Wild Optimism in Learning and Decision Theory

• Wild optimism is a multifaceted concept where constructive estimators, such as wild refitting in penalized ERM, provide computable upper bounds on excess risk.
• It appears in reinforcement learning and bandits as overoptimistic value estimation or exploration that can lead to instability and inefficiency.
• In game theory and medical ML, wild optimism manifests as aggressive best-case reasoning or misleading comparisons, emphasizing the need for precise uncertainty calibration.

Searching arXiv for the cited paper and related uses of “wild optimism.” “Wild optimism” is used in several distinct technical senses across current arXiv literatures. In statistical learning, it is the name of a computable estimator produced by wild refitting and used to upper-bound excess risk for penalized ERM under Bregman losses (Hu et al., 2 Sep 2025). In multiple reinforcement-learning papers, it denotes an undesirable regime in which optimism is too aggressive, fixed, or noise-insensitive, so that value estimates or exploration become unstable or wasteful (Tasdighi et al., 2024); (Liu et al., 2023); (Cheung et al., 2019). In decision theory and game theory, it is closely related to maxmax reasoning under coarse beliefs or ignorance, where actions are evaluated by their best-case consequences (Sinander, 2023); (Guarino et al., 2021). In medical machine learning, an analogous phenomenon appears as baseless optimism induced by weak baselines and misleading comparisons (Demasi et al., 2017).

1. Statistical learning: wild optimism as a model-free excess-risk estimator

In "Wild Refitting for Model-Free Excess Risk Evaluation of Opaque ML/AI Models under Bregman Loss" (Hu et al., 2 Sep 2025), wild optimism is the central estimator for excess-risk evaluation of black-box penalized ERM under Bregman losses. The setting uses outputs in $\mathcal{Y}\subset \mathbb{R}^d$, predictors $f:\mathcal{X}\to\mathcal{Y}$, and a differentiable convex potential $\phi:\mathbb{R}^d\to\mathbb{R}$, with loss

$$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$

The paper assumes $\phi$ is $\beta$-smooth and $\alpha$-strongly convex, with unique minimizer. Under strict convexity, the optimal predictor is the conditional mean in Bregman geometry: $f^*(x_i)=\mathbb{E}[y_i\mid x_i]$ in fixed design and $f^*(x)=\mathbb{E}[Y\mid X=x]$ in random design (Hu et al., 2 Sep 2025). The learned predictor $\hat f$ is obtained by a black-box training procedure $f:\mathcal{X}\to\mathcal{Y}$0 that approximately solves penalized ERM on one dataset.

The paper decomposes empirical excess risk into a Bregman discrepancy term and a stochastic inner-product term, the latter called the true optimism. The resulting objective is to upper-bound this optimism term without invoking VC dimension, covering numbers, or Rademacher complexity of the global function class (Hu et al., 2 Sep 2025). This is the point at which wild refitting enters.

Wild refitting computes residuals $f:\mathcal{X}\to\mathcal{Y}$1, flips them coordinatewise using vector-valued Rademacher noise $f:\mathcal{X}\to\mathcal{Y}$2, forms wild responses

$f:\mathcal{X}\to\mathcal{Y}$3

and retrains once on the perturbed labels to obtain $f:\mathcal{X}\to\mathcal{Y}$4 (Hu et al., 2 Sep 2025). The corresponding wild optimism is

$f:\mathcal{X}\to\mathcal{Y}$5

This quantity is entirely computable from the two fits and the observed data. The paper proves that, with an appropriately chosen wild noise scale $f:\mathcal{X}\to\mathcal{Y}$6, wild optimism upper-bounds a local symmetrized empirical process and, with additional correction terms, yields high-probability upper bounds on empirical excess risk in fixed design and population excess risk in random design (Hu et al., 2 Sep 2025). The method is explicitly described as model-free and black-box because it uses only one dataset, requires only oracle access to $f:\mathcal{X}\to\mathcal{Y}$7, and does not rely on explicit structural information about $f:\mathcal{X}\to\mathcal{Y}$8.

