Counterfactual Action Guidance (CAG)

• Counterfactual Action Guidance (CAG) is a framework that employs structural causal models to identify minimal, causally consistent interventions.
• It combines logic-based planning, rule-based surrogates, and RL-based methods to generate actionable recourse across diverse applications.
• Empirical evaluations in domains like credit risk, healthcare, and robotics demonstrate improved decision outcomes with cost-effective, interpretable interventions.

Counterfactual Action Guidance (CAG) is a research area and suite of methodologies that aim to generate actionable, causally-grounded interventions—typically as minimal or feasible edits to actions, features, or policies—to achieve desired counterfactual outcomes in complex machine learning and decision-making systems. CAG is foundational in domains where transparency, recourse, and interpretability are needed for automated predictors, sequential policies, or human-in-the-loop analytics. Implementations span rule-based logic planning, continuous control, offline RL, sequential recourse, causal data augmentation, and interactive visualization frameworks.

1. Theoretical Foundations and Problem Formulation

At its core, CAG is rooted in the formalism of structural causal models (SCMs) and counterfactual reasoning. Given an initial state—such as an input feature vector for a classifier, a trajectory in a Markov decision process (MDP), or an event-driven process—CAG seeks an intervention (or sequence thereof) such that, once applied, the model yields a favorable outcome, all while satisfying causal constraints imposed by the domain (Dasgupta et al., 2024, Dasgupta et al., 24 Aug 2025, Tsirtsis et al., 2021, 2505.12701, Schulam et al., 2017).

For static tabular models with features $X_1,\dots,X_n$, CAG considers the SCM $G=(V,E)$ (a causal DAG with structural equations $X_i \leftarrow f_i(pa(X_i), U_i)$). Interventions replace certain equations with fixed values (do-operations), and a search is performed for a minimally invasive sequence of changes $do(A=a)$ that leads from an observed state $I$ (undesired outcome) to some $S_{goal}$ (desired outcome), under the overall constraint $S_{goal}\in S_C$ (causally consistent subspace).

In sequential or RL settings, the goal is to minimally alter a trajectory $\tau=(s_0,a_0,\dots,s_T)$ (finite-horizon MDP or POMDP) so the resulting (counterfactual) outcome trajectory achieves improved expected reward or matches a desired property. Typically, this is formalized as

$$\max_{a'_{0:T-1}:\; \mathrm{Ham}(a',a)\le k} \; \mathbb{E}\left[\sum_{t=0}^{T-1} R(s'_t, a'_t)\right],$$

subject to state transition and edit budget constraints (Tsirtsis et al., 2021, 2505.12701).

Cost models are integral; they quantify the user or system burden of each intervention, distinguishing between direct (user-initiated) and indirect (causally-propagated) changes (Dasgupta et al., 24 Aug 2025).

2. Core Methodologies and Algorithmic Paradigms

Counterfactual action guidance spans a wide algorithmic spectrum:

• Logic-based Causality-Constrained Planning: Frameworks such as CFGs (Dasgupta et al., 2024) encode decision rules and the causal graph in Answer Set Programming (ASP), using a goal-directed ASP system (s(CASP)) for symbolic planning. Interventions are actions in the logic program, and the system searches for a sequence ("Plan") leading from the initial to the goal state, always maintaining causal consistency at each step.
• Rule-based Surrogates with User-Effort Costs: MC3G (Dasgupta et al., 24 Aug 2025) approximates black-box models by extracting rule-based surrogates, then finds interventions that both flip the decision and respect causal constraints. Only user-initiated changes accrue cost; automatically induced (causal) changes are cost-free. The algorithm optimizes

$$\min_x C_{\rm user}(x_0, x) = \sum_k w_k |x_k - x_{0,k}|^p,$$

where $G=(V,E)$0 if feature $G=(V,E)$1 is only changed by propagation.

