Doubly-Adaptive Conformal Update

• Doubly-adaptive conformal update is a method that couples two distinct adaptive mechanisms to calibrate uncertainty in evolving, non-exchangeable data.
• It simultaneously adapts the target (e.g., pseudo-outcome or model representation) and the miscoverage threshold via online updates for robust inference.
• Applications include causal effect estimation, trajectory prediction, and multi-model selection, achieving improved interval efficiency and stable coverage.

Doubly-adaptive conformal update is a research pattern in conformal inference in which two distinct adaptive mechanisms are coupled within a single uncertainty-quantification procedure. In recent work, the first mechanism typically adapts the object being calibrated—such as a pseudo-outcome, a model, a representation space, a subset of calibration residuals, or an example-specific coverage policy—while the second mechanism adapts the conformal control itself, usually through an online update of a miscoverage level or threshold. The clearest explicit uses of this idea include doubly robust adaptive conformal inference for causal effects under temporal dependence, horizon-wide online calibration for trajectory ensembles, multi-model adaptive conformal prediction in dynamic environments, and spectral adaptive conformal prediction for structured non-exchangeable data (Koukorinis et al., 29 Jun 2026, Li et al., 18 Aug 2025, Hajihashemi et al., 2024, Opoku et al., 14 Jun 2026).

1. Conceptual scope and relation to adaptive conformal inference

Adaptive conformal inference (ACI) was introduced to maintain coverage under non-exchangeability by updating calibration online. A notable backdrop is that the finite-sample coverage guarantees associated with ACI do not require the use of conformal predictors: they hold for the broader class of confidence predictors, defined by the requirement of producing nested prediction sets, which was argued to be essential for meaningful confidence statements. In online settings, non-conformal confidence predictors were reported to offer significant computational advantages while maintaining comparable predictive efficiency, and in batch settings inductive non-conformal confidence predictors can outperform inductive conformal predictors by using the full training dataset without a separate calibration set, particularly when data are limited (Szabadváry et al., 2024).

Within that broader ACI lineage, “doubly-adaptive conformal update” is best understood as a methodological pattern rather than a single named algorithm. In one line of work, the two adaptivities are double robustness in the target construction and temporal adaptivity in the conformal calibration. In another, they are time-wise and horizon-wise threshold adaptation. Elsewhere they are calibration adaptation plus model selection, or task-adaptive representation learning plus local neighborhood-based calibration. This suggests that the term denotes a family of coupled update schemes whose common feature is simultaneous adaptation along two non-identical axes (Koukorinis et al., 29 Jun 2026, Li et al., 18 Aug 2025, Hajihashemi et al., 2024, Feng et al., 24 Feb 2026).

2. Core update law and two-axis architecture

A recurrent control law is the ACI recursion

$\alpha_{t+1} = \alpha_t + \gamma(\alpha - M_t),$

where $M_t$ is a miss indicator. In doubly robust adaptive conformal inference for causal effects, the update is written as

$$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$

with the interpretation that non-coverage decreases $\alpha_{t+1}$ and makes future intervals wider, whereas coverage increases $\alpha_{t+1}$ and allows future intervals to shrink. Spectral adaptive conformal prediction uses the same feedback form with $$M_t = \mathbf{1}\{Y_t \notin \widehat C_t^{\text{SA}}(X_t)\}$$, and retrospective online conformal inference likewise couples its interval construction to an ACI-type update of $\alpha_t$ (Koukorinis et al., 29 Jun 2026, Opoku et al., 14 Jun 2026, Jun et al., 6 Nov 2025).

The second axis of adaptation varies by application. In the trajectory-ensemble setting, standard ACI is generalized from a single threshold to an interval of admissible thresholds: $$I_t^h = [\alpha_t^h - \delta_t,\ \alpha_t^h + \delta_t],$$ with $\delta_t = o(t)$. For any $\alpha^{h*} \in I_t^h$, long-term coverage is preserved, which leaves a second-stage optimization free to choose horizon-specific thresholds that minimize a user-specified multi-step objective. In that construction, the online update adapts over time, and the optimization reallocates uncertainty across horizons at each time $M_t$0 (Li et al., 18 Aug 2025).