This usage is unusual within the broader “optimism” vocabulary. Here the phrase is not pejorative. It denotes a constructive estimator that inflates observed training error by a computable, refitting-based term in order to upper-bound true excess risk. A plausible implication is that the phrase “wild optimism” in this paper is best understood as a data-driven, randomized analogue of classical optimism penalties, specialized to opaque learners and Bregman geometry.

2. Reinforcement learning: uncontrolled optimistic exploration and value inflation

In several reinforcement-learning papers, “wild optimism” has the opposite valence. It refers to a regime in which optimistic estimates are too aggressive relative to uncertainty, leading to instability, over-exploration, or overestimation.

In "Exploring Pessimism and Optimism Dynamics in Deep Reinforcement Learning" (Tasdighi et al., 2024), wild optimism is the regime where optimism in value estimates is strong, fixed, and insensitive to uncertainty. The paper studies deep off-policy actor-critic methods and introduces Utility Soft Actor-Critic (USAC), which controls optimism and pessimism separately for actor and critic by means of an entropic-utility construction

$f:\mathcal{X}\to\mathcal{Y}$9

For a Laplace $\phi:\mathbb{R}^d\to\mathbb{R}$0-distribution, this becomes

$\phi:\mathbb{R}^d\to\mathbb{R}$1

so $\phi:\mathbb{R}^d\to\mathbb{R}$2 is optimistic and $\phi:\mathbb{R}^d\to\mathbb{R}$3 is pessimistic (Tasdighi et al., 2024). The paper’s empirical message is that overestimation consistently hurts performance, best critic settings are always pessimistic, and actor optimism is task-dependent. This suggests that uncontrolled optimism is specifically dangerous in the critic, while calibrated actor optimism can be beneficial when tied to uncertainty.

In "OVD-Explorer: Optimism Should Not Be the Sole Pursuit of Exploration in Noisy Environments" (Liu et al., 2023), wild optimism is the tendency of optimism-in-the-face-of-uncertainty to over-explore regions with high aleatoric noise. The paper decomposes uncertainty in distributional RL into epistemic and aleatoric components and constructs an optimistic value distribution

$\phi:\mathbb{R}^d\to\mathbb{R}$4

Its exploration signal is based on a mutual-information quantity involving the CDF of a pessimistic current value distribution (Liu et al., 2023). The explicit aim is to retain optimism about epistemic uncertainty while down-weighting attraction to irreducible noise. In this literature, wild optimism is therefore a failure to distinguish learnable uncertainty from stochastic variability.

In "Non-Stationary Reinforcement Learning: The Blessing of (More) Optimism" (Cheung et al., 2019) and the closely related "Reinforcement Learning for Non-Stationary Markov Decision Processes: The Blessing of (More) Optimism" (Cheung et al., 2020), optimism becomes “wild” in a different sense. Standard confidence sets in drifting MDPs can force optimistic planning over empirical models with very large diameter, even when each true $\phi:\mathbb{R}^d\to\mathbb{R}$5 has small diameter. The remedy is confidence widening: $\phi:\mathbb{R}^d\to\mathbb{R}$6 The extra slack $\phi:\mathbb{R}^d\to\mathbb{R}$7 is deliberate additional optimism. The papers show that this widening is necessary to recover dynamic regret of order

$\phi:\mathbb{R}^d\to\mathbb{R}$8

for both the tuned SWUCRL2-CW algorithm and the parameter-free BORL meta-algorithm (Cheung et al., 2019); (Cheung et al., 2020). In this case, “more optimism” is not a pathology but a structural correction for drifting transitions.

These RL usages collectively separate two notions. Uncertainty-aware, calibrated optimism is treated as beneficial. Wild optimism appears when optimistic mechanisms are insensitive to critic uncertainty, insensitive to aleatoric noise, or too tightly constrained by misleading empirical models.

3. Bandits, contextual bandits, and related optimism principles

Several papers in adjacent online-learning settings help clarify how wild optimism differs from more disciplined optimism principles.

