CREDO: Epistemic-Aware Conformalized Credal Envelopes for Regression

2026-03-10 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, extended

Summary

CREDO introduces an epistemic-aware conformal prediction method for regression that generates prediction intervals with distribution-free coverage guarantees while explicitly representing epistemic uncertainty. This approach, termed "credal-then-conformalize," first constructs an interpretable credal envelope that expands in regions with weak local evidence. It then applies split conformal calibration to this envelope, ensuring marginal coverage without additional assumptions. The resulting prediction intervals are interpretable, allowing their width to be decomposed into aleatoric noise, epistemic inflation, and a distribution-free calibration slack. CREDO's implementation uses endpoint-trimmed posterior predictive distributions and can adapt trimming levels based on local data density. Experiments on 12 regression benchmarks show CREDO maintains target 90% marginal coverage, improves sparsity adaptivity, and offers competitive efficiency compared to state-of-the-art conformal methods like CQR and UACQR-S/P.

Key takeaway

For AI Scientists and Research Scientists developing robust uncertainty quantification methods, CREDO offers a principled way to integrate epistemic awareness into conformal prediction. You should consider adopting CREDO to generate prediction intervals that not only guarantee marginal coverage but also provide a clear, decomposable understanding of uncertainty sources, especially in data-sparse or extrapolative regions. This can lead to more reliable and diagnostically rich models.

Key insights

CREDO combines credal sets and conformal prediction to deliver calibrated, interpretable prediction intervals that explicitly quantify epistemic uncertainty.

Principles

• Separate epistemic modeling from conformal calibration.
• Credal envelopes widen with weak local evidence.
• Interval width decomposes into aleatoric, epistemic, and calibration components.

Method

CREDO constructs a local credal set, derives a credal quantile envelope, and then applies split conformal calibration using a distance-to-envelope score to guarantee marginal coverage.

In practice

• Use endpoint-trimmed posterior credal sets for epistemic uncertainty.
• Implement adaptive trimming levels based on kNN scarcity scores.
• Decompose interval width for diagnostic insights into uncertainty sources.

Topics

• Conformal Prediction
• Epistemic Uncertainty
• Credal Sets
• Regression Analysis
• Uncertainty Quantification

Code references

• Monoxido45/CREDO

Best for: AI Researcher, AI Scientist, Research Scientist

Related on AIssential

• Learning Credal Ensembles via Distributionally Robust Optimization — CreDRO quantifies epistemic uncertainty by modeling disagreement from potential train-test distribution shifts via distributionally robust optimization.
• Online Conformal Prediction: Enforcing monotonicity via Online Optimization — New online conformal prediction methods ensure nested prediction sets across multiple confidence levels.
• Concerning Uncertainty -- A Systematic Survey of Uncertainty-Aware XAI — UAXAI integrates uncertainty into AI explanations to improve transparency, reliability, and human trust in high-stakes decision-making.
• ReSS: Learning Reasoning Models for Tabular Data Prediction via Symbolic Scaffold — ReSS combines decision trees and LLMs to generate faithful, explainable reasoning for tabular data predictions.
• Introduction to Deep Evidential Regression for Uncertainty Quantification — Evidential Deep Learning quantifies both epistemic and aleatoric uncertainty in a single inference pass.
• Uncertainty Quantification for Multimodal Large Language Models with Incoherence-adjusted Semantic Volume — UMPIRE quantifies MLLM uncertainty by measuring semantic volume and response incoherence without additional training or external tools.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.