Self-Generated Counterfactual Explanations

• Self-Generated Counterfactual Explanations (SCEs) are methods that autonomously generate minimal input modifications to flip a model's prediction, ensuring both validity and minimality.
• They leverage internal model capacities—using techniques like latent traversal and generative modeling—to provide actionable, contrastive insights and bridge the gap between prediction and explanation.
• SCEs are applied across various domains, including vision, tabular data, sequential planning, and large language models, enhancing transparency, fairness, and recourse in high-stakes decision-making.

Self-Generated Counterfactual Explanations (SCEs) are a class of explainability methods wherein the model itself, or an architecture tightly coupled with it, autonomously generates explanations by identifying minimal changes to an input that would alter its prediction. The paradigm aims to provide actionable, contrastive insight into model decisions by producing “what-if” scenarios, such as identifying which evidence is missing for a classification or how to change a feature vector to cross a decision boundary. The theoretical underpinnings, methodological spectrum, evaluation standards, and limitations of SCEs vary remarkably across tasks, input modalities, integration with planning and sequential models, and—in the case of LLMs—the very nature of generative reasoning.

1. Formal Definition and Core Principles

Self-Generated Counterfactual Explanations are distinguished from post-hoc or surrogate-based counterfactual methods by the use of native model capacities (prompting, internal modules, or joint training) to directly or indirectly generate counterfactuals. In its canonical form, given an input $x$ with prediction $y = f(x)$, the SCE task is to output $x'$ such that $f(x') = y'$ with $y' \neq y$ and, ideally, $x'$ is minimally different from $x$ under a task-suitable distance function. In vision, SCEs point out absent discriminative evidence (“this bird is not a Scarlet Tanager because it does not have black wings.”) (Hendricks et al., 2018); in tabular or sequence modeling, SCEs identify changes to feature values or action choices (Tsirtsis et al., 2020, Tsirtsis et al., 2021, Belle, 13 Feb 2025).

The SCE objective formally decomposes into two constraints: validity (counterfactual achieves the target label) and minimality (smallest possible change in input space or latent representation).

2. Methodological Spectrum

Visual and Structured SCEs

In computer vision, three methodological threads dominate:

• Missing Evidence Textual SCEs: First, attribute-level textual SCEs mine discriminative noun phrases from counter-classes, then inspect input images for the presence of these attributes, and generate natural language explanations by negating absent class-defining evidence. Candidate features are extracted from an explanation model, checked via phrase-critic or multimodal classifier, and negated to form a counterfactual sentence, e.g. “This is not a Bobolink because it does not have a yellow nape” (Hendricks et al., 2018). This modular pipeline integrates language chunkers, visual grounding, and rule-based generation.
• Discriminant Attribution SCEs: Optimization-free discriminant SCEs, such as SCOUT, combine predicted-class attributions, counter-class attributions, and self-aware confidence maps to localize image regions that are informative for the prediction but not for the alternative class. Explanations are formed by elementwise combination and thresholding of attribution maps (Wang et al., 2020).
• Generative and Manifold-Constrained SCEs: In visual domains and high-dimensional tabular data, more recent techniques employ autoencoders or generative models (GANs, diffusion models, VAEs) whose latent spaces are regularized to reflect class structure (Zhao et al., 2023, Madaan et al., 2023, Pegios et al., 4 Nov 2024, Bender et al., 17 Jun 2025). Counterfactuals are generated by interpolation (in Gaussian mixture models for classification, or disentangled representations for regression) or by guided diffusion within the data manifold, optimizing for both validity and plausibility with proximity and diversity losses. The Riemannian latent traversal approach leverages decoder and classifier Jacobians to constrain optimization paths to realistic and robust trajectories (Pegios et al., 4 Nov 2024).

Sequential and Structured Planning SCEs

In planning and sequential decision-making, SCEs generalize from inputs to sequences of actions or even domain-level modifications:

• Plan-Based SCEs: Action sequence counterfactuals are formulated as plans (or subplans) that, if executed from a given state or situation, would produce different outcomes. Modal fragments of situation calculus are used to formalize agent knowledge, beliefs, and counterfactual branches, enabling reconciliation of an agent's plan with user corrections or model amendments (Belle, 13 Feb 2025).
• Domain Modification SCEs: At the domain level, existential and universal counterfactual scenarios determine minimal changes to the problem definition (initial state, goal, or action preconditions) so that plans satisfying user-specified $LTL_f$ properties become possible, or all plans must satisfy them, respectively (Gigante et al., 29 Aug 2025).

Time Series and Temporally Constrained SCEs

• Causal Time Series SCEs: CounTS formalizes SCE generation in time series via variational Bayesian modeling, with explicit abduction-action-prediction steps that account for confounders. Counterfactuals are generated by fixing exogenous variables and minimally intervening on actionable time-series components (Yan et al., 2023).
• Temporal Logic-Constrained SCEs: In process mining, temporal constraints expressed in LTLp are compiled into DFAs and injected into genetic SCE search via knowledge-aware crossover/mutation operators, ensuring that generated counterfactual traces are both classifier-valid and rule-compliant (Buliga et al., 3 Mar 2025).

LLM SCEs

• LLMs are prompted to generate SCEs by revising their inputs to flip their predictions (validity) while making minimal edits. Evaluation involves both unconstrained and minimal-change instruction settings. Metrics include success (prediction flip) and edit distance (input modification extent) (Dehghanighobadi et al., 25 Feb 2025, Mayne et al., 11 Sep 2025).

