation must be matched to the object one actually cares about—decision quality, reasoning integrity, fairness of deployment, or normative compliance—rather than to a convenient proxy or a self-referential formal test.