Cognitive Taxonomy Overview
- Cognitive taxonomy is a structured classification of mental operations, organizing thinking into hierarchical and multidimensional frameworks.
- It underpins educational assessment frameworks like Bloom’s Taxonomy and supports automated evaluation tasks in software and AI research.
- Recent advances apply cognitive taxonomies in LLM internal probing, AGI profiling, and human–AI collaboration to enhance reasoning and performance.
Cognitive taxonomy is a structured classification of mental operations, faculties, or cognitive phenomena. In its classical educational form, it organizes thinking into ordered levels of complexity, most prominently Bloom’s progression from Knowledge to Evaluation; in later work it becomes two-dimensional, dual-mode, or domain-specific, and in contemporary AI research it also functions as a framework for prompt design, internal representation analysis, cognitive profiling, vulnerability assessment, and explanation evaluation (Kumar et al., 2010, Laddha et al., 2021, Raimondi et al., 19 Feb 2026, Burnell et al., 27 May 2026).
1. Classical foundations and conceptual scope
Bloom’s original cognitive taxonomy, as applied in software-design research, is hierarchical and dynamic: Knowledge, Comprehension, Application, Analysis, Synthesis, and Evaluation. The levels are associated with characteristic behaviors such as recall, identify, recognize, define, and describe at the base; interpret and translate at Comprehension; apply, construct, and solve at Application; analyze, categorize, and compare at Analysis; design, develop, formulate, and organize at Synthesis; and appraise, assess, judge, and evaluate at the highest level (Kumar et al., 2010). In that formulation, the taxonomy is not merely a fixed ladder. Complex tasks move iteratively and recurrently across levels as understanding deepens and decisions are made.
A later educational reformulation, Revised Bloom’s Taxonomy, preserves the ascending structure but separates two axes: six cognitive process categories—Remember, Understand, Apply, Analyze, Evaluate, Create—and four knowledge dimensions—Factual, Conceptual, Procedural, and Metacognitive. This changes the original one-dimensional, noun-oriented scheme into a matrix in which an assessment item may be located, for example, at Apply–Procedural or Analyze–Conceptual (Laddha et al., 2021). Across these versions, cognitive taxonomy serves two related purposes: ordering mental demand and specifying what kind of knowledge is being processed.
Not all later frameworks map directly onto Bloom, even when they resemble it. The Hierarchical Prompting Taxonomy, for example, is explicitly described as cognitively inspired and aligned with “principles that are crucial for humans to address the majority of reasoning and interpretation tasks,” but it does not claim a formal mapping to Bloom or SOLO (Budagam et al., 2024). This suggests that “cognitive taxonomy” has expanded from a single canonical hierarchy into a broader family of classificatory schemes for mental work.
2. Educational hierarchies and assessment frameworks
Educational applications remain the most visible setting for cognitive taxonomies. In automated assessment based on Revised Bloom’s Taxonomy, one study classified 844 Software Engineering examination questions into a 6-class cognitive process task and a 3-class knowledge-dimension task. On this dataset, a CNN achieved 80% testing accuracy for cognitive-process classification and 66.67% for knowledge-dimension classification, outperforming the reported LSTM results in generalization (Laddha et al., 2021). The result is important less as a universal benchmark than as evidence that taxonomy labels can be operationalized for large-scale assessment auditing.
A different line of work operationalizes cognitive progression through prompting rather than question labeling. The Hierarchical Prompting Taxonomy introduces four criteria—Basic Recall and Reproduction, Understanding and Interpretation, Analysis and Reasoning, and Application of Knowledge and Execution—and makes them actionable through a five-level Hierarchical Prompt Framework: Role Prompting, Zero-Shot Chain-of-Thought Prompting, Three-Shot Chain-of-Thought Prompting, Least-to-Most Prompting, and Generated Knowledge Prompting. Its evaluation metric is the HP-Score rather than the “HPI” sometimes attributed to the work. Across BoolQ, CommonSenseQA, IWSLT-2017 en–fr, and SamSum, manual level-wise prompt escalation produced lower HP-Scores and higher evaluation metrics than adaptive prompt selection, while adaptive HPF was reported as vulnerable to prompt-selector hallucinations (Budagam et al., 2024).
The educational benchmark EduEval extends this logic by unifying Bloom’s Taxonomy and Webb’s Depth of Knowledge into six dimensions: Memorization, Understanding, Application, Reasoning, Creativity, and Ethics. The benchmark contains 24 task types and 11,150 validated items across Chinese K–12 education, including exam questions, classroom dialogue, essays, and expert-designed prompts. Its empirical pattern is uneven rather than uniformly progressive: models perform strongly on factual tasks, struggle on classroom dialogue classification, show mixed behavior on multi-step reasoning, and often lose creative quality under few-shot prompting, while ethical scenario accuracy is comparatively robust (Ma et al., 29 Nov 2025).
