TMK Models: Task, Method, Knowledge

• TMK models are formal frameworks that distinctly define Tasks, Methods, and Knowledge to enable systematic, explainable AI with causal, teleological, and compositional dynamics.
• They integrate finite-state mechanisms and hierarchical task decomposition, supporting modular reasoning in applications like intelligent tutoring and multi-task learning.
• Empirical studies demonstrate that TMK-based architectures enhance factual correctness, causal coherence, and transfer performance across diverse AI systems.

A Task-Method-Knowledge (TMK) model is a formalism for representing and operationalizing the structure of agent behavior, systems design, or skill acquisition, by capturing the fundamental distinction between goals (Tasks), mechanistic realization (Methods), and domain grounding (Knowledge). TMK models encode explicit relations among these components—specifically, causal transitions (state-change dynamics), goal hierarchies (teleology), and decompositions (compositionality)—to support interpretability, systematic explanation, and compositional reasoning in AI systems. TMK underlies recent advances in explainable AI, multi-task learning, knowledge-based AI, and parameter-efficient transfer, as exemplified by its adoption in intelligent tutoring systems, interactive agents, and deep learning transfer frameworks.

1. Formal Structure of TMK Models

A TMK model is commonly defined as a tuple

$$\mathrm{TMK} = (T, M, K, R_{\text{causal}}, R_{\text{goal}}, R_{\text{decomp}})$$

where:

• $T = \{t_1, t_2, \dots\}$: the set of Tasks, representing goals or subgoals.
• $M = \{m_1, m_2, \dots\}$: the set of Methods, each a procedural mechanism.
• $K = \{\kappa_1, \kappa_2, \dots\}$: the set of Knowledge entities (concepts, relations, instances).
• $R_{\text{causal}} \subseteq S \times S$: the state-transition relation, typically defined within each Method as a guarded finite-state machine.
• $R_{\text{goal}} \subseteq T \times T$: encodes task hierarchies (goal/subgoal relations).
• $R_{\text{decomp}} \subseteq T \times M$: maps each Task to one or more implementing Methods (Dass et al., 26 Nov 2025).

In many instantiations, Tasks are further described as tuples $(G, I, \text{Pre}, \text{Post}, O)$, denoting a logical goal, inputs, pre- and post-conditions, and explicit outputs (Dass et al., 10 Apr 2025). Methods are modeled as finite-state machines: $$m = (S_m, s_0, S_{\text{succ}}, S_{\text{fail}}, \Delta_m)$$, with guarded transitions $\Delta_m \subseteq S_m \times C \times S_m$, $C$ being boolean predicates over $K$. Knowledge forms a supporting ontology of concepts, properties, and axioms referenced by both Tasks and Methods.

TMK formalisms may explicitly include a vocabulary component $V$ (TMKV) to support organizational and glossary functions (Goel et al., 2022).

2. Encoding Causality, Teleology, and Decomposition

• Causality ($R_{\text{causal}}$): Methods are formalized as finite-state machines, where transitions $(s_i, \text{guard}_{i,j}, s_j)$ are only permitted if guard conditions (predicates over Knowledge) are satisfied. This models how actions or operations effect state change, enforcing explicit causal justification at the symbolic level (Dass et al., 26 Nov 2025).
• Teleology ($R_{\text{goal}}$): Tasks are organized as a tree or DAG, with edges denoting subgoal relationships. Each method state can be tagged with a reference indicating which subgoal is advanced, supporting stepwise teleological linkage within explanations (Dass et al., 26 Nov 2025, Dass et al., 10 Apr 2025).
• Compositionality ($R_{\text{decomp}}$): Tasks decompose into one or more Methods, and complex Tasks may invoke sub-Tasks. Decomposition functions $\delta: M \rightarrow 2^T$ and mappings $\mu: T \rightarrow 2^M$ make the hierarchical structure explicit, supporting scalable, modular reasoning (Goel et al., 2022, Dass et al., 10 Apr 2025).

TMK thereby enables formal tracing—at each decision or explanation step—of not only what is done, but why (relative to goals), and how (via procedural, causal, and compositional breakdown).

3. TMK in Hybrid Symbolic–Generative Architectures

Recent research operationalizes TMK models within hybrid symbolic-generative architectures, such as Ivy and VERA (Dass et al., 26 Nov 2025, Sushri et al., 2024), in which a two- or three-layer pipeline enforces fidelity between underlying symbolic models and natural language explanations:

• Symbolic Control Layer: Encodes $T$, $M$, $K$ and associated relations, performs question scope classification, and retrieves relevant fragments (Task, Method, and Knowledge).
• Generative Interpretation Layer: An LLM or similar generative model is constrained to synthesize explanations that mention only valid transitions (from $R_{\text{causal}}$), relate steps to their parent goals ($R_{\text{goal}}$), and decompose as specified in $R_{\text{decomp}}$. Constraint enforcement is performed via hard template prompting, rejection filtering, or reranking.

A canonical four-stage pipeline includes (1) scope classification, (2) structured retrieval, (3) constrained generation, and (4) coherence optimization. Example pseudocode for enforcement of TMK constraints in generation is provided in (Dass et al., 26 Nov 2025).

