Explicit-Action Modelling

• Explicit-action modelling is a method of representing actions as clear, human-readable symbols that enhance transparency and transferability in AI systems.
• It employs mapping techniques—ranging from natural-language spans to logical rule schemas—to facilitate robust error diagnosis and compositional planning.
• Its practical applications span dialogue systems, program synthesis, and symbolic planning, demonstrating improved sample efficiency and measurable performance gains.

Explicit-action modelling refers to representing the semantics, effects, or decision-making structure of actions in system models, learning algorithms, and intelligent agents in a directly interpretable, human-readable, or compositional form. This stands in contrast to latent-action approaches, where actions are encoded as unstructured or non-interpretable latent vectors. In explicit-action frameworks, each action is mapped to a discrete and semantically rich symbol, span, logical rule, or transformation, enabling direct analysis, explanation, composition, and generalization. Across modern AI—including task-oriented dialogue, planning, program synthesis, motion modeling, and conceptual behavioral modeling—the use of explicit-action representations systematically improves explainability, generalization, error diagnosis, and sample efficiency.

1. Conceptual Foundations: Explicit vs. Latent Action Representations

Explicit-action modelling emerged to address the limitations of latent-action methods that represent actions as unstructured embeddings (e.g., latent $z$ in VAEs), which suffer from opacity, poor compositionality, and domain overfitting (Huang et al., 2020). In contrast:

• Latent-action approaches: Introduce an unobserved variable $z$ for each action (e.g., dialogue turn), typically learned by reconstructing observable effects from $z$ and context. These suffer from entanglement, poor alignment with human-understandable actions, and limited transferability to new domains or tasks (Huang et al., 2020).
• Explicit-action approaches: Map each action or utterance to a discrete, interpretable representation—ranging from natural-language phrase spans and rule-based schemas to logical or symbolic models. This mapping usually leverages a finite vocabulary (e.g., domain slot types, action names, or content words) and mechanisms for extracting only the most salient tokens driving state transitions or decisions.

The explicit representation makes the plan, behavior policy, or procedural logic of a system transparent and compositional, facilitating error tracing (via intermediate symbols) and supporting out-of-distribution generalization by recombining basic action units (Huang et al., 2020).

2. Explicit-Action Model Architectures and Algorithms

Several canonical families of explicit-action models exist across domains:

a. Memory-Augmented Saliency and Compositional Spans for Dialogue

The MASP approach (Huang et al., 2020) in dialogue generation uses a key-value memory component with a pre-specified vocabulary (all slot names/values and content words) to summarize utterances into natural-language action spans. Queries derived from state-tracking models retrieve $k$ words by iterative attention hops, gated to variable length; these words form an explicit content plan $l(x)$ for the subsequent surface realization step. The model is trained with objectives for reconstruction (action span must sufficient for state update), compactness (matching gold state text spans), and dialogue state tracking stability.

b. Discrete Program Synthesis and Action Sequence Optimization

In model explanation and recourse, explicit actions are defined as domain-specified atomic feature modifications (e.g., "increase income by $500", "set employment to full-time") and sequences thereof (Ramakrishnan et al., 2019). The approach synthesizes minimal-cost sequences of such actions that will shift a classifier's output, using discrete search (beam search over programs) plus continuous adversarial optimization for parametric actions, subject to symbolic pre- and post-conditions.

c. Symbolic Planning and Propositional Action Models

Action models in planning (PDDL, STRIPS) are explicit in the sense that each action is a schema with well-defined preconditions and effects over a set of propositional variables or predicates (Arora et al., 2018, Bolander et al., 2015, Asai, 2019). Learning approaches either employ update-based restriction (pruning candidate operator schemas inconsistent with observations) or deep models (LSTM sequence tagging) to isolate the correct action rules that explain demonstration traces. Neural-symbolic methods can extract such explicit representations from raw data, including images, by learning discrete latent state variables and rule-structured classifiers mapping them.

d. Explicit-Action Version-Space Learning

Recent work formalizes learning explicit-action models as maintaining boundaries in precondition/effect hypothesis space, supporting sound (maximally specific) and complete (most general, nondeterministic) under-/over-approximations that provably converge to the true action model as more transitions are observed (Aineto et al., 15 Apr 2024). This process is online, supports both deterministic and non-deterministic action models, and enables confident inference under data uncertainty.

e. Explicit Behavioral and Conceptual Modeling

In conceptual and software engineering modeling, a minimal set of primitive explicit actions (e.g., "create", "process", "release", "transfer", "receive") is sufficient to reconstruct the semantics of system behaviors, business processes, or software workflows (Al-Fedaghi, 2022). The thinging machine (TM) framework formalizes all behaviors as compositions of these primitives, enabling cross-disciplinary mappings to UML/BPMN and advocating transparency in event and action decompositions.

3. Empirical Impact and Performance Characteristics

Empirical analyses across domains consistently demonstrate several advantages for explicit-action modeling:

• Sample efficiency and robustness: MASP achieves higher "Inform" and "Success" rates on dialogue benchmarks (MultiWOZ) than latent-action models, especially in low-data and transfer (novel domain) regimes (Huang et al., 2020).
• Interpretability and error diagnosis: Intermediate explicit spans (e.g., { "request", "departure" }) decouple content planning and surface realization, enabling localization of failures in human-interpretable terms (Huang et al., 2020).
• Model completeness and learnability: In symbolic planning, explicit-action hypotheses can be finitely identified for deterministic actions—any consistent action model can be exactly recovered from finite observation sets, an impossible task for non-deterministic/latent approaches (Bolander et al., 2015).
• Practical Synthesis: Synthesis of action sequences for classifier recourse (e.g., credit prediction) finds minimal, valid modifications that are cheaper and shorter than those found by greedy or purely adversarial methods, demonstrating the utility of explicit design of actionable transformations (Ramakrishnan et al., 2019).

