Contrastive Reasons

• Contrastive reasons are explanations that distinguish a fact from a salient foil by highlighting the unique difference-making conditions.
• They are formalized through logical decompositions and operationalized in AI models to separate shared from discriminative features.
• Their applications range from explainable AI and argument mining to decision support systems, addressing challenges like foil selection and computational complexity.

Contrastive reasons are reasons given relative to a foil: they answer not merely why some fact $P$ obtained, but why $P$ obtained rather than some alternative $Q$. Across explainable AI, formal argumentation, description logics, retrieval-augmented generation, rule-based automation, and argument mining, the common structure is a decomposition into a fact, a foil, and the difference-making conditions that support the fact while excluding, defeating, or failing to realize the foil (Jacovi et al., 2021, Mahmood et al., 2 May 2026, Geibinger et al., 11 Jul 2025).

1. Conceptual structure

The basic contrastive schema is the question “Why $P$ rather than $Q$?” In this schema, $P$ is the fact and $Q$ is the foil. The foil is not an arbitrary negation; it is a salient alternative, often tied to user expectation, surprise, or a nearby competing classification. This is why contrastive reasons are typically more selective than non-contrastive reasons: they suppress causes that are common to both fact and foil and retain those that distinguish the two (Jacovi et al., 2021, Mahmood et al., 2 May 2026).

A recurring theme is that contrastive reasons are user-centered. In rule-based smart environments, the foil is the alternative event the user likely expected; in retrieval-augmented generation it is the set of retrieved passages that should be discounted; in formal argumentation it may be an explicit rival argument or an implicit attacker; and in description logics it may be either another individual instantiating the same concept or the same individual instantiating a rival concept (Herbold et al., 2024, Ranaldi et al., 2024, Borg et al., 2021, Mahmood et al., 2 May 2026).

The term also has a broader use in argument mining. In online polarized debate, contrastive reasons are short sentential excerpts that express premises, arguments, rebuttals, or other persuasive content from divergent viewpoints, with the additional requirement that the reason should allow a reader to infer the writer’s viewpoint. In that usage, the contrast is organized across opposing camps rather than around a single prediction or entailment (Trabelsi et al., 2019).

2. Formal models and logical structure

A common logical pattern is to separate commonality from difference. In a propositional framework, several canonical contrastive explanation problems are defined by outputs of the form $(\theta,\theta',\chi)$, where $\theta$ is the fact-specific part, $\theta'$ is the foil-specific part, and $P$0 is shared context. The optimization criterion minimizes $P$1 and then maximizes $P$2, forcing common information into $P$3 and leaving $P$4 as the discriminative residue. In the counterfactual difference problem, if the input $P$5 defines an assignment, the output corresponds to a cardinality-minimal set of literal flips leading from $P$6 to a foil assignment satisfying $P$7 (Geibinger et al., 11 Jul 2025).

In description logics, a contrastive explanation problem can be written as $P$8, where $P$9 and $Q$0. The explanation is a tuple $Q$1, where $Q$2 captures what fact and foil already share, $Q$3 captures what the fact has and the foil lacks, and $Q$4 is a minimal conflict set needed when the foil-side completion is inconsistent with the current ABox. A related formulation treats facts as $Q$5 and allows foils of the form $Q$6 or $Q$7, thereby distinguishing entity contrast from concept contrast (Koopmann et al., 14 Nov 2025, Mahmood et al., 2 May 2026).

Formal argumentation gives a graph-theoretic version of the same idea. If $Q$8 is accepted and $Q$9 is not, then contrastive reasons are defined through the intersection of the explanation for $P$0’s acceptance with the explanation for $P$1’s non-acceptance. Using defenders $P$2 and undefeated attackers $P$3, the contrastive explanation isolates arguments that simultaneously support the fact and block the foil. In structured argumentation, the same mechanism can be lifted from abstract arguments to premises or subarguments, making the contrastive reason more explicit at the level of domain propositions (Borg et al., 2021).

These logical frameworks also show that contrastive reasoning is not merely the juxtaposition of “why $P$4?” and “why not $P$5?” A separate justification for $P$6 and a separate abductive repair for $P$7 can be formally correct yet contrastively poor if they use unrelated derivational routes. This is why several frameworks enforce one shared pattern plus a difference component rather than two independent explanations (Koopmann et al., 14 Nov 2025, Mahmood et al., 2 May 2026).