• Sequential/Consequence-Aware Recourse: CSCF and LocalFACE (Naumann et al., 2021, Small et al., 2023) model recourse as a path-planning problem through feasible, data-dense regions, often employing multi-objective genetic algorithms to optimize action sequences under direct and consequence-discounted costs, while maximizing diversity and satisfying ordering, actionability, and domain-specific constraints.
• Continuous-Action RL Counterfactuals: In continuous-control MDPs, CAG is instantiated via deep reinforcement learning (RL), e.g., with TD3 variants that optimize for improved reward subject to minimal deviation in the action sequence, possibly under hard or soft constraints such as mandatory policy adherence in subsets of states (2505.12701). The optimization criterion is typically

$G=(V,E)$2

where $G=(V,E)$3 is a distance metric.

• Offline RL Data Augmentation: CAIAC (Urpí et al., 2024) evaluates per-entity causal influence using conditional mutual information, then constructs counterfactual data points by swapping action-unaffected entity subspaces between trajectories, expanding the support of the offline data in a causally faithful manner.
• Vision-Language-Action Model Correction: CAG is used for inference-time regularization of VLAs (Fang et al., 19 Feb 2026): a dual-branch, classifier-free guidance interpolation combines language-conditioned and language-unconditioned policies to correct vision-based shortcutting and restore proper language grounding in action selection.
• LLM-Based Autonomous Control: The Conformal Counterfactual Generation (CCG) framework (Farzaneh et al., 27 Jan 2026) applies SCMs to the closed-loop system (User → LLM → Environment) and, via abduction and test-time scaling of latent environment noise, generates sets of counterfactual outcomes or action plans with statistical coverage guarantees, supporting causal introspection in LLM-driven agentic systems.

3. Empirical Benchmarks and Illustrative Domains

CAG methodologies are validated across numerous benchmarks:

• Loan Approval and Credit Risk: CFGs and MC3G are benchmarked on UCI Adult, German Credit datasets, where optimal intervention sequences produce actionable, causally-plausible recourse (Dasgupta et al., 2024, Dasgupta et al., 24 Aug 2025).
• Sequential Healthcare and Therapy: Dynamic programming-based CAG (Tsirtsis et al., 2021) is applied to cognitive behavioral therapy (CBT) data, providing interpretable, high-impact recommendations (e.g., when and how to alter therapy themes) and validated with real patient trajectories.
• Robotics and Control: CAIAC is tested on D4RL Franka-Kitchen, Fetch-Pick&Lift, and Fetch-Push environments: swapping action-unaffected entity states improves robustness to distributional shift and increases out-of-distribution (OOD) success rates over baselines (Urpí et al., 2024).
• Vision-Language Manipulation: LIBERO-CF benchmark quantifies counterfactual failures in VLAs; CAG improves language grounding and task success by up to 17.2% in real-world robot trials (Fang et al., 19 Feb 2026).
• Motion Guidance in Sports: Joint-level counterfactual motion guidance for badminton (Seong et al., 2024) generates plausible, classifier-flipping movement trajectories that are closer to the user's original motion than nearest-neighbor/expert imitation.
• Autonomous Control with LLMs: Wireless network scheduling via LLM-based agents with conformal CAG achieves lower KPI errors (throughput MAE: 0.15 vs. 0.28 Mbps for CAG vs. interventional generation), with semantic evaluation showing 92% preference for CAG outputs (Farzaneh et al., 27 Jan 2026).

4. Causal Constraints, Actionability, and Recourse

A unifying principle of CAG across domains is that counterfactuals must be causally-consistent and actionable. Counterfactuals violating the SCM-imposed structure are considered invalid, as they may suggest impossible or non-actionable recommendations. Methods like CFGs and MC3G enforce causal closure by encoding causal relationships as explicit constraints—either in logic rules (ASP, Horn clauses) or via the design of the cost functional (zero effort for causally-propagated effects).

Actionability is further nuanced in path-based and sequential frameworks, with explicit modeling of feature- or action-specific preconditions, consequence-discounts (how earlier actions modulate the cost/effect of later ones), and modification frequencies. Pareto-optimal sets of solutions offer diversity in recourse, respecting user constraints such as feature immutability or sequence order (Naumann et al., 2021, Small et al., 2023).

