Human-Centered Symbolic Explanations

• Human-centered symbolic explanations are transparent, ontology-grounded representations of AI reasoning that map user queries to explicit, causal inference chains.
• They utilize symbolic architectures, semantic frames, and rule-based logic to generate detailed, step-by-step reasoning traces.
• These explanations enhance troubleshooting, decision-making, and trust by exposing every reasoning step through user-tailored interfaces.

Human-centered symbolic explanations are formally grounded, transparent, and user-aligned representations of AI agent reasoning, designed to enable genuine understanding, accountability, and trust for human operators and stakeholders. Unlike opaque black-box rationales or shallow post hoc “explanations,” human-centered symbolic explanations articulate the actual internal state and inference process of an agent, using domain ontologies, explicit plans, interpretable rules, or logical abstractions. The defining criterion is that every step in the reasoning chain is accessible to human inspection and can be causally or correlationally related to observed behavior, supporting robust error detection, operational safety, and shared decision-making responsibility (Nirenburg et al., 2024, Watkins et al., 27 Feb 2025, Wu et al., 2023).

1. Defining Principles of Human-Centered Symbolic Explanation

The foundational requirement is a mapping from user-oriented questions to explicit, human-readable accounts of an agent’s internal symbolic reasoning—specifically as causal chains or correlational summaries in domain-relevant terms. Three non-negotiable criteria distinguish this class:

• Ontologically Grounded World Model: A comprehensive domain ontology (e.g., including high-level task, object, and agent types) forms the basis for all representations.
• Explicit Reasoning Traces: Full disclosure of the state transitions, plan hierarchies, precondition/effect propagation, or rule firings involved in accomplishing a task.
• User-Tailored Presentation: Explanations must answer user queries with relevant detail, in a format that is actionable and context-sensitive.

These principles ensure that explanations are not ad hoc rationalizations, but direct exposures of the agent's epistemic process. For instance, answering a question such as “Why did you ask ‘What do they look like?’” involves showing the parsed Text-Meaning Representation (TMR), the Agenda of unmatched preconditions, and the resulting plan structure (Nirenburg et al., 2024).

2. Formal Representation and Reasoning Frameworks

Human-centered symbolic explanations require formal machinery that supports full introspection and traceability. Exemplars include:

• Symbolic Knowledge Architectures: Such as the LEIA architecture, which uses a two-tiered representation combining domain ontologies and fully explicit plan libraries:
  • Entities: Concepts (#C.i), properties (slot–value), and structured relations.
  • Text-Meaning Representation (TMR): Structured records for each user utterance (e.g., { type: REQUEST-ACTION, agent: #HUMAN, theme: SEARCH-FOR-LOST-OBJECT, ... }).
  • Visual-Meaning Representation (VMR): Symbolic coding of perceptual features (type, subclass, color, etc.).
  • Plan Library: Tuples of the form $$P_k = (\mathrm{name}_k,\,\mathrm{Pre}_k,\,\mathrm{Eff}_k)$$, with backward-chaining inference for goal satisfaction.
• Semantic Frames and Symbolic Graphs: In human–robot and multimodal AI, explanations reduce to “semantic frames”—explicit slot–value graphs mapping actions, tools, and receivers, constructed via LLMs from multimodal streams (Watkins et al., 27 Feb 2025).
• Symbolic Rules and Fuzzy Logic: For visual activity reasoning, explanations invoke a symbol–rule hypergraph where symbols (English phrases) and rules (if–then hyperedges) describe activity semantics. Fuzzy logic combines symbol detection probabilities for rule satisfaction (Wu et al., 2023).
• Logical Abstractions and Irrelevance Formalizations: In logic programming explanations (e.g., ASP), abstraction operators like removal and clustering of irrelevant atoms yield simplified, cognitively efficient explanations without loss of correctness (Saribatur et al., 3 Feb 2026).

Each of these representations supports runtime queries about why a system acted, what knowledge led to a recommendation, and how perceptual or data-driven modules contributed.

3. Explanation Workflows and Interaction Design

Workflows for generating human-centered symbolic explanations are characterized by stepwise integration of user interaction, automatic reasoning, and machine learning as a supporting—not dominant—modality. Canonical steps include:

1. User Input Parsing: Natural language utterances or interface signals are semantically parsed into TMR, semantic frames, or predicates, rooted in the domain ontology (Nirenburg et al., 2024, Watkins et al., 27 Feb 2025).
2. Goal/Plan Management: Parsed themes become goals on a symbolic Agenda, with plan selection and precondition checking following classical AI planning norms.
3. Perceptual/ML Module Mediation: Low-level perception is handled by ML modules, but their outputs are elevated to symbolic constructs (e.g., VMR, symbolic predicates) via lightweight classifiers or LLM-based frame fusion (Watkins et al., 27 Feb 2025).
4. Action Execution and Logging: All invocations, transitions, and perceptual matches are logged symbolically.
5. Explanation Generation on Demand: At any point, under-the-hood panels or query interfaces reveal relevant TMR/VMR, Agenda/plan trees, rules fired, and natural language paraphrases of the symbolic trace (Nirenburg et al., 2024).
6. Bidirectionality and Correction: In human–robot interaction, explanations prompt human feedback, enabling mutual alignment and model updating (Lera et al., 8 Apr 2025).

