XplainAct: Actionable Explanations in AI

• XplainAct is an explainable AI paradigm that integrates explanation with action, intervention, and real-time query handling across varied domains.
• It operationalizes explanation through methods like natural-language interfaces, contrastive reasoning, causal inference, and model manipulation.
• Applications range from remote autonomous missions and AI planning to personalized intervention analytics and generative art, demonstrating its versatile impact.

XplainAct is a recurring label in explainable AI research for systems that couple explanation with action, intervention, or interactive use. Rather than denoting a single canonical algorithm, it appears across several technically distinct lines of work: a natural-language interface for scrutable autonomous robots, a contrastive framework for explaining planner decisions, a visual analytics system for personalized intervention reasoning, and a craft-based approach to explainability in diffusion-model art practice. Several syntheses also map related work in discourse-driven explanation, explanation-as-action, causal concept-based explanation, and interpretable manipulation onto the same label, indicating a shared emphasis on explanation that is operational, situated, and queryable rather than purely descriptive (Garcia et al., 2018, Borgo et al., 2018, Zhang et al., 19 Jul 2025, Abuzuraiq et al., 10 Aug 2025, Akula et al., 2022, Cope et al., 2023, Bjøru et al., 2 Dec 2025).

1. Terminological scope and research uses

The name is not used uniformly. In some papers, XplainAct is the name of the system or methodology itself; in others, it is used as an overview label for an explanation paradigm centered on actionable, contrastive, or intervention-oriented reasoning.

Context	XplainAct denotes	Characteristic mechanisms
Remote autonomous missions	Natural-language explanation interface	Chat-based “what/why/why not,” expert think-aloud model
AI planning	Contrastive explanation of action choice	Foil injection, replanning, $\Delta C$, $\Delta M$, causal support
Personalized intervention analytics	Visual analytics framework	Local SCMs, subgroup simulation, LIME/SHAP, PCP/map views
Generative art	Explainability-in-Action	ComfyUI model bending, feature-map visualization, CLIP adjustments

In the autonomous-systems line, XplainAct addresses operator queries during remote missions and is realized through MIRIAM alongside SeeTrack and Neptune (Garcia et al., 2018). In AI planning, the term is attached to the question “Why action $a$ rather than $\hat a$?” and is operationalized in XAI-Plan through controlled replanning and plan comparison (Borgo et al., 2018). In visual analytics, XplainAct is a framework for simulating and explaining interventions at the individual level within subpopulations (Zhang et al., 19 Jul 2025). In generative art, “Explainability-in-Action” designates a craft-based approach in which diffusion-model internals are exposed and manipulated during artistic practice (Abuzuraiq et al., 10 Aug 2025).

Several adjacent papers explicitly state that the original work does not use the name “XplainAct,” but map their own framework to it in exposition. This applies to the UCLA natural-language explanation architecture, the explanation-as-action formalization, and the causal concept-based post-hoc framework (Akula et al., 2022, Cope et al., 2023, Bjøru et al., 2 Dec 2025). A plausible implication is that XplainAct has become a family-resemblance term for explanation systems in which explanation is inseparable from querying, intervention, or downstream decision support.

2. Interactive explanation in autonomy and planning

In remote autonomous missions, XplainAct is designed for settings in which a robot is physically remote, autonomy is high, and misunderstanding a behavior can be costly. The system is deployed as MIRIAM, a multimodal chat interface connected to SeeByte’s SeeTrack interface and Neptune autonomy software. Operators query the vehicle in natural language with “what,” “why,” and “why not” questions; dialogue management tracks conversation state and follow-ups, NLU identifies behavior targets and intent, an autonomy-model adapter retrieves current and historical mission state, and an explanation engine matches that state against an interpretable model derived from expert speak-aloud rationalizations. Explanations are generated through template-based NLG and verbalized according to confidence tiers. If $c(r_i \mid s_t, H_t)\in[0,1]$ is the confidence in reason $r_i$, then the label is high for $c\ge 0.80$, medium for $0.40 \le c < 0.80$, and low for $c<0.40$. The paper explicitly contrasts completeness, in which all candidate reasons are shown, with soundness, in which only high-confidence reasons are returned, and notes that the trade-off remains open in high-risk contexts (Garcia et al., 2018).