In "Optimism in the Face of Ambiguity Principle for Multi-Armed Bandits" (Li et al., 2024), the optimistic object is not the reward estimate itself but an ambiguous perturbation law in an FTPL scheme. The central potential is

$\phi:\mathbb{R}^d\to\mathbb{R}$9

where $$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$0 is a marginal ambiguity set. The paper shows that this “optimism in the face of ambiguity” is equivalent to FTRL with a separable regularizer and yields best-of-both-worlds regret in adversarial and stochastic bandits (Li et al., 2024). This provides a contrast: optimism can be deliberately “more favorable” than a fixed perturbation law while remaining analytically controlled.

In "Upper Counterfactual Confidence Bounds: a New Optimism Principle for Contextual Bandits" (Xu et al., 2020), optimism is moved from action space to policy space. The UCCB principle uses policy-level confidence terms and implements them through counterfactual action trajectories at the current context. The resulting regret for finite $$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$1 and finite $$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$2 is

$$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$3

and the paper emphasizes that the complexity no longer scales with $$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$4 (Xu et al., 2020). This is not called wild optimism, but it is explicitly an optimism principle that becomes broader in scope while remaining controlled by contextual potential arguments.

In "Optimistic optimization of a Brownian" (Grill et al., 2019), optimism is interval-based and UCB-like. For each interval $$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$5, the algorithm uses

$$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$6

and repeatedly refines the interval with largest upper bound. The main theorem gives

$$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$7

for optimizing a Brownian motion on $$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$8 (Grill et al., 2019). This is another case in which highly aggressive optimistic search is mathematically safe because the confidence construction is tailored to the process.

A plausible synthesis is that “wild optimism” becomes problematic primarily when optimism is detached from the correct uncertainty geometry. The ambiguity-based FTPL construction, the policy-space UCCB construction, and Brownian optimistic refinement all enlarge the optimistic object while retaining a precise support-function, counterfactual, or Gaussian-process structure.

4. Games and strategic interaction: optimism as aggressive best-case reasoning

In strategic-interaction papers, optimism is formulated as a choice rule rather than an estimator. In "Optimism and Pessimism in Strategic Interactions under Ignorance" (Guarino et al., 2021), players hold coarse beliefs $$l(x,y)=D_\phi(x,y), \qquad D_\phi(x,y)=\phi(x)-\phi(y)-\langle \nabla \phi(y),x-y\rangle .$$9 and evaluate actions by either maxmax or maxmin. The optimistic best reply is

$\phi$0

while the pessimistic best reply is

$\phi$1

The paper identifies optimism with maxmax and pessimism with maxmin, and proves that common belief in optimism characterizes Point Rationalizability, whereas common belief in pessimism characterizes Wald Rationalizability (Guarino et al., 2021). In this setting, wild optimism corresponds to best-case focus under coarse beliefs, not to statistical overestimation.

A related but dynamical perspective appears in "Chaos, Extremism and Optimism: Volume Analysis of Learning in Games" (Cheung et al., 2020). There the issue is not a belief rule but optimistic multiplicative weights. In zero-sum games, OMWU contracts volume and stabilizes, while in coordination games the roles reverse and OMWU expands volume exponentially fast (Cheung et al., 2020). The paper’s “no-free-lunch” message is that optimistic updates can be stabilizing in adversarial environments and explosively unstable in cooperative ones. This suggests that in games, whether optimism is “wild” depends on the game class as much as on the update rule.

In matrix games with unknown payoff matrix and bandit feedback, "Randomised Optimism via Competitive Co-Evolution for Matrix Games with Bandit Feedback" (Lin, 19 May 2025) introduces CoEBL, which implements randomised optimism by mutating empirical payoffs: $\phi$2 The mean shift is UCB-like, but the optimistic perturbation is sampled rather than fixed. The paper proves worst-case Nash regret

$\phi$3

matching deterministic optimism-based methods (Lin, 19 May 2025). Here aggressive optimism is evolutionary and stochastic, but still formally controlled.