3. Contrastive reasons in model interpretability

In neural classification, contrastive reasons are often defined relative to a target label $P$8 and an explicit foil label $P$9. A central formulation uses a classifier $Q$0, where $Q$1 is a latent representation and $Q$2 is the final linear layer. The pairwise discriminative direction is

$Q$3

Projecting the representation onto that direction,

$Q$4

removes latent components that support fact and foil equally and retains only what differentiates $Q$5 from $Q$6. Factor importance is then evaluated by interventions on tokens, spans, or concepts and scored by how much the intervention changes the normalized fact-versus-foil preference (Jacovi et al., 2021).

A related operationalization in text classification uses the contrastive target quantity

$Q$7

where $Q$8 is the logit of the predicted label and $Q$9 is the logit of the foil. Post-hoc methods such as Layer-wise Relevance Propagation, Gradient $P$0 Input, and Gradient Norm can then attribute the margin between fact and foil rather than the fact alone. In this setting, contrastive explanations are explicitly answers to questions such as “Why class $P$1 rather than class $P$2?” (Eberle et al., 2023).

For deep convolutional image classifiers, one model-intrusive approach represents contrastive reasons with internal filters rather than input perturbations. The method identifies Minimum Correct (MC) filters, a sparse set of top-layer filters sufficient to preserve the original class, and Minimum Incorrect (MI) filters, a sparse nonnegative set of additive activation changes sufficient to obtain a specified alternative class. The contrastive question “Why class $P$3 rather than class $P$4?” is then answered by the concepts encoded in MC filters for $P$5 together with the missing or weak MI concepts for $P$6 (Tariq et al., 12 Jan 2025).

These formulations share a common assumption: contrastive reasons are not the total causal basis of a prediction, but the subset of features, concepts, or latent directions that are useful for the fact and against the foil. This makes them more discriminative than ordinary saliency maps or label-agnostic feature attributions (Jacovi et al., 2021, Tariq et al., 12 Jan 2025).

4. Operational systems and application domains

Several systems turn contrastive reasons into interactive or task-specific procedures. REASONX does this for decision trees and tree surrogates using Constraint Logic Programming. It computes factual decision rules, contrastive decision rules, and closest contrastive examples, while allowing background knowledge, feature immutability, and under-specified user profiles to be represented as linear constraints. In that framework, a contrastive reason can be a rule, a region, or a nearest feasible counterexample, and the explanation process is explicitly dialogue-like and iterative (State et al., 2023).

In robotic planning over Markov decision processes, contrastive explanations are generated for action choices of the form “Why take action $P$7 rather than $P$8?” The formalization uses three factors: selectiveness, constrictiveness, and responsibility. A state is critical when the spread in action impact exceeds a threshold, responsibility is based on the relative impact $P$9, and constrictiveness measures how many future critical decision points remain after taking an action. The resulting reasons are comparative performance reasons such as “because it leads to the shortest route” or “because it leads to the most flexible future route” (Chen et al., 2020).

In retrieval-augmented LLMs, contrastive reasons are explanations of why some retrieved passages should be trusted and others discounted. Contrastive-RAG partitions retrieved documents into a relevant set $Q$0 and an irrelevant set $Q$1, then generates an explanation $Q$2 that states why $Q$3 is relevant to the query rather than $Q$4. The explanation is organized around passage-to-passage comparison and then used to support the final answer (Ranaldi et al., 2024).

In rule-based smart environments, the main technical problem is foil prediction. The system first classifies the situation into one of three confusing cases: an action occurred but another action was expected, no action occurred but an action was expected, or an error occurred but an action was expected. In the first two cases, candidate expected rules are ranked with TOPSIS using precondition similarity, ownership, frequency, and explanation occurrence; in the first case, precondition similarity is the Jaccard similarity

$Q$5

The selected expected rule supplies the foil, and the explanation then contrasts the actual or missing action with that user-expected alternative (Herbold et al., 2024).

Outside explanation proper, contrastive reasons are also operationalized for debate analysis. A pipeline built around the Phrase Author Interaction Topic-Viewpoint model detects and clusters phrase-level reasons from opposing viewpoints, producing a digest in which rows correspond to recurring facets and columns correspond to divergent camps. Here the contrast is viewpoint-relative rather than fact-versus-foil for a single decision, but the organizing principle is again the differential expression of reasons across alternatives (Trabelsi et al., 2019).