The planning formulation is contrastive rather than status-descriptive. XAI-Plan answers “Why action $a$ rather than $\Delta M$0?” by locating the state $\Delta M$1 at which the planner chose $\Delta M$2, enumerating the alternative applicable actions

$\Delta M$3

forcing the user-selected foil $\Delta M$4 into the planning problem, and comparing the resulting counterfactual plan $\Delta M$5 with the original plan $\Delta M$6. Four strategies are described: replanning from the post-foil state, replanning from the initial state with the user action forced, replanning from the initial state with a timed-initial-literal window, and separately planning initial and later segments. Explanations are grounded in feasibility, cost difference $\Delta M$7, makespan difference $\Delta M$8, edit distance, preconditions, effects, and weakest conditions for later-plan viability. In the DriverLog example, the original two-truck plan has $\Delta M$9, a single-truck alternative has $a$0, and therefore $a$1; a poor counterfactual that preserves irrelevant early commitments yields $a$2 and $a$3, illustrating how context preservation can distort explanation quality (Borgo et al., 2018).

Taken together, these two lines instantiate two different meanings of scrutability. In remote autonomy, the explanandum is the live behavior of an executing system; in planning, it is the planner’s choice among locally applicable alternatives. Both nevertheless rely on state reconstruction, causal constraints, and user-driven queries rather than static post hoc summaries.

3. Explanation as communicative action

A more abstract formulation treats explanation itself as an action sequence whose success depends on changes in the explainee’s internal state. In that formalization, explanation is a sequence of observations made by the explainee, explanatory actions are those of the explainer that generate those observations, and explanatory effectiveness is defined as

$a$4

where $a$5 is an understanding measure for explanandum $a$6 at time $a$7. The paper frames explanation as a two-player cooperative game $a$8 and builds the understanding measure from Algorithmic Information Theory and Universal Intelligence Theory, combining self-report, mutual information, compression, integration, and utilization over fair $a$9-tests. It also makes two points that are central to later XplainAct-style systems: explanations can have negative effectiveness, and explanation quality cannot be assessed independently of the explainee’s state (Cope et al., 2023).

The UCLA interaction framework operationalizes a related idea at the dialogue-system level. An Explainer receives the user’s natural-language question, performs dialog-act recognition and contrastive question-type classification, selects an explanation type with a rule-based controller, queries an Explainable Performer and an Atomic Performer for evidence, and generates a response while tracking discourse with Rhetorical Structure Theory. The evidence store is an And-Or Graph, and the explanation repertoire includes Direct Explanation, Part-based Explanation, Causal Explanation, Temporal Explanation, Common-sense Explanation, Counter-factual Explanation, and Discourse Entailment based Explanation. The framework also distinguishes contrastive question types such as NOT-X, NOT-X1-BUT-X2, and NOT-X-BUT-Y, and intervention types such as DO-X, DO-NOT-X, and DO-X-NOT-Y. The paper does not report quantitative results, but it specifies planned evaluation by precision/recall/F1 for intent recognition, METEOR and CIDEr for relevancy, and parse-graph fidelity through comparison of X-pg and STC-pg (Akula et al., 2022).

These formulations move XplainAct away from the artifact-centric view of explanation. The explanation is not merely a text or visualization attached to a prediction; it is a sequence of communicative acts shaped by user intent, discourse context, and the explainee’s evolving representation.