For domain-specific applications (e.g., robotics, software agent patching, or visual scene understanding), additional formal elements such as executable property-based tests, infection/outcome conditions (Kang et al., 30 Jul 2025), and time-series symbolic pattern swaps (Płudowski et al., 28 Mar 2025) may be included to maximize fidelity and user actionability.

4. Practical Implementations and Human-Interface Modalities

Symbolic explanations are made practically accessible through multimodal and visually structured interfaces:

• Under-the-Hood Panels: Parallel, scrollable panels displaying TMR (for user utterance parsing), VMR (for perception), Agenda (goal/plan status, with unmet preconditions highlighted), and “Thoughts” (natural language paraphrase of reasoning), as in the LEIA deployment (Nirenburg et al., 2024).
• Semantic Frame Displays: Slot–value graphical displays indicating action roles, tool usage, and receiver elements, with missing slots flagged for user input (Watkins et al., 27 Feb 2025).
• Pictogram Boards (Asterics Grid + ARASAAC): For diverse and accessibility-critical users, symbolic explanations use highly-iconic pictograms mapped to semantic concepts, organized into action, context, and query categories, supporting both robot-to-user and user-to-robot explanations (Lera et al., 8 Apr 2025).
• Executable Certificates: In coding agents, explanations instantiate as executable property-based tests and quantifier-free predicates, giving both symbolic and empirical evidence of causal chains for developer inspection and automated patch validation (Kang et al., 30 Jul 2025).

These interfaces support rapid, real-time explanation requests, enable human-in-the-loop corrections, and promote user adaptability by tailoring level-of-detail and terminology to user profiles and contexts.

5. Human-Centered Effects, Empirical Validation, and Design Criteria

Human-centered symbolic explanations deliver empirically verified benefits across understanding, operational efficiency, and trust. Metrics and results include:

• Transparency and Accountability: All reasoning steps and perceptual interpretations are open to inspection and directly attributable to agent knowledge sources and explicit rules (Nirenburg et al., 2024).
• Cognitive Load and Effectiveness: Cognitive studies show that clustering irrelevant details in symbolic explanations significantly improves participant understanding (accuracy), while removal of irrelevant details reduces cognitive effort (faster responses), as measured in randomized experimental settings using Answer Set Programming explanations (Saribatur et al., 3 Feb 2026).
• User Preference and Task Performance: User studies with human–robot and symbolic planning systems demonstrate strong user preference and better task completion times for symbolic versus black-box or saliency-based explanations (Sreedharan et al., 2020, Lera et al., 8 Apr 2025).
• Faithfulness and Fidelity: The “fidelity” metric quantifies the congruence of symbolic explanations with the agent’s internal, possibly sub-symbolic, logic, with domain-specific results reporting fidelities above 0.8 for chemically meaningful classes (Himmelhuber et al., 2021).
• Participatory and Adaptive Design: Participatory workshops and co-design with domain experts (e.g., clean-room technicians) directly shape role ontologies, template vocabulary, and interruption rules in semantic frame–driven systems, improving real-world adoption and trust (Watkins et al., 27 Feb 2025).

A synthesis of design guidelines emerges: anchor explanations in abstract but task-relevant ontologies, maintain the minimal sufficient detail for understanding and action, provide mechanisms for iterative user correction, and empirically validate abstraction operators to avoid over-pruning of meaningful detail (Saribatur et al., 3 Feb 2026).

6. Domains of Application and Limitations

Human-centered symbolic explanations have been instantiated in broad domains:

• Autonomous Robotics: Providing explicit plan-based traces, goal–precondition–effect trees, and transparent user-facing reasoning for tasks such as search-and-rescue or clinical assistants (Nirenburg et al., 2024).
• Industrial Multimodal Systems: LLM-driven semantic-framing of multimodal input to support collaborative task execution and error correction in manufacturing (Watkins et al., 27 Feb 2025).
• Human–Robot Interaction for Accessibility and Pedagogy: ARASAAC-driven pictogram board explanations for neurodiverse users, deployed in education and assistive robotics (Lera et al., 8 Apr 2025).
• Visual Activity Understanding: LLM-extracted, diagrammatic, and fuzzy rule-based symbolic explanations for human activity in visual scenes (Wu et al., 2023).
• Code Analysis Agents: Property-based test generation and symbolic infection chain explanations for software repair and patch validation (Kang et al., 30 Jul 2025).
• Structured Decision Support: Abstraction-based symbolic explanations for planning, sequential decision-making, and post hoc contestability (Sreedharan et al., 2020).
• Time Series, Music, and Science: Counterfactual reasoning with symbolic pattern swaps or graph edit operations for time series and symbolic music graphs (Płudowski et al., 28 Mar 2025, Hilaire et al., 30 Sep 2025).

Reported limitations include the necessity for robust and relevant domain ontologies, sensitivity to over-abstraction or under-explanation, and dependence on participatory tuning of explanation detail-levels. Scalability and interface adaptation for highly complex or real-time settings remain open challenges.

---

Key References:

(Nirenburg et al., 2024, Watkins et al., 27 Feb 2025, Wu et al., 2023, Lera et al., 8 Apr 2025, Kang et al., 30 Jul 2025, Sreedharan et al., 2020, Saribatur et al., 3 Feb 2026, Himmelhuber et al., 2021)