Directive Counterfactuals (DCFs)

• Directive Counterfactuals are actionable explanations defined by a target outcome and a concrete sequence of feasible actions to transform the factual state.
• They employ methods like model inversion, structural causal modeling, and MDP planning to generate cost-effective and context-aware intervention paths.
• DCFs are applied in machine learning recourse and robotics to provide explicit, user-adapted guidance that balances feasibility, action cost, and real-world constraints.

Directive Counterfactuals (DCFs) are actionable counterfactual explanations defined by both a desired hypothetical outcome and an explicit, feasible sequence of interventions or actions to achieve that outcome. Unlike classical counterfactuals, which merely describe alternate realities, DCFs prescribe or diagnose concrete pathways to change the factual state. Across causal inference, recourse in ML, and embodied agent control, DCFs operationalize the inversion of predictive or causal models to support decision making, recourse, and robustness. Their formalization integrates model inversion (to find antecedents for specified consequents), feasibility, action costs, and human or contextual constraints (Marwala, 2014, Balke et al., 2013, Singh et al., 2021, Yan et al., 29 Sep 2025).

1. Formal Definitions and Taxonomy

A Directive Counterfactual extends the classical counterfactual (a “what if” statement) by incorporating a directive plan—a concrete set of interventions (or actions) transforming the factual state $x$ to a counterfactual state $c$ that realizes the domain-specified behavioral or decision objective.

• Generic counterfactual: “If $x$ had been $c$ instead, then $y$ would have been $y^*$.”
• Directive counterfactual: “To achieve $y^*$, enact sequence $T$ of actions/changes mapping $x$ to $c$ so that $c$0.”

Formally, in supervised ML or structural models:

• Instance space: $c$1. Classifier or causal model $c$2.
• Directive Explanation (Singh et al., 2021): 6-tuple $c$3, where $c$4 is a policy or plan transforming $c$5 to $c$6. $c$7 is chosen so that $c$8, and $c$9 is cost- and feasibility-constrained.
• Structural Causal Models (SCM) (Balke et al., 2013): Given endogenous variables $x$0, exogenous $x$1, and functions $x$2, a DCF is the evaluation of $x$3 under the policy intervention $x$4, representing an enforced action or treatment.

In robotics and embodied instruction following, DCFs arise when prescribed directives are at odds with environmental or perceptual reality, creating compliance conflicts detected by discrepancy between human priors, static robot knowledge, and real-time perceptual data (Yan et al., 29 Sep 2025).

2. Model-Based Approaches to DCF Computation

Model Inversion and Prescriptive Counterfactual Generation

Rational counterfactuals (Marwala, 2014):

• Given $x$5, invert $x$6 to find $x$7 (e.g., squared error) subject to domain constraints.
• Use simulated annealing to optimize $x$8 over a feasible region.

SCM-based DCFs (Balke et al., 2013):

• Represent interventions as $x$9 surgically replace equations in the causal model.
• Compute $c$0 via three-step procedure: (1) abduction (infer likely exogenous $c$1 given evidence $c$2), (2) action (mutilate model to impose policy), (3) prediction (compute $c$3 under $c$4).
• Supports both deterministic and stochastic models, with guarantees for identified SCMs.

ML Recourse as DCFs (Singh et al., 2021):

• Counterfactual search: $c$5.
• Extend to directive planning: Explicitly generate an action path $c$6 (sequence of feasible atomic actions $c$7 with costs) connecting $c$8 to $c$9.

Actionability, Feasibility, and Optimization

• Feasibility constraints: Only mutable, accessible features may be acted upon; cost functions encode practical difficulty or social acceptability.
• Optimization objective: Find $y$0, where $y$1 is the set of reachable states via feasible actions, and $y$2 sums action costs along $y$3 (Singh et al., 2021).
• Policy modeling: Planner cast as an MDP $y$4, where $y$5 is state, $y$6 is action space, $y$7 describes transitions, and $y$8 encodes negative action cost plus terminal reward for $y$9.

3. DCFs in Human-Centric Recourse and Explanation

Directive counterfactuals are critical in ML recourse, actionable explanations, and user-facing decision support (Singh et al., 2021). Unlike non-directive counterfactual queries, DCFs:

• Provide explicit stepwise action policies, not just endpoints.
• Account for individual user feasibility and context—e.g., prohibit recommendations such as “change age”.
• Offer both specific directives (concrete actions) and generic directives (action types).
• Support cost-sensitive optimization (minimize time, money, effort), potentially via interactive user feedback to adapt plans.