4. Counterfactual and causal intervention frameworks

In personalized intervention analytics, XplainAct is a visual analytics system for simulating, explaining, and reasoning about interventions at the individual level within subpopulations. Its motivation is that average treatment effects can conceal substantial heterogeneity, making population-level summaries inadequate in settings such as public health and political campaigns. The framework combines the potential-outcomes and structural-causal-model traditions, using quantities such as the ATE $\hat a$0, the individual treatment effect $\hat a$1, and the subgroup-conditional effect $\hat a$2. For a focal unit, a subgroup is built by LSH-based nearest-neighbor retrieval, a local structural model is parameterized within that subgroup, and a user intervention $\hat a$3 is propagated through dependent attributes to generate a counterfactual profile and outcome. The interface integrates a choropleth map, a parallel coordinates plot, and explanation panels with LIME and SHAP. In the opioid case study, the data cover more than 3,000 US counties and 11 attributes; in the voting case study, the SHAP waterfall for Webb County, Texas reports that a $\hat a$4 minority share contributes a $\hat a$5 shift from a national baseline of $\hat a$6 (Zhang et al., 19 Jul 2025).

A more formal causal version of XplainAct explains black-box models by intervening on human-meaningful concepts $\hat a$7 in a triplet $\hat a$8 consisting of a black-box predictor $\hat a$9, a concept-to-data mapping $c(r_i \mid s_t, H_t)\in[0,1]$0, and a concept-level SCM $c(r_i \mid s_t, H_t)\in[0,1]$1. The central score is the probability of sufficiency. In the binary case,

$c(r_i \mid s_t, H_t)\in[0,1]$2

and the local explanation of a model output is expressed as

$c(r_i \mid s_t, H_t)\in[0,1]$3

Global or subgroup explanations average these local quantities over a set $c(r_i \mid s_t, H_t)\in[0,1]$4 consistent with the subgroup context. In the CelebA proof-of-concept, the age classifier reports for one local query that $c(r_i \mid s_t, H_t)\in[0,1]$5 under the FSCM, with a canonical PSCM interval $c(r_i \mid s_t, H_t)\in[0,1]$6 and an independence baseline of $c(r_i \mid s_t, H_t)\in[0,1]$7. For the global age task conditioned on $c(r_i \mid s_t, H_t)\in[0,1]$8, singleton interventions yield $c(r_i \mid s_t, H_t)\in[0,1]$9 for $r_i$0, $r_i$1 for $r_i$2, $r_i$3 for $r_i$4, and $r_i$5 for $r_i$6; the corresponding gray-hair subgroup score is $r_i$7 for males and $r_i$8 for females (Bjøru et al., 2 Dec 2025).

Both systems make counterfactual reasoning central, but they differ in explanandum and interface. The visual-analytics framework targets intervention planning over observed populations with interactive subgroup definition; the concept-based framework targets post hoc explanation of a fixed predictive model through concept-level sufficiency scores. In both cases, explanation is tied to explicit interventions rather than correlation alone.

5. Explainability-in-Action in generative art

In the generative-art literature, XplainAct is explicitly defined as a craft-based approach to explainable AI that emphasizes learning by doing. It is grounded in Schön’s “reflection-in-action” and treats large text-to-image diffusion models not as opaque systems to be sampled, but as materials to be bent, inspected, and shaped. The implementation is a suite of model-bending and inspection nodes integrated into ComfyUI. The core nodes are Model Bending (SD Layers), which injects transformations such as rotation, scaling, additive or multiplicative perturbations, noise, and erosion or dilation into selected layers; Model Inspector, which exposes UNet and VAE architectures as nested trees; Visualize Feature Map, which displays intermediate activations; and CLIP conditioning adjustments, which enable local movements in the prompt-embedding space. The interaction cycle is immediate and iterative: select a layer, apply a bend, sample with fixed seed and parameters, inspect feature maps and outputs, and repeat (Abuzuraiq et al., 10 Aug 2025).