A comparable separation appears in limited-feedback conformal selection. There the ACI recursion is applied to a control parameter or dual variable, for example

$M_t$1

so that validity is enforced through the success process while an outer bandit or semi-bandit learner adapts the efficiency side of the problem. The control update and the structural learner therefore play distinct adaptive roles (Gollapudi et al., 14 May 2026).

3. Orthogonalized targets and doubly robust calibration

The most explicit formulation of a doubly-adaptive conformal update appears in doubly robust adaptive conformal inference for causal effects under temporal dependence. The data are a single temporally dependent observational sequence $M_t$2, and the target is the conditional average treatment effect

$M_t$3

Because $M_t$4 is latent, the conformal procedure calibrates a doubly robust pseudo-outcome

$M_t$5

Under correct nuisance specification, $M_t$6, and the leading bias is of product order $M_t$7, rather than a sum of nuisance errors. The paper describes this as “double robustness in dependent data,” obtained through $M_t$8-mixing assumptions and temporal block cross-fitting with guard bands (Koukorinis et al., 29 Jun 2026).

The conformal score is then

$M_t$9

or, in the variance-standardized version,

$$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$0

The corresponding intervals are centered at $$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$1, with radius determined by an empirical quantile at the current $$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$2. This is the second adaptivity: the interval width is adaptively calibrated online, while the calibrated quantity is itself adaptively debiased by the doubly robust construction. The paper explicitly characterizes this pairing as a “doubly-adaptive conformal update” in which “the value we calibrate” is adaptively debiased and “the interval width we choose” is adaptively calibrated (Koukorinis et al., 29 Jun 2026).

The main coverage theorem decomposes the finite-$$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$3 coverage gap into a mixing gap, a nuisance-bias tax, and an adaptation term: $$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$4 For pseudo-outcomes, the theorem gives long-run coverage, while for the latent CATE the intervals are asymptotically conservative under consistency of $$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$5. The simulations further report that variance-standardized DR-ACI achieves roughly 63% narrower intervals than split conformal at the same coverage, and in the non-stationary regime D it is the only method to maintain near-nominal coverage (~89.9%) and stable width (Koukorinis et al., 29 Jun 2026).

4. Structured calibration under dependence, trajectories, and retrospective adjustment

A second major instantiation appears in trajectory prediction. In adaptive conformal prediction over trajectory ensembles, the two adaptivities are over time and across forecast horizons. The method uses the probabilistic conformal prediction nonconformity score

$$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$6

so the prediction set at each horizon is a union of balls around ensemble members. This yields discontinuous prediction intervals around each trajectory and naturally captures temporal dependencies. Online ACI-style recursions are maintained separately for each horizon, producing admissible intervals $$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$7, and a horizon-wide optimization then selects $$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$8 to minimize a multi-step objective over coverage and interval width. The paper states that this is exactly the core idea of a doubly-adaptive conformal update: adapt thresholds over time and across forecast horizons or trajectory structure (Li et al., 18 Aug 2025).

In structured non-exchangeable time series, spectral adaptive conformal prediction uses local spectral similarity to adapt which calibration residuals matter: $$\alpha_{t+1} = \alpha_t + \gamma\big(\alpha - \mathbf{1}\{\psi_t^{\mathrm{DR}} \notin \hat C_t\}\big),$$9 These weights define a weighted empirical CDF and a weighted conformal quantile $\alpha_{t+1}$0, while the ACI recursion updates $\alpha_{t+1}$1 online. The two adaptive elements are therefore adaptive localization through $\alpha_{t+1}$2 and adaptive level through $\alpha_{t+1}$3. The approximate coverage result for the fixed weighted quantile depends on a spectral mismatch term $\alpha_{t+1}$4 and a concentration term $\alpha_{t+1}$5, while the adaptive update gives the deterministic identity

$\alpha_{t+1}$6

The paper emphasizes that spectral weighting must be monitored through effective sample size diagnostics, with

$\alpha_{t+1}$7

because overly aggressive localization can collapse coverage (Opoku et al., 14 Jun 2026).