5. Evaluation Metrics and Comparative Outcomes

Empirical evaluations of CAG frameworks employ a suite of quantitative metrics tailored to the underlying domain:

• Causal Compliance/Consistency: Percentage of counterfactuals residing in $G=(V,E)$4, the causally-allowed state space (Dasgupta et al., 24 Aug 2025).
• Distance Metrics: $G=(V,E)$5 norms or domain-specific proximities (e.g., Fréchet, DTW distances for motions) to capture recourse burden and plausibility (Seong et al., 2024, Dasgupta et al., 24 Aug 2025, Naumann et al., 2021).
• Action Set Size: Number of user-initiated interventions—smaller, sparser sets are preferred (Dasgupta et al., 24 Aug 2025).
• Reward Improvement/Performance Gain: Proportions of trajectories where counterfactual action sequences raise cumulative return or outperform baselines (e.g., $G=(V,E)$6, $G=(V,E)$7 in RL benchmarks) (2505.12701).
• Feasibility and Diversity: Number of Pareto-optimal solutions, diversity of action pathways (Naumann et al., 2021).
• User Study Outcomes: Accuracy, ranking offset, efficiency, and exploration pattern metrics in interactive analytics settings (Wang et al., 2024).

6. Limitations, Open Problems, and Future Directions

Open technical and theoretical challenges in CAG research include:

• Scalability: Extending to high-dimensional, continuous-control/action spaces—especially with complex, entangled causality or long-horizon planning—remains challenging (2505.12701).
• Theoretical Guarantees: Approximations (e.g., expectation surrogates for essential suprema) lack tight performance guarantees in highly stochastic or adversarial settings (2505.12701).
• Faithful Causal Discovery: When only partial domain causal knowledge is available, inferring the correct SCM or causal influence remains a nontrivial bottleneck (Urpí et al., 2024, Dasgupta et al., 24 Aug 2025).
• User-Focused Recourse: Automatic suggestion of initial recourse path, integration of user-driven preferences, and adaptive weighting of costs in interactive systems are active areas (Wang et al., 2024).
• Generalization: Robustness to OOD counterfactuals, as well as efficient transfer to real-world deployment, needs further exploration, particularly in safety-critical domains (Urpí et al., 2024, Fang et al., 19 Feb 2026).
• Interpretability Beyond Trajectories: Methods to summarize, visualize, or explain counterfactual action sequences and strategies in more accessible formats (e.g., language, natural explanations) are underdeveloped (2505.12701).

7. Representative Implementations and Comparative Summary

The following table summarizes key CAG frameworks, their domains, and unique methodological features:

Framework	Domain	Unique Mechanism / Constraint
CFGs (Dasgupta et al., 2024)	Tabular recourse	Goal-directed ASP, dual rules for alternate worlds
MC3G (Dasgupta et al., 24 Aug 2025)	Black-box classifiers	Rule-based surrogate, zero user-cost for causal propagation
CSCF (Naumann et al., 2021)	Recourse sequences	Multi-objective BRKGA, consequence discounts, diversity
CAG-RL (2505.12701)	Continuous-action RL	TD3-based, min-distance improved-reward actions
CAIAC (Urpí et al., 2024)	Offline RL & Robotics	Action influence (CMI), controllable entity swapping
LocalFACE (Small et al., 2023)	Algorithmic recourse	Locally-guided, path-based, data-density constraints
VLA-CAG (Fang et al., 19 Feb 2026)	Vision-language-action	Dual-branch inference for counterfactual policy
CCG (Farzaneh et al., 27 Jan 2026)	LLM-based control	SCM abduction, conformal risk calibration

In all, Counterfactual Action Guidance provides a unified, causality-aware framework for bridging the gap between predictive ML models, sequential decision-makers, and practical recourse—enabling actionable, plausibly minimal, reliably justified interventions across diverse application domains.