When deploying machine learning models in risk-sensitive domains like clinical forecasting or financial valuation, knowing how much to trust a prediction is as important as the prediction itself. Standard Conformal Prediction (CP) has emerged as a premier framework for this, providing distribution-free prediction intervals with rigorous mathematical guarantees.
However, standard CP offers a marginal coverage guarantee: it promises that, on average across the entire data distribution, the true value will fall within the predicted interval $95\%$ of the time. The fatal flaw of marginal coverage is that it permits a model to be wildly overconfident on difficult subgroups as long as it compensates by being overly conservative on easy ones.
To solve this, a rapidly growing body of research focuses on localized calibration and conditional coverage optimization. The goal is simple yet challenging: ensure that the $95\%$ coverage guarantee holds not just on average, but locally for every individual input neighborhood.
The Core Problem: The Illusion of Average Safety
Imagine a medical model predicting patient survival times. If the data is heteroscedastic—meaning the noise level or uncertainty varies across the input space—a global calibration strategy fails. It computes a single global quantile of past errors and uses it to size all future prediction intervals.
As a result, for a high-risk patient subgroup with volatile outcomes, the global interval will be too narrow, failing to meet the safety threshold. Conversely, for a stable, predictable subgroup, the interval will be unnecessarily wide, destroying its practical utility.
True conditional coverage—guaranteeing valid uncertainty bounds for every exact point in the input space—is theoretically impossible to achieve in finite samples without restrictive parametric assumptions. Localized conformal prediction bypasses this impossibility by targeting approximate conditional validity. It achieves this by shifting the calibration mechanism from a global pool to a localized, data-dependent neighborhood.
Mechanism 1: Distance and Kernel-Based Localization
The most intuitive way to localize uncertainty estimation is to weight historical calibration data based on its proximity to the current test instance. Instead of treating every historical error equally, the system constructs a localized empirical distribution of errors.
[Global Calibration] --> Uses all calibration points equally --> One-size-fits-all interval
[Localized Calibration] --> Applies distance/kernel weights --> Input-adaptive interval
Lin et al. (2021) formalized this with Locally Valid and Discriminative (LVD) prediction intervals. When a new test point arrives, LVD calculates a weighted quantile of historical errors, assigning higher weights to calibration samples that lie in the immediate neighborhood of the test point. This prevents models from being overconfident on difficult subgroups while maintaining high sharpness (narrow intervals) on easier cases. As Lin et al. (2021) and Yan (2023) demonstrate, this neighborhood-aware approach consistently maintains coverage guarantees even for high-error tail cases where global methods collapse.
Researchers define these neighborhoods using several distinct geometries:
- Feature-Space Proximity: Wu et al. (2025) utilize a kernel function to weight calibration data points based on their proximity in the feature space, allowing the uncertainty quantification to adapt dynamically to local data density and heteroscedasticity.
- Geographic Proximity: In domains like real estate or climate modeling, feature distance is secondary to physical distance. Hjort et al. (2025) introduced spatially weighted conformal prediction, which applies a kernel function to prioritize geographically proximate calibration data, successfully mitigating local coverage gaps caused by regional market volatility.
- Geometric Data Structures: For complex, multivariate outputs, Thurin et al. (2025) leverage a $k$-nearest neighbors approach on the calibration set to allow prediction regions to adjust both their size and non-convex shape based on the local input context.
Mechanism 2: Model-Driven and Structural Neighborhoods
Defining neighborhoods purely by Euclidean distance in raw feature space can fail when dealing with high-dimensional data or complex, non-linear relationships. To counter this, several frameworks derive localized neighborhoods directly from the internal representations or structural properties of the underlying machine learning models.