Retrospective online conformal inference provides another variant in which the second adaptivity is not localization but score recomputation. The method uses kernel ridge regression in a moving window and efficiently recomputes all leave-one-out residuals

$\alpha_{t+1}$8

whenever a new point arrives. Prediction intervals are then formed by a Jackknife+ construction using these retrospectively adjusted residuals, while $\alpha_{t+1}$9 is updated by ACI, DtACI, SFOGD, or SAOCP. The paper’s claim is that existing online conformal methods adapt only forward in time, whereas retrospective adjustment aligns the entire set of residuals with the most recent data distribution and thereby achieves faster coverage recalibration and improved statistical efficiency (Jun et al., 6 Nov 2025).

5. Predictor, representation, and policy adaptation

A different interpretation of double adaptivity arises when the conformal layer is coupled to model selection. In multi-model ensemble conformal prediction for dynamic environments, each candidate model $\alpha_{t+1}$0 maintains its own adaptive miscoverage sequence $\alpha_{t+1}$1, updated by Scale-Free Online Gradient Descent on a pinball loss, while model weights $\alpha_{t+1}$2 are updated by multiplicative weights. In the dynamic setting, SAMOCP adds a second expert layer with staggered lifetimes and expert weights $\alpha_{t+1}$3. The resulting procedure adapts both the conformal calibration and the model or expert used to define the scores, and is proved to achieve strongly adaptive regret over all intervals while maintaining valid coverage (Hajihashemi et al., 2024).

DANCE supplies a representation-level version of the same template. It learns a task-specific kernel $\alpha_{t+1}$4 on top of pretrained embeddings by Recursive Feature Machine kernel ridge regression, then defines two local nonconformity scores in the learned geometry: a rank-based $\alpha_{t+1}$5-nearest-neighbor score $\alpha_{t+1}$6 and a contrastive density score $\alpha_{t+1}$7. Their calibrated sets are intersected,

$\alpha_{t+1}$8

with the total miscoverage budget split as $\alpha_{t+1}$9 and $$M_t = \mathbf{1}\{Y_t \notin \widehat C_t^{\text{SA}}(X_t)\}$$0. The paper describes DANCE as a doubly adaptive neighborhood conformal algorithm because it adapts both the representation space used for local geometry and the local calibration or nonconformity scoring performed within that space (Feng et al., 24 Feb 2026).

Adaptive coverage policies move the second axis from model or representation to the coverage level itself. In that setting, the miscoverage is chosen per example as

$$M_t = \mathbf{1}\{Y_t \notin \widehat C_t^{\text{SA}}(X_t)\}$$1

and conformal e-prediction with soft-rank e-values preserves a post-hoc guarantee even when $$M_t = \mathbf{1}\{Y_t \notin \widehat C_t^{\text{SA}}(X_t)\}$$2 depends arbitrarily on the data. The policy $$M_t = \mathbf{1}\{Y_t \notin \widehat C_t^{\text{SA}}(X_t)\}$$3 is learned by leave-one-out pseudo-episodes, so the system adapts coverage per example and adapts the mapping from scores to coverage through training. The paper presents this as a route to varying the coverage level and the resulting prediction set size with the difficulty of each individual example while maintaining valid statistical guarantees (Gauthier et al., 5 Oct 2025).

