Cognitive & Metacognitive Acts in Intelligent Systems

Updated 27 January 2026

Cognitive and metacognitive acts are distinct yet interdependent components where cognitive acts perform perception, reasoning, and action, while metacognitive acts monitor and adjust these processes.
They are implemented in architectures like MIDCA to detect expectation failures and trigger adaptive meta-level interventions based on performance discrepancies.
Their integration enhances robustness and adaptability in autonomous agents and human-AI systems, leading to improved risk management and self-improvement capabilities.

Cognitive and Metacognitive Acts

Cognitive and metacognitive acts are distinct but interdependent components within both human and artificial intelligence. Cognitive acts constitute the primary processes by which agents perceive, represent, reason, and act on the world. Metacognitive acts operate at a higher level, monitoring, regulating, and modifying cognitive processes to improve performance, adaptiveness, and robustness. This architecture is manifest in cognitive systems such as MIDCA and is pivotal to advanced autonomous agents and human-AI co-creative systems. Below, the formal definitions, architectural patterns, mechanistic implementations, and empirical manifestations of these acts are detailed, together with their role in adaptive behavior and agent self-improvement.

1. Formal Distinction: Cognitive vs. Metacognitive Acts

Cognitive acts are the basic operations of perception, evaluation, memory retrieval, planning, and action execution. Formally, a cognitive system may be modeled as a state-transition system $E=(S, A, \gamma)$ , where $S$ is the set of world states, $A$ the set of actions, and $\gamma: S \times A \rightarrow S$ the successor function. A planning problem $P=(E, S_0, g)$ seeks a plan $\pi = (a_1, \ldots, a_n)$ achieving $g$ from $S_0$ (Cox et al., 2022).

Metacognitive acts are “cognition about cognition”—second-order operations that monitor, evaluate, and control the cognitive loop itself. These acts include introspective monitoring (detecting expectation failures), meta-level goal formation, meta-planning, and execution of strategies that alter cognitive processes. In architectures like MIDCA, metacognitive acts are realized as a parallel loop:

Cognitive cycle: Perceive $\rightarrow$ Interpret $\rightarrow$ Plan $\rightarrow$ Act
Metacognitive cycle: Monitor $\rightarrow$ Interpret $\rightarrow$ Evaluate $\rightarrow$ Intend $\rightarrow$ Plan $\rightarrow$ Control (Amos-Binks et al., 2019, Cox et al., 2022)

Taxonomy Example:

Level	World Axis	Modeller Axis
Object-level	$W^{\bullet}$	$M^{\bullet}$
Meta-level	$W^{\bullet\bullet}$ (meta-model of world model)	$M^{\bullet\bullet}$ (meta-model of modeller state)

At the cognitive level, the agent operates directly on states and actions; at the metacognitive level, the agent operates on histories of cognitive acts and meta-knowledge to produce adjustments or meta-actions (0807.4417).

2. Mechanisms and Representation in Cognitive Architectures

Cognitive architectures explicitly operationalize the distinction and interaction between cognitive and metacognitive acts. In MIDCA, metacognitive acts use traces of cognitive activity, stored as sequences of mental states $s^M$ and mental actions $a^M$ , to detect expectation failures and trigger meta-level goals (Cox et al., 2022). A typical architecture involves:

Cognitive loop: Direct execution of plans, (e.g., in a plant-protection domain, spraying weeds).
Metacognitive monitoring: Comparing observed transitions $(s_j, a_j, s_{j+1})$ to expected transitions $E^M$ ; triggering meta-goal formation when discrepancies are found.
Meta-level plan synthesis: Building meta-plans (e.g., “perform-learning”) that, if necessary, invoke learning algorithms to repair flawed action models.

Similarly, anticipatory thinking (AT) operates at the boundary: the agent reflects on a generated object-level plan, analyzes precondition vulnerability, anticipates possible plan failures (conditioning events), and inserts anticipatory actions to reduce risk prior to execution. Numeric risk assessment of these meta-level interventions quantifies recall and cost-benefit improvements (Amos-Binks et al., 2019).

3. Metacognitive Control, Monitoring, and Learning

Metacognitive control encompasses mechanisms by which agents not only introspect on their cognitive traces but also intervene adaptively. The computational process involves:

Monitoring: Extracting traces $t=(s_0, a_0, s_1, \ldots, s_k)$ from the cognitive cycle.
Meta-evaluation: Detecting when $w(s_j, a_j) \neq s_{j+1}$ for the cognitive transition function $w$ .
Goal selection and meta-planning: When failures arise, meta-explanation patterns generate goals such as switching reasoning methods, altering goals, or, most crucially, repairing knowledge via learning (e.g., FOIL learns missing action effects).
Meta-control: Applies learned changes or adjustments back to the cognitive layer's planning and execution modules (Cox et al., 2022).