The paper frames these manipulations within standard diffusion-model notation. It writes the forward process as

$r_i$9

with $c\ge 0.80$0, and the reverse model as a conditional denoiser $c\ge 0.80$1. It also gives the classifier-free guidance combination

$c\ge 0.80$2

The qualitative centerpiece is a rotation study on Stable Diffusion v1.4 with prompt “analog style portrait of a person,” seed $c\ge 0.80$3, $c\ge 0.80$4 steps, CFG $c\ge 0.80$5, sampler dpmpp_2m, and Karras scheduler. Holding those parameters fixed while rotating selected middle-block layers exposes different functional roles for early and late UNet regions: early bends often disrupt coarse structure and layout, while later bends can inject stylization or texture shifts without destroying subject identity. The paper does not report formal user studies or quantitative metrics, but argues that such controlled, repeatable perturbation makes the model feel like a deformable material and supports tacit understanding (Abuzuraiq et al., 10 Aug 2025).

This use of XplainAct is notably different from decision-support and counterfactual-planning variants. Its explanandum is not a planner’s decision or a causal effect estimate, but the internal structure of a generative model as encountered through practice. Explainability is therefore embedded in manipulation, not in a separate explanatory report.

6. Related actionable-explanation lines

Several related systems in the provided corpus do not primarily use the XplainAct name but align closely with its action-oriented logic. In software engineering, CounterACT mines action rules directly from historical defect data instead of explaining a black-box predictor post hoc. A plan recommends changing flexible code metrics so that a buggy instance moves into a not-buggy region. Across 9 software projects, the reported release-level median overlap score is $c\ge 0.80$6, the commit-level median overlap is $c\ge 0.80$7, and the median F1-score is $c\ge 0.80$8; in an LLM-guided code-edit experiment on 40 Kafka bugs, 28 of 40 plan-guided edits passed tests, and the McNemar test gives $c\ge 0.80$9 (Oueslati et al., 2024).

In language-conditioned manipulation, Ex-PERACT introduces a discrete skill-code bottleneck above a voxel-based policy, making language-grounded skills inspectable and reusable. The high-level module maps language and temporally ordered observations to a quantized skill code, and the low-level module predicts discretized translation, rotation, and gripper actions over a $0.40 \le c < 0.80$0 voxel grid. Interpretability is qualitative rather than metricized: per-code word clouds reveal skill semantics such as color-focused targeting, count-driven placement, and screwing-related subskills. On eight RLBench tasks with 10 demonstrations each, the multi-language Ex-PERACT variant reports success rates of $0.40 \le c < 0.80$1, while PERACT reports $0.40 \le c < 0.80$2 (Zheng et al., 2024).

In multimodal misinformation detection, EXCLAIM contributes an explainable and actionable agentic architecture that mirrors human fact-checking. Its pipeline comprises a Retrieval Agent, Detective Agent, and Analyst Agent operating over three FAISS indices for visual entities, textual entities, and event-level context. The final output is $0.40 \le c < 0.80$3, where $0.40 \le c < 0.80$4 is the out-of-context judgment and $0.40 \le c < 0.80$5 is the evidence-grounded explanation. On NewsCLIPpings Merged/Balance, EXCLAIM reports $0.40 \le c < 0.80$6 accuracy, compared with $0.40 \le c < 0.80$7 for SNIFFER and $0.40 \le c < 0.80$8 for GPT-4o few-shot (Wu et al., 1 Mar 2025).

These adjacent systems reinforce a broad pattern. Whether the domain is code maintenance, robotic execution, or multimodal verification, the common design move is to make explanations intervene on the problem structure itself: alter metrics, select skills, retrieve counterevidence, or propagate causal changes. That pattern is consistent with the more formal explanation-as-action formulation and helps clarify why the XplainAct label continues to recur across otherwise heterogeneous research programs (Cope et al., 2023).

Across these uses, XplainAct names a persistent shift in explainability research: from retrospective description toward explanation that is embedded in dialogue, replanning, model manipulation, or intervention simulation. The precise technical realization varies widely—rule-based autonomy models, PDDL replanning, local SCMs, diffusion-layer bending, concept-level probability-of-sufficiency, or agentic retrieval—but the recurring ambition is stable: to make explanations usable at the point where a person must query, compare, modify, or decide.