A broader operational generalization appears in efficient online conformal selection with limited feedback. There, the ACI update is applied to a threshold, a probing budget, or a dual variable rather than only to a quantile parameter, while a bandit or semi-bandit module adapts the efficiency side of the selection. The main claim is that the simple ACI update rule, when applied to the appropriate control parameter or dual variable, is adversarially valid and stochastically efficient under canonical limited-feedback models (Gollapudi et al., 14 May 2026).

6. Guarantees, diagnostics, and limitations

The guarantees attached to doubly-adaptive conformal updates are heterogeneous. Some are exact finite-sample split-conformal guarantees under exchangeability, as in DANCE and conformal e-prediction with adaptive coverage policies. Some are long-run average guarantees under arbitrary or non-exchangeable sequences, as in ACI-style online methods, multi-model dynamic conformal prediction, and limited-feedback conformal selection. Others are approximate local-in-time statements, as in spectral weighting under spectral smoothness and effective sample size control (Feng et al., 24 Feb 2026, Gauthier et al., 5 Oct 2025, Hajihashemi et al., 2024, Gollapudi et al., 14 May 2026, Opoku et al., 14 Jun 2026).

This diversity creates a recurring misconception: double adaptivity does not imply a single kind of validity statement. In DR-ACI, exact finite-sample, distribution-free control is for doubly robust pseudo-outcomes, whereas latent CATE containment is asymptotically conservative under consistency of $$M_t = \mathbf{1}\{Y_t \notin \widehat C_t^{\text{SA}}(X_t)\}$$4. In CP-Traj, the guarantee is long-term coverage for each horizon, not a joint pathwise guarantee over the entire trajectory. In spectral ACI, the fixed weighted quantile has an approximate coverage bound and the online update gives deterministic long-run calibration. In adaptive coverage policies, the guarantee is marginal in an e-value sense and explicitly permits data-dependent $$M_t = \mathbf{1}\{Y_t \notin \widehat C_t^{\text{SA}}(X_t)\}$$5, but it is not conditional coverage (Koukorinis et al., 29 Jun 2026, Li et al., 18 Aug 2025, Opoku et al., 14 Jun 2026, Gauthier et al., 5 Oct 2025).

Another misconception is that more localization or more structure-aware adaptation is automatically beneficial. Spectral ACI shows that localization must be monitored by effective sample size diagnostics; in the U.S. utilities example, very small median $$M_t = \mathbf{1}\{Y_t \notin \widehat C_t^{\text{SA}}(X_t)\}$$6 was associated with severe undercoverage. DR-ACI states that strict stationarity excludes structural breaks and long-range dependence, and its good performance in drift regimes is empirical rather than theorem-covered. Retrospective adjustment improves adaptation speed, but its current construction is tailored to regression with real-valued labels and linear smoothers admitting efficient leave-one-out formulas. Multi-model and expert aggregation improve efficiency in dynamic environments, but their coverage theorem is asymptotic and in expectation. These points indicate that the second adaptive layer frequently introduces a new bias-variance or validity-efficiency tradeoff that must itself be controlled (Opoku et al., 14 Jun 2026, Koukorinis et al., 29 Jun 2026, Jun et al., 6 Nov 2025, Hajihashemi et al., 2024).

Taken together, the recent literature supports a precise but non-uniform usage of the term. A doubly-adaptive conformal update is not a canonical algorithmic primitive; it is a compositional design in which conformal calibration is coupled to a second adaptive mechanism that changes what is being calibrated, how calibration data are weighted, which model supplies the score, how thresholds are allocated across tasks or horizons, or how coverage is individualized. The principal open directions stated across these works include sharper dependence bounds, non-stationary theory beyond empirical success, alternative dependence-aware splitting schemes, multiple doubly robust layers, smoother or regret-minimizing ACI variants, adaptive bandwidth selection for spectral localization, and broader extensions to structured outputs and limited-feedback decision problems (Koukorinis et al., 29 Jun 2026, Opoku et al., 14 Jun 2026, Li et al., 18 Aug 2025, Gollapudi et al., 14 May 2026).