3. Evaluation, Metrics, and Limitations

Key Evaluation Metrics

• Validity: The fraction of SCEs that, when evaluated with the original model, yield the desired target outcome.
• Minimality/Excess Distance: The difference between the distance traversed by the SCE and the minimal distance necessary to effect a decision change, often computed via metrics such as Gower’s Distance, $L_1$/$L_2$ norms, or normalized edit distance (for text inputs) (Mayne et al., 11 Sep 2025, Zhao et al., 2023).
• Sparsity, Proximity, and Diversity: Additional metrics include the number of features changed, distance to target, feature overlap, and diversity among multiple SCEs (for sufficiency).
• Plausibility/Manifold Proximity: Realism measures include negative log-likelihood under an autoregressive model, in-distribution metric like MMD, or reconstruction loss through generative model decoders (Madaan et al., 2023, Zhao et al., 2023).

Empirical Findings and Trade-offs

Across studies, several systematic limitations have been identified, particularly in LLM-driven SCEs:

• Validity–Minimality Trade-off: Unconstrained prompting yields high validity but poor minimality—SCEs overshoot the decision boundary, making excessive changes that obscure the true frontier. Conversely, explicit instructions for minimal modification result in small edits that often fail to flip the prediction, demonstrating a robust trade-off across multiple datasets, models, and distance functions (Mayne et al., 11 Sep 2025).
• Causal Misalignment: Standard ML-based SCE methods that ignore the true causal structure of data can generate counterfactuals that are invalid when evaluated using Pearl’s SCM-based abduction-action-prediction process. Empirical discrepancies reach ~30% even under controlled conditions, highlighting the need for causally-aware generation (Smith, 2023).
• Disagreement Among Methods: There is high disagreement among different SCE algorithms regarding which features to modify, independent of classifier type but driven by optimization (proximity vs. plausibility) and dataset characteristics. This disagreement exposes SCEs to risks of manipulation and “fairwashing,” undermining trust and transparency (Brughmans et al., 2023).

4. Formal and Algorithmic Foundations

Mathematical Formulations

• SCEs are often framed as a constrained optimization problem: minimize $d(x, x')$ s.t. $f(x') = y'$ and (optionally) $x'$ lies on the data manifold $\mathcal{M}$ (Bender et al., 17 Jun 2025).
• For structured latent-space approaches:

$z_{\text{cf}} = (1 - \alpha)z + \alpha \mu_{y'}$

with $z$ as latent embedding of $x$, $\mu_{y'}$ the target class centroid (Zhao et al., 2023).

• Riemannian traversal incorporates decoder Jacobian in the latent space metric:

$$M_Z(z) = J_\mu(z)^\top J_\mu(z) + J_\sigma(z)^\top J_\sigma(z)$$

and the update rule:

$$z' = z - \eta \cdot M_Z^{-1}(z)\nabla_z f_y(z)/\|M_Z^{-1}(z)\nabla_z f_y(z)\|_2$$

(Pegios et al., 4 Nov 2024).

Algorithmic Summaries

• Evidence-checking for attribute SCEs uses elementwise-multiplied, L2-normalized fusion of visual and textual features, scored by a linear layer (Hendricks et al., 2018).
• Counterfactual generation in planning is formalized as finding a minimal modification $P \rightarrow P'$ such that $\exists \pi' \in \Pi(P')$ or $\forall \pi' \in \Pi(P'), \pi' \models \psi$ for some temporal property $\psi$ (Gigante et al., 29 Aug 2025).
• For LLM-based SCEs, the prompt-based SCE generation process is:

1. Query prediction $y = f(x)$.
2. Select target $y'$.
3. Instruct model to generate $x'$ minimally different from $x$ such that $f(x') = y'$.
4. Compute normalized edit distance and excess distance for $x, x'$ (Mayne et al., 11 Sep 2025).

5. Applications and Impact

SCEs have direct applications in high-stakes and regulatory domains—including finance (loan recourse), medicine (treatment recommendations), strategic behavior modeling (with individuals adapting feature values to cross decision boundaries) (Tsirtsis et al., 2020), sequential guidance in therapy (Tsirtsis et al., 2021), automated planning and domain repair (Gigante et al., 29 Aug 2025), as well as education and model debugging in visual and time series domains (Hendricks et al., 2018, Yan et al., 2023).

Their impact on user trust, bias detection, actionable recourse, and model reliability is contingent on the faithful balance between validity, minimality, plausibility, and sufficiency, as well as transparency in the methodology and algorithm selection (Brughmans et al., 2023).

6. Limitations, Open Issues, and Future Directions

Persistent challenges include the observed validity–minimality trade-off in LLM SCEs, disagreement across counterfactual algorithms, and the risk of causally-invalid explanations. There is increasing recognition—both via empirical studies and theoretical analysis—that SCEs without constrained optimization, causal grounding, or explicit manifold regularization are likely to be unreliable, especially in high-stakes settings (Mayne et al., 11 Sep 2025, Smith, 2023, Brughmans et al., 2023).

Recommendations for future work include:

• Incorporation of causally-aware objectives and structural constraints in SCE generation.
• Joint or end-to-end architectures that align predictive and explanatory tasks (e.g., VCNet, Smooth Counterfactual Explorer) (Guyomard et al., 2022, Bender et al., 17 Jun 2025).
• Development of evaluation metrics that jointly capture validity, minimality, diversity, plausibility, and trustworthiness.
• Research into more advanced prompt engineering, fine-tuning objectives, and explicit self-consistency regularization in LLMs for robust SCEs (Dehghanighobadi et al., 25 Feb 2025, Mayne et al., 11 Sep 2025).
• Enhanced transparency and documentation of the SCE generation process, especially in applications subject to legal or regulatory scrutiny (Brughmans et al., 2023).

Self-Generated Counterfactual Explanations remain a central—yet challenging—component of explainable AI, with active research focused on aligning technical, causal, and operational desiderata for both researchers and practitioners.