5. Human evidence, evaluation, and controversy

A central empirical controversy is whether humans actually explain in a contrastive manner. On four English text-classification datasets, with RoBERTa, GPT-2, and T5 models and post-hoc explainers, one study found that model-based explanations computed in contrastive and non-contrastive settings align equally well with human rationales. In the BIOS annotation study, cross-setting agreement between human contrastive and non-contrastive rationales was $Q$6 overall in Cohen’s Kappa, compared with $Q$7 within the non-contrastive setting and $Q$8 within the contrastive setting; humans also selected fewer words in the contrastive condition, $Q$9 on average versus $(\theta,\theta',\chi)$0, with lower sentence-level entropy $(\theta,\theta',\chi)$1 versus $(\theta,\theta',\chi)$2. The paper’s summary is that “humans do not necessarily explain in a contrastive manner” (Eberle et al., 2023).

A similar caution appears in reinforcement learning. In a user study on global explanations of RL policies, complete explanations were generally more effective when they were the same size as, or smaller than, contrastive explanations, and they were no worse when they were larger. The study therefore concluded that contrastive explanations are not sufficient to solve the problem of effectively explaining RL policies and require additional careful study for use in this context (Narayanan et al., 2022).

By contrast, in robotic planning, a user study with $(\theta,\theta',\chi)$3 Amazon Mechanical Turk participants found that explanations using responsibility and constrictiveness can improve understanding and trust relative to minimal rule tracing, while selectiveness reduces cognitive burden but may also reduce perceived completeness. This suggests that the usefulness of contrastive reasons depends strongly on representation format, task, and the cognitive cost of maintaining the foil (Chen et al., 2020).

Taken together, these results undermine a blanket claim that contrastive explanations are automatically more human-like or more helpful. They can be more selective, more class-specific, or more useful in high-conflict settings, but those gains are not uniform across domains (Eberle et al., 2023, Narayanan et al., 2022).

6. Scope, limitations, and open problems

One recurring limitation is foil selection. In some domains, such as rule-based automation or retrieval-augmented generation, the foil must be predicted from context; in others, such as argumentation, it can be extracted from attack structure; in description logics it may be either an alternative individual or an alternative concept. This suggests that contrastive reasons are only as informative as the foil is appropriate (Herbold et al., 2024, Borg et al., 2021, Mahmood et al., 2 May 2026).

Another limitation is that contrastive explanations are often post hoc. In NLP rationale studies, highlighted tokens are proxies for human explanation rather than direct observations of internal reasoning. In surrogate-based systems such as REASONX, explanation quality depends on surrogate fidelity. In neural contrastive attribution, the contrastive reason is tied to model-internal decision geometry, not necessarily to world-level causation (Eberle et al., 2023, State et al., 2023, Jacovi et al., 2021).

Formal frameworks expose additional computational limits. In description logics, subset difference-minimal contrastive explanations can be comparatively manageable, but conflict-minimal explanations are much harder: with fresh individuals, verification is $(\theta,\theta',\chi)$4-complete for $(\theta,\theta',\chi)$5 and $(\theta,\theta',\chi)$6-complete for $(\theta,\theta',\chi)$7 and $(\theta,\theta',\chi)$8. Commonality-maximal and cardinality-optimal variants are also intractable in expressive settings. This suggests that the most intuitively appealing notion of contrastive reason is not always the most computationally accessible (Koopmann et al., 14 Nov 2025).

Open problems therefore cluster around three themes. The first is better foil construction, including multi-foil or user-adaptive foils. The second is richer structural representations of commonality and difference, especially beyond pairwise or single-label settings. The third is evaluation: current evidence shows that contrastive reasons can be more focused, but not necessarily more faithful, more human-like, or more effective in every interface. Future work in description logics, logic-based explanation, and interactive XAI explicitly points toward quantified patterns, additional optimality criteria, improved computation methods, and broader empirical validation (Mahmood et al., 2 May 2026, Geibinger et al., 11 Jul 2025, Koopmann et al., 14 Nov 2025).

Contrastive reasons thus occupy a distinctive position in contemporary explanation research. They are not merely reasons for truth, reasons for falsity, or counterfactual repairs in isolation. They are reasons organized around a salient alternative, with the explicit aim of isolating the difference between fact and foil. Whether formulated as $(\theta,\theta',\chi)$9, as $\theta$0 and $\theta$1, as defenders intersecting undefeated attackers, or as discriminative latent directions and internal concepts, the unifying principle is the same: explanatory adequacy is measured not by how much support one can enumerate for a fact, but by how precisely one can say why this, and not that.