In high-stakes environments like medical diagnosis or credit scoring, a machine learning model that guesses when it is uncertain is a liability. To mitigate this, researchers use selective classification (or "learning to reject"), which allows a model to abstain from making a prediction when its confidence falls below a certain threshold.
However, a simple "I don't know" is often insufficient for human experts. If a model refuses to diagnose a patient, the clinician needs to know why: Is the case an outlier the model hasn't seen before? Or are the symptoms so ambiguous that they overlap with multiple conditions? Andrea Pugnana argues that without explainable rejection criteria, users cannot distinguish between data gaps and internal algorithmic failures, undermining the trust required for human-AI collaboration Topics in Selective Classification.
The emerging research direction focuses on transforming this "black-box" rejection into an informative, transparent signal.
Decomposing the Source of Silence
The first step in explaining a rejection is identifying the type of uncertainty. Johannes Brinkrolf and Barbara Hammer propose a geometric framework to distinguish between two fundamental causes of abstention Causes of Rejects in Prototype-based Classification Aleatoric vs. Epistemic Uncertainty:
- Aleatoric Uncertainty (Ambiguity): The input lies on the boundary between two classes. The model "knows" the domain but finds the specific instance inherently confusing.
- Epistemic Uncertainty (Novelty): The input is an outlier or "out-of-distribution." The model recognizes that it lacks the training data to make a competent judgment.
By communicating this distinction, a system can guide a human expert to either look for more discriminating features (in the case of ambiguity) or treat the sample as a rare, novel case (in the case of novelty).
Mechanisms for Transparent Abstention
Researchers have developed several architectural patterns to make these rejections interpretable.
Set-Based Explanations
Instead of a binary "predict or reject" signal, the SCRIB framework developed by Zhen Lin, Cao Xiao, and Jimeng Sun 0001 outputs a set of plausible labels SCRIB: Set-classifier with Class-specific Risk Bounds for Blackbox Models. If the model returns {Arrhythmia, Normal}, it effectively explains its rejection by showing exactly which classes it is confusing. This provides structured feedback that helps experts focus their manual review.
Audit Models and Feature Attribution
Another approach involves training a secondary "audit" classifier specifically to predict when the primary model will abstain. Carlos Mougan and Dan Saattrup Nielsen demonstrate that by applying feature attribution methods like SHAP to this audit model, we can identify which specific input features triggered the decision to withhold a prediction Model Agnostic Explainable Selective Regression via Uncertainty Estimation.
Counterfactual Rejections
The L2loRe framework, introduced by Clara Punzi and Roberto Pellungrini, uses counterfactuals to explain the "distance to certainty" L2loRe: a method for explaining the reject option. It generates human-readable explanations by showing the minimal feature modifications required to move a rejected sample back into a region where the model would feel confident.
Reasoning and Knowledge Gaps
In the context of Large Language Models (LLMs), abstention is often a tool to prevent hallucinations. Hyuhng Joon Kim and colleagues developed Contrastive Decoding with Abstention (CDA) to help models stay silent when they lack parametric knowledge When to Speak, When to Abstain: Contrastive Decoding with Abstention.
To make these refusals more than just generic "I don't know" messages, Junfeng Liao and Ling Chen propose Targeted Reasoning Unlearning (TRU), which trains models to generate internal reasoning traces before a refusal Explainable LLM Unlearning through Reasoning. This "internal monologue" allows practitioners to verify that the model is identifying sensitive or missing information correctly rather than failing randomly. Similarly, the KnowGuard framework by Xilin Dang generates transparency reports that detail specific evidence gaps in medical knowledge graphs, explaining exactly what information was missing to reach a diagnosis KnowGuard: Knowledge-Driven Abstention for Multi-Round Clinical Reasoning.
Rejecting Based on Explanation Quality
A final, critical facet of this research is the idea that a model should abstain not just when it is uncertain about the label, but when it cannot provide a trustworthy reason.
Luca Stradiotti and Stefano Teso introduce LtX (Learning to Reject Low-Quality Explanations), which uses human feedback to train a rejector that evaluates the quality of the model's underlying reasoning Learning to Reject Low-Quality Explanations via User Feedback. This ensures that even if a model is statistically confident in its prediction, it will still abstain if its explanation is nonsensical or misleading—a vital safeguard against "right for the wrong reasons" scenarios.