← News

ResearchSources: arXiv · 18 Nov 2026

Formal XAI paper links counterfactual explanations to abductive and contrastive reasoning

A new arXiv paper proposes conjunctive counterfactual explanations, connecting closest-world counterfactuals with the abductive and contrastive explanations used in formal explainability, and reports efficient extraction on benchmarks.

In plain words

When an AI model makes a prediction, counterfactual explanations describe what would need to change for a different outcome. The paper defines a new restricted form of these explanations, built only from conjunctions of feature conditions, that provably changes the prediction. The authors show it connects to existing explanation types and can be computed with current tools.

What was announced

The paper, posted to arXiv, addresses a gap between two views of counterfactual explanations. One view treats them as counterfactual conditionals under closest-world semantics, which requires reasoning over all minimally distant counterexamples; the other, formal explainable AI, relies on feature-based abductive and contrastive explanations that are efficient to compute and linked by a minimal hitting-set duality. To bridge the two, the authors characterize closest-world counterfactual explanations as a disjunctive normal form over all minimally distant counterexample instances, and introduce conjunctive counterfactual explanations (CFXps), a subclass of counterfactual antecedents restricted to conjunctions of feature literals that guarantee a prediction change.

The authors establish structural connections among the three explanation types: every contrastive explanation induces an abductive explanation for its witness instance, and every contrastive explanation combined with a witness abductive explanation induces a CFXp whose support is determined by that abductive explanation. They say these results enable a modular way to compute CFXps using existing abductive and contrastive machinery. In experiments on a representative set of benchmarks with tree ensembles and neural networks, CFXp extraction was practically efficient, and compared with popular point-based counterfactual explainers, CFXps generally required no more feature changes while guaranteeing the counterfactual prediction beyond a single instance.

What it does not show

  • The source is only the paper's abstract, so the full paper's methods, proofs and experimental detail are not covered here.
  • The benchmark results are the authors' own and have not been independently confirmed.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.