VERA’s Ask-TMK module blends TMK formalism with LLM-based retrieval and generation using question classification, FAISS-based nearest-neighbor search, and Chain-of-Thought prompting for method explanations. Only information present in the TMK store is exposed to generation, yielding high factual fidelity (Sushri et al., 2024).

4. Empirical Applications and Case Studies

Skill Learning and Educational Agents:

In Ivy, TMK models are used to power AI coaching systems that deliver explanations with explicit structural, causal, and teleological content. Empirical evaluation demonstrates substantial improvements in explanation quality, as measured by expert annotation on dimensions such as causal coherence (Ivy+TMK-Structured: 1.6, baseline GPT: 0.7), teleological linkage, and compositional decomposition. Factual correctness rates (65%) considerably exceed unconstrained baselines (15–35%), and novice annotators display high agreement with expert-expected structure (F1 ≈ 0.75–0.85) (Dass et al., 26 Nov 2025).

Interactive Agents and Self-Explanation:

Skillsync’s AskJill leverages a TMK model comprising structured goal, method, knowledge, and vocabulary tables. The agent answers natural-language queries about tasks and system vocabulary with near-perfect accuracy (100% in vitro, 91% in situ), organizing stakeholder questions along the TMK ladder and providing domain-precise responses (Goel et al., 2022). VERA demonstrates that TMK-structured self-explanation modules can achieve high recall, precision, and accuracy across a large and diverse question set in intelligent learning environments (Sushri et al., 2024).

Multi-task Deep Learning and Parameter-efficient Transfer:

In MultGT (Zhao et al., 2023), TMK principles are integrated into a multi-task Graph Transformer, with task-specific “knowledge injection” and domain-driven pooling modules. Tasks (tumor typing, staging), methods (graph-transformer with task-aware heads), and knowledge (learned task subspace, domain pooling rules) are cleanly separated, yielding superior performance across all tasks.

The aTLAS algorithm (Zhang et al., 2024) represents an instantiation of TMK for task vector arithmetic in neural networks: each fine-tuning dataset defines a “Task” via its task vector, a compositional “Method” (aTLAS scaling and combination), and a resulting “Knowledge” embedding. aTLAS empirically improves disentanglement and generalization in few-shot and multi-task scenarios.

5. Principles, Best Practices, and Design Guidelines

Theoretical grounding for TMK rests on three foundational principles:

• Teleological Principle: Every method is selected with explicit reference to the task’s goal formula.
• Causal Principle: Procedural steps (state transitions) are justified by knowledge-based preconditions and effects, enforcing a cause-effect structure.
• Compositional Principle: Tasks and methods are hierarchically decomposable, and modular explanations are composable via explicit sub-task and sub-method linkages (Dass et al., 10 Apr 2025).

Best-practice guidelines for TMK modeling include:

1. Explicit enumeration of task goals, inputs, pre-/post-conditions, and outputs.
2. Full specification of method FSMs with well-defined transitions and guard predicates.
3. Construction of a sufficiently expressive ontology of domain knowledge.
4. Hierarchical decomposition into sub-TMK triples.
5. Peer review of models for coverage and consistency (Dass et al., 10 Apr 2025).

6. Domain Scope and Current Limitations

TMK models have been deployed in educational technology, agent-based systems, explainable multi-task deep learning, and parameter-efficient adaptation. Extant empirical evaluations demonstrate high gains in explainability, coverage, user satisfaction, and compositional generalization.

Limitations include the scalability of constructing TMK models in domains with very large numbers of methods or rapidly changing knowledge, the coverage of procedural explanations in domains where methods are not well-understood, and preliminary evaluations relying mainly on internal developer annotation rather than large-scale user studies (Dass et al., 26 Nov 2025, Sushri et al., 2024). A plausible implication is need for further work in automated TMK extraction from system codebases or logs, and broader external validation.

7. Representative TMK Model Variants

System/Paper	Tasks (T)	Methods (M)	Knowledge (K)
Ivy (Dass et al., 26 Nov 2025)	Skill goals, subgoals	FSMs with causal guards	Domain predicates, logical state configs
Skillsync (Goel et al., 2022)	User-facing workflow tasks, goals	UI/action procedures, subtasks	Structured KB tables, vocab glossary
MulGT (Zhao et al., 2023)	Tumor typing, staging	Graph-Transformer, pooling	Latent tokens, domain-driven rules
aTLAS (Zhang et al., 2024)	Task vectors (fine-tunes)	Block-scaling algorithms	Pretrained representations, task vectors
VERA (Sushri et al., 2024)	Science inquiry tasks, modeling	FSMs for agent operations	JSON concepts, experiment definitions

Each system makes the T–M–K separation explicit, enables stepwise compositional explanations or transfer, and enforces structure either in symbolic pipelines, hybrid generation, or parameter composition. The explicit modeling of $R_{\text{causal}}$, $R_{\text{goal}}$, and $R_{\text{decomp}}$ is a recurring feature across the landscape.

---

For details on TMK definitions and empirical assessments in hybrid architectures, see (Dass et al., 26 Nov 2025); skill-based QA (Dass et al., 10 Apr 2025); design-based agent explainability (Goel et al., 2022); multi-task learning (Zhao et al., 2023); parameter-efficient transfer learning (Zhang et al., 2024); and self-explaining agents (Sushri et al., 2024).