Empirical user studies confirm a strong preference for directive-specific explanations (49% first choice), especially when seeking recourse after adverse decisions (e.g., loan denials). Preferences are influenced by scenario context, feasibility, and social factors, underscoring the importance of adaptable, context-aware DCF generation (Singh et al., 2021).

4. DCFs in Causal Policy Analysis and Interventional Semantics

Within structural causal modeling, DCFs embody interventions encoded as $y^*$0 or $y^*$1, effectuating deliberate manipulations of endogenous variables (Balke et al., 2013). The key distinctions are:

• DCFs prescribe actions as direct interventions, not mere conditioning.
• The abduction-action-prediction workflow supports integration of factual evidence with hypothetical policies.
• Soundness and completeness (identifiability): If $y^*$2 and $y^*$3 are fully specified, $y^*$4 is uniquely determined and computable.

Illustrative examples include policy interventions in econometric models (e.g., fixing price to analyze demand) and diagnostic reasoning in deterministic systems (e.g., “What if Bob hadn’t fired?” in firing squad counterfactuals).

5. Embodied Agents and DCF Recognition in Robotics

In robotics, DCFs arise particularly in task execution conditioned on natural language directives that may be misaligned with sensorimotor reality (Yan et al., 29 Sep 2025). The DynaMIC framework operationalizes DCFs as follows:

• The robot maintains $y^*$5 (human prior), $y^*$6 (static robot schema), and dynamic $y^*$7 (perceptual data).
• A discrepancy function $y^*$8 signals contradiction between instruction assumptions and observed environment.
• DCF detection involves: semantic-level feasibility checks, multimodal grounding pipelines, trajectory planning that inserts perception steps, and human-in-the-loop feedback.
• DCFs trigger natural-language feedback—proactively halting or replanning execution if the directive is found to be infeasible, hazardous, or ill-aligned with physical reality (e.g., “The bowl is too small for that fruit.”).
• Experiments show robust DCF detection (44/50 correct) and critical roles for multi-level grounding and trajectory refinement modules.

6. Algorithmic Realizations and Integration in Practice

Practical implementation of DCFs involves several algorithmic modules and pipeline integration strategies:

• Candidate generation: Off-the-shelf counterfactual generators produce plausible $y^*$9 given $y^*$0.
• Action modeling: Construct catalogues of domain-relevant atomic actions with feasibility and cost metrics.
• Plan synthesis: MDP solvers (value iteration, Q-learning, deterministic planners) derive cost-optimal, feasible action policies $y^*$1 that realize recourse or intervention.
• User adaptation: Interactive modules allow users to constrain/disallow certain actions, thereby personalizing DCFs.
• Explanation rendering: Translate action plans into user-appropriate specifics—either as ordered action lists or higher-level recommendations.
• Integration points: Deploy at decision time in ML pipelines or at planning/replanning junctures in autonomous agents; optionally support fairness and sensitive-feature protections, as well as pre-computation for frequent DCFs.

7. Comparative Summary and Empirical Validation

DCFs differ from standard counterfactuals as summarized below:

Aspect	Standard Counterfactual	Directive Counterfactual
Output	End-state $y^*$2	Action sequence $y^*$3, end-state $y^*$4
Actionability	Implicit	Explicit, planned
Feasibility/Social Cost	Not modeled	Explicit constraints and costs
User adaptation	Rare	Central
Causal structure	May be lost	Policies as interventions

Empirical results in both recourse-oriented ML (Singh et al., 2021) and robot instruction following (Yan et al., 29 Sep 2025) demonstrate the efficacy and necessity of DCFs for actionable guidance, robustness to spurious commands, and maintaining alignment with user intentions.

---

References:

• "Rational Counterfactuals" (Marwala, 2014)
• "Directive Explanations for Actionable Explainability in Machine Learning Applications" (Singh et al., 2021)
• "Counterfactuals and Policy Analysis in Structural Models" (Balke et al., 2013)
• "DynaMIC: Dynamic Multimodal In-Context Learning Enabled Embodied Robot Counterfactual Resistance Ability" (Yan et al., 29 Sep 2025)