5. Economics and moral hazard: optimism as an up-shift in perceived output distributions

In "Optimism, overconfidence, and moral hazard" (Sinander, 2023), optimism is defined behaviorally and then identified structurally. The paper rewrites the moral-hazard problem in parsimonious form: $\phi$4 where $\phi$5 is the minimum cost of producing output distribution $\phi$6 (Sinander, 2023). Relative confidence is characterized by pointwise downward shifts in $\phi$7; relative optimism is characterized by $\phi$8 being up-shifted from $\phi$9 in the sense of first-order stochastic dominance comparisons.

In the paper’s binary-output example,

$\beta$0

decreasing $\beta$1 is a vertical downward shift corresponding to confidence, while increasing $\beta$2 is a horizontal shift making high probabilities of success relatively cheaper, corresponding to optimism (Sinander, 2023). The text explicitly interprets large horizontal shifts as “wild optimism”: the agent behaves as if very high-output distributions are cheap. This can make steep, performance-based contracts appear attractive even when the underlying technology would not support such outcomes.

The paper’s broader result is that optimism and overconfidence can be diagnosed from choice over contracts without invoking true probabilities at the definitional stage. A plausible implication is that wild optimism in this literature is a structural distortion in the cost of generating success, rather than a simple tendency to prefer risky contracts.

6. Evaluation pathologies: false optimism in medical machine learning

In "Meaningless comparisons lead to false optimism in medical machine learning" (Demasi et al., 2017), optimism is neither an estimator nor a choice rule. It is a field-level evaluation pathology created by weak baselines. The paper’s meta-analysis of sensor-based prediction of mental wellbeing reports that the bulk of the literature, approximately $\beta$3, uses comparisons that ignore patient baseline state (Demasi et al., 2017). Population baselines are reported in roughly half of publications, whereas personal baselines are reported in only about $\beta$4.

The paper’s central point is that for longitudinal personalized prediction, a personal baseline can already be very strong. In the authors’ words, it is possible to explain over $\beta$5 of the variance of some mood measures by simply guessing that each patient has their own average mood (Demasi et al., 2017). This motivates the proposed “user lift,” defined as personal-baseline error minus model error. Positive user lift indicates genuine gain over simply predicting each individual’s typical state; zero or negative user lift indicates no practical improvement.

This is a different but closely related meaning of wild optimism. The optimism is not in the model’s internal mechanism; it is in the interpretation of results. High apparent performance relative to a population baseline can coexist with no improvement over a patient-specific constant predictor. In this usage, wild optimism is baseless optimism induced by meaningless comparisons.

7. Cross-domain structure and recurring themes

Across these literatures, the phrase “wild optimism” consistently marks an asymmetry toward favorable possibilities, but the technical object being inflated varies sharply. In excess-risk theory it is a computable, refitting-based upper bound on a local complexity term (Hu et al., 2 Sep 2025). In deep RL it is excessive optimism in value estimation or exploration, especially when detached from uncertainty calibration (Tasdighi et al., 2024); (Liu et al., 2023). In drifting MDPs it is deliberate confidence widening, added because standard confidence sets are too narrow in a control-theoretic sense (Cheung et al., 2019). In strategic interaction it is maxmax reasoning under ignorance (Guarino et al., 2021). In moral hazard it is an up-shift in perceived feasibility of high-output distributions (Sinander, 2023). In medical ML it is the false appearance of progress generated by weak baselines (Demasi et al., 2017).

Two contrasts recur. First, some papers treat wild optimism as a pathology to be tamed, while others formalize it as a legitimate analytic tool. Second, the distinction between epistemic uncertainty and misleading variability appears repeatedly. The deep-RL and noisy-exploration papers make this distinction explicit; the medical-ML paper makes an analogous point through baselines; the excess-risk paper makes it through a refitting construction that replaces global class complexity by a local, computable quantity (Hu et al., 2 Sep 2025); (Liu et al., 2023); (Demasi et al., 2017).

This suggests a useful general reading. “Wild optimism” names a family resemblance rather than a single concept: a favorable distortion that becomes technically acceptable only when paired with the right geometry, the right benchmark, or the right control term. Where those are absent, the same asymmetry becomes a source of instability, over-exploration, or false inference.