PACE: feasible counterfactual explanations via neuro-symbolic AI

A scoring model denies a loan and the explanation system suggests the applicant should «be five years younger» or «change their zip code». Impossible or unacceptable recommendations like these are the problem tackled by PACE, a neuro-symbolic framework for generating plausible and actionable counterfactual explanations, published on arXiv on July 3, 2026.

The problem: counterfactuals nobody can act on

Counterfactual explanations answer the question «what would have to change in the input for the model to decide otherwise». They are one of the most valued explainability formats because they speak the language of the affected user: instead of showing weights or attention maps, they offer an action. The problem, as the paper itself points out, is that many existing methods generate alternatives that do change the prediction but turn out to be unrealistic or infeasible, because they incorporate no explicit mechanisms for domain knowledge or constraints on which interventions are possible.

What PACE proposes

The proposal splits the problem into two components. On one side, a neural predictive model handles classification, with no changes from the usual setup. On the other, a symbolic reasoning layer enforces domain-specific constraints during counterfactual generation: human-readable rules that define which actions are feasible and which are not. By explicitly modeling feasible interventions, the system produces explanations consistent with domain knowledge and, according to the authors, interpretable and actionable.

Two design details deserve attention. The framework is modular and model-agnostic: it does not require retraining the classifier or accessing its internals, which makes it easier to apply on systems already in production. And the separation between prediction and reasoning follows the line of neuro-symbolic AI, which combines data-trained models with explicit symbolic rules, a current gaining traction precisely in the problems where pure learning falls short on guarantees.

Why it matters

Explainability is no longer an academic topic. The European AI Act imposes transparency obligations on high-risk systems, and sectors such as banking, insurance and healthcare have been required to justify automated decisions for years. In that context, an infeasible counterfactual is not just a bad explanation: it is a legal and reputational risk. Telling a customer their loan would be approved «at another age» is useless; telling them it would be approved by bringing their debt ratio below a specific threshold is an explanation you can defend before a regulator.

The trend also connects with the rise of agents: as systems automate decisions end to end, explaining what would have changed the outcome stops being a courtesy and becomes an operational requirement.

Who this is useful for

For ML teams operating decision models in regulated domains, PACE points to a practical architecture: keep the existing classifier and add the symbolic layer with the business rules that usually already exist in internal policy documents. For auditors and risk officers, the approach offers something verifiable: the constraints are explicit and readable, not buried in the weights of a network.

Reasonable limits

Expectations should be tempered. This is a freshly announced paper (version 1) and the abstract includes no quantitative comparisons with previous methods, so the empirical evidence remains to be examined in the full text. The approach also inherits the classic bottleneck of the symbolic world: someone has to write and maintain the domain rules, expert work that does not scale on its own and that goes stale if the business changes and nobody updates the constraints.

At EP we have long argued that the feasible counterfactual is one of the few branches of explainability with direct product utility: it does not just say why the model rejected, but what the user can do about it. If PACE delivers on its premise, the value will lie less in the algorithm and more in forcing organizations to make their rules explicit.