Concrete implementations include self-regulation (modifying policy selection based on self-assessed competence), metacognitive sensitivity (quantifying the reliability of confidence in predictions via meta- $d'$ ), and active metacognitive particles that engage in mental action to update lower-level belief parameters (Valiente et al., 2024, Trinh et al., 11 Dec 2025, Sandved-Smith et al., 2024).

4. Cognitive and Metacognitive Acts in Adaptive and Co-creative Systems

Metacognition has been extended from classic planning domains to agentic adaptation in novel or ill-structured environments:

Agents in unknown environments: The MUSE framework integrates dynamically learned competence awareness and self-regulation; agents estimate $C_t(\pi) \approx P(\textrm{success}|\textrm{history},\pi)$ for candidate plans and select strategies maximizing predicted competence, not just reward (Valiente et al., 2024).
Generative AI design workflows: Cognitive acts include intent formulation, problem exploration, design generation, and evaluation. Metacognitive scaffolds—Socratic questioning, planning checklists, reflection prompts—enable designers to plan, self-monitor, reflect, or recalibrate strategy, significantly improving design feasibility (Gmeiner et al., 15 Jun 2025).
Knowledge editing in MLLMs: Cognitive-level editing modifies model parameters to fit new facts. Meta-cognitive editing layers—self-awareness (tracking “what” was changed), boundary monitoring (tracking “when” to apply changes), and reflective label refinement—allow models to reason about the scope, impact, and reliability of edits (Fan et al., 6 Sep 2025).

5. Empirical and Algorithmic Examples

Metacognitive acts are directly measurable and operationalized in various settings:

Learning performance: In a plant-protection planning domain, agents with metacognitive learning achieve significantly higher goal rates (up to $\sim$ 85%) versus non-metacognitive baselines ( $\sim$ 40%), underlining the utility of metacognitive learning-goal operations (Cox et al., 2022).
Self-monitoring in memory: Humans' judgments of learning (JOLs) are accurate meta-predictions of future memory, but GPT-based LLMs demonstrate no such item-level predictive correlation—highlighting the mechanistic absence of meta-level monitoring in current AI models (Huff et al., 2024).
Decision-making under uncertainty: Metacognitively-rational agents optimize their sampling distributions $q^\ast_{\mathrm{meta}}(o)$ given resource constraints, explaining availability bias and the fourfold pattern of risk preference as outcomes of meta-level optimization under bounded resources (Nobandegani et al., 2018).

Algorithmic sketches (see below) clarify the computational locus of metacognition:

def metacognitive_cycle(cognitive_trace, normative_model):
    m = monitor(cognitive_trace)
    delta = eval_meta(m, normative_model)
    if delta > threshold:
        c_next = control(delta, cognitive_trace)
        return c_next
    return cognitive_trace

(Kennedy, 2022)

6. Theoretical Generalizations and Applications

The cognitive vs. metacognitive act distinction generalizes across domains and agent types:

Resource-rational metareasoning formalizes the cost-benefit calculation of meta-level acts: for meta-action $m \in M$ , maximize $\mathbb{E}[\textrm{utility}] - \textrm{cost}(m)$ (Johnson et al., 2024).
Hierarchical modular systems: In the CRMN model, a global “responsibility signal” $\alpha_{ik}$ , derived from lower-level prediction and reward errors, selects which sub-modules are subject to metacognitive updating and hence conscious access (Kawato et al., 2021).
Knowledge taxonomies: Cognitive acts operate on $W^{\bullet}$ (world models) and $M^{\bullet}$ (modeller state); metacognitive acts operate on $W^{\bullet\bullet}$ and $M^{\bullet\bullet}$ (meta-models about domain and self) (0807.4417).
Self-regulated learning: Cognitive acts (e.g., “describe,” “compare”) are sequenced and abstracted to higher-level metacognitive strategies via mining of clickstream or problem-solving traces (Tian et al., 2019).

7. Significance and Implications

Distinguishing and integrating cognitive and metacognitive acts has systemic implications:

Robustness and adaptability: Metacognitive monitoring enables agents to detect unfamiliar or high-risk situations, triggering risk mitigation, additional learning, or dynamic adjustment of strategy (Amos-Binks et al., 2019, Valiente et al., 2024).
Explainability and alignment: Metacognitive modules facilitate the construction of explanations (“why was this strategy chosen?”), intellectual humility (recognizing knowledge gaps), and explicit calibration of confidence—traits fundamental to safe, trustworthy AI (Johnson et al., 2024, Trinh et al., 11 Dec 2025).
Human-AI interaction: In co-creative design, reflection scaffolds and metacognitive prompts improve not only the quality of cognitive output but also the depth of engagement and learning by human users (Gmeiner et al., 15 Jun 2025).
Theoretical grounding: The formalization of meta-level control—via monitoring, evaluation, and regulation—anchors a unified cross-disciplinary understanding spanning AI, cognitive science, and computational neuroscience.

This conceptual and technical apparatus provides a foundation for engineering autonomous, adaptive, and self-improving agents across domains, from metacognitive reinforcement learning to LLM-based reasoning, collaborative design, and beyond.