Mixture-of-Experts Gating
In heterogeneous data environments, data points might belong to hidden subpopulations. Kong et al. (2026) developed a framework that treats the gating vectors of a Mixture-of-Experts (MoE) model as soft domain assignments. By localizing uncertainty estimates through these model-internal representations, they ensure that calibration tracks the specific variance of distinct data segments. Similarly, Wong et al. (2025) demonstrated that using data-dependent weights within predictive ensembles enhances local validity across diverse data slices, preventing systematic under-performance on challenging subsets.
Error-Pattern Clustering
Rather than grouping data by input features, Rakhshaninejad et al. (2025) proposed cluster-conditioned conformal prediction (CCP). This approach groups instances based on their prediction error patterns. By performing dynamic neighborhood calibration against reference points that share similar model vulnerabilities, the resulting uncertainty intervals remain valid even when encountering entirely new categories or out-of-distribution inputs.
Tree-Based Partitioning
Instead of post-hoc weighting, localization can be baked directly into the model structure. Johansson et al. (2023) introduced Conformal Predictive Distribution Trees (CPDTs), which embed conformal distributions directly into the leaves of a regression tree. This provides a localized calibration framework where distinct categories or regions within the tree structure receive granular, mathematically guaranteed uncertainty estimates.
Mechanism 3: Score Rectification and Density Estimation
An alternative strategy to localized weighting is score rectification. Instead of changing how the quantile is calculated from the calibration set, these methods learn a secondary model to map or "rectify" standard conformity scores so that they inherently reflect local difficulty.
[Standard Score] --> [Rectification Function (Learns Local Variance)] --> [Rectified Score]
|
Produces Locally Valid
Prediction Interval
Plassier et al. (2025) developed Rectified Conformal Prediction (RCP) to address this. RCP adjusts prediction set sizes by learning to transform conformity scores using conditional quantile estimation. By adjusting the scores based on a local estimate of model uncertainty, the framework prevents under-coverage in difficult regions and over-coverage in easy ones. In a parallel direction, Plassier et al. (2025) introduced Probabilistic Conformal Prediction ($CP^2$), which adapts prediction sets to the local behavior of an estimated data distribution, proving highly effective in multimodal or highly heteroscedastic settings.
This aligns closely with the non-parametric approach proposed by Luo et al. (2022), who introduced Local Calibration Error (LCE) to measure reliability within soft neighborhoods and developed a recalibration method (LoRe) that corrects localized biases. By moving beyond global fixes, Kuleshov et al. (2022) showed that framing recalibration as a low-dimensional density estimation problem allows intervals to expand where local uncertainty is high while maintaining sharpness where the model is confident.
Why It Works: The Intuition of Shifting Error Distributions
To understand why localized conformal prediction works, visualize the calibration errors as a landscape of varying heights.
Standard conformal prediction flattens this entire landscape into a single, global histogram. It finds the 95th percentile of that combined histogram and draws a uniform boundary. If the test point happens to sit in a deep valley (a low-variance region), the boundary is far too wide. If the test point sits on a high peak (a high-variance region), the boundary is too narrow, and the true value escapes.
Localized conformal prediction acts as a dynamic spatial filter. By applying a kernel function (Wu et al. (2025)), a spatial weight (Hjort et al. (2025)), or a model-derived gating weight (Kong et al. (2026)), it isolates only the slice of the error landscape immediately surrounding the test point.
- In a high-variance neighborhood, the localized error distribution stretches out. The calculated quantile naturally pushes outward, generating a wider, safer interval that preserves the coverage guarantee.
- In a low-variance neighborhood, the localized error distribution compresses. The calculated quantile pulls inward, generating a sharp, highly precise interval.
By allowing the empirical error distribution to stretch, shrink, and shift depending on the local context, these methods guarantee that the safety margins remain honest, adaptive, and reliable for every individual prediction.
Key Foundational Work
Deep Kernel Survival Analysis and Subject-Specific Survival Time Prediction Intervals (2020) — contributed 2 concept elements to this direction.
Local and global sparse Gaussian process approximations (2007) — contributed 1 concept elements to this direction.