The following table summarizes key findings from several explicit-action learning frameworks:

Domain	Explicit-Action Representation	Key Result or Metric	Comparative Baseline	Gain
Dialogue (Huang et al., 2020)	Spans over slot/content vocabulary	Inform: 70.2% (20% data, MASP)	MALA (latent): 63.5%	+6.7pp Inform (low data); +5.3pp in transfer
Planning (Asai, 2019)	Boolean symbolic rules (PDDL)	Per-bit successor accuracy ≈99%	Black-box NN: ≈99%	Competitive, plus explicit rule extraction
ModRecourse (Ramakrishnan et al., 2019)	Sequences of feature-modification actions	Success: 98% (credit tasks)	Greedy: 85%	20–40% reduction in action costs

4. Theoretical Properties: Learnability, Guarantees, and Limitations

a. Learnability and Convergence

• Finite identifiability: Deterministic propositional action models are finitely identifiable from transitions; after seeing a "defining tell-tale" set of state transitions, the hypothesis space collapses to a unique model (Bolander et al., 2015). Non-deterministic actions, however, are only identifiable in the limit.
• Version-space convergence: In the fully observable case with a finite action/environment vocabulary, both sound (specific) and complete (generalized, possibly non-deterministic) models are uniquely defined by the boundaries of version spaces and provably converge to the true model given sufficient data (Aineto et al., 15 Apr 2024).

b. Limitations

• Vocabulary restriction: Fixed vocabularies may limit coverage of rare or paraphrased actions (Huang et al., 2020).
• Complexity and scaling: Explicit rule extraction from latent (e.g., image-based) representations can yield large, entangled rule sets that overwhelm standard symbolic planners (PDDL→SAS+ translation bottlenecks) (Asai, 2019).
• Ambiguity in high-variance or under-specified domains: When supervision is minimal or contextual confounds are strong (as in weakly supervised video localization), explicit subspace decomposition requires access to reliable stream-based features (Liu et al., 2021).

5. Applications and Extensions

Explicit-action models are foundational in several key areas:

• Dialogue systems: Enabling explainable and generalizable task-oriented dialogue agents via explicit span-based action plans (Huang et al., 2020, White et al., 2023).
• Planning and model acquisition: Extraction of PDDL/STRIPS action schemas from symbolic trajectories, images, or noisy demonstrations for application in classical and neural-symbolic planners (Arora et al., 2018, Bolander et al., 2015, Asai, 2019).
• Recourse and decision reversal: Generating user-understandable action plans for changing algorithmic decisions under domain constraints (Ramakrishnan et al., 2019).
• Activity recognition: Disentangling action from context or motion for improved classification and localization performance in visual domains (Zhuang et al., 21 Oct 2025, Liu et al., 2021).
• Conceptual and software modeling: Providing a minimal, interpretable basis for behavioral semantics across UML, BPMN, and static/dynamic process views (Al-Fedaghi, 2022).

Several extensible directions have been suggested: moving to structured or hierarchical action representations, expanding/varying the vocabulary through unsupervised mining, integrating explicit-action models into end-to-end reinforcement learning with human-interpretable rewards, and developing planning/translation engines capable of handling large-scale Boolean or logic circuit representations directly (Huang et al., 2020, Asai, 2019).

6. Synthesis and Outlook

Explicit-action modelling synthesizes decades of progress in symbolic reasoning, program synthesis, interpretable ML, and cognitive modeling by structuring agent behavior around discrete, compositional, and human-comprehensible primitives. This enables transparent system design, accountable automated decisions, and robust transfer across domains and tasks. Remaining open challenges encompass representation scalability, ability to handle paraphrastic richness and unstructured environments, and closing the loop between neural representation learning and explicit action model extraction for complex, high-dimensional tasks.

---

Principal Sources:

• "Generalizable and Explainable Dialogue Generation via Explicit Action Learning" (Huang et al., 2020)
• "Synthesizing Action Sequences for Modifying Model Decisions" (Ramakrishnan et al., 2019)
• "Action Model Acquisition using LSTM" (Arora et al., 2018)
• "Learning Action Models: Qualitative Approach" (Bolander et al., 2015)
• "Neural-Symbolic Descriptive Action Model from Images: The Search for STRIPS" (Asai, 2019)
• "Action Model Learning with Guarantees" (Aineto et al., 15 Apr 2024)
• "Conceptual Modeling of Actions" (Al-Fedaghi, 2022)
• "Leveraging Explicit Procedural Instructions for Data-Efficient Action Prediction" (White et al., 2023)
• "A Renaissance of Explicit Motion Information Mining from Transformers for Action Recognition" (Zhuang et al., 21 Oct 2025)
• "Weakly Supervised Temporal Action Localization Through Learning Explicit Subspaces for Action and Context" (Liu et al., 2021)
• "Learning Partially Observable Deterministic Action Models" (Amir et al., 2014)