3. Domain-specific cognitive taxonomies in technical practice
Outside education, cognitive taxonomy is often adapted to the structure of a specific professional task. In software design, the GIRA case study—“gprs based Intranet Remote Administration”—was used to argue that the system-design phase of the software development life cycle activates all six Bloom levels. The study mapped Knowledge to recalling prerequisites such as client-server architecture, Java, GPRS, RMI, sockets, HTTP, and design notations; Comprehension to interpreting requirements such as remote mobile administration; Application to constructing flowcharts, DFDs, pseudocode, and block diagrams; Analysis to decomposing the system into mobile, server, and client components; Synthesis to integrating J2ME/GCF, GPRS, HTTP, Tomcat, servlets, RMI, and sockets; and Evaluation to judging suitability and feasibility of design choices. Its correlation graph reported activity at all six levels, with higher frequency at Synthesis and Evaluation, supporting the claim that software design is a complex cognitive process (Kumar et al., 2010).
A broader software-engineering survey adopts a different sense of taxonomy: not levels of difficulty but top-level cognitive concepts. It organizes five decades of empirical SE research around perception, attention, memory, cognitive load, reasoning, cognitive biases, knowledge, social cognition, cognitive control, and errors, and also classifies qualitative and quantitative procedures used to assess them. Across 311 papers, the most developed areas were requirements, design, construction, and maintenance, while the overall state of the art was described as fragmented from the perspective of cognition (Fagerholm et al., 2022).
Command-line computing introduced a further adaptation. CogTax defines taxonomy level by combining cognitive complexity and operational impact through the rule . Its four levels range from safe read-only inspection to advanced system management, and its operational dimension distinguishes observational, reversible, structural, and administrative actions. On 585 expert-annotated Linux/bash commands, a classifier that fused AST-derived syntactic features with semantic embeddings achieved 89% accuracy, outperforming either representation alone (Alonso-Carracedo et al., 30 Jun 2026). A related grading study used a four-level taxonomy for Linux/bash examinations and showed that human–AI agreement declined as taxonomy level increased; under rubric-guided prompting, Gemini 3.0 Pro achieved ICC(3,1) = 0.888, MAE = 0.10, and Bland–Altman bias = -0.014 on 1200 real student responses (Alonso-Carracedo et al., 2 Jul 2026).
4. Cognitive taxonomy in LLM evaluation and AGI measurement
In recent LLM research, cognitive taxonomy is no longer only an external rubric for items or tasks; it is also treated as a property of internal representations. A mechanistic-interpretability study used Bloom’s Taxonomy as a hierarchy of prompt demand and linearly probed residual-stream activations in four open-weight decoder-only models. On a balanced corpus of 1,128 questions, multiclass logistic probes reached approximately 95% mean accuracy across Bloom levels, with a Cognitive Separability Onset around layer $l^\* \approx 5$ and errors concentrated between adjacent levels, summarized by . The authors emphasize, however, that linear decodability does not establish causal use of the decoded information during generation (Raimondi et al., 19 Feb 2026).
A more fine-grained reasoning taxonomy expands beyond Bloom entirely. “Cognitive Foundations for Reasoning and Their Manifestation in LLMs” defines 28 cognitive elements spanning reasoning invariants, meta-cognitive controls, knowledge representations, and transformation operations. Using 171,485 model traces from 17 models across text, vision, and audio, plus approximately 54 human think-aloud traces, it reports that humans more often exhibit hierarchical nesting and meta-cognitive monitoring, whereas models rely heavily on shallow sequential forward chaining. On ill-structured problems, the divergence is strongest. The same work reports that test-time reasoning guidance derived from successful cognitive subgraphs can improve performance by up to 60% on complex tasks (Kargupta et al., 20 Nov 2025).
At a higher level of abstraction, a proposed AGI evaluation framework decomposes general intelligence into 10 faculties: perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, and social cognition. It recommends held-out, targeted cognitive tasks, broad human baselines, and uncertainty quantification, then summarizes systems as cognitive profiles located within the human performance distribution (Burnell et al., 27 May 2026). This suggests a shift from single-score benchmarking toward profile-based intelligence measurement.
5. Human–AI systems, metacognition, and cognitive risk
As AI becomes embedded in knowledge work, several taxonomies reconceive cognition as distributed rather than purely individual. The Augmented Cognition Framework retains the six Bloom levels but gives each a distinct verb in two modes—Individual and Distributed—and adds a seventh level, Orchestration, devoted to mode-switching, trust calibration, degradation detection, and partnership optimization. Its central claim is asymmetrical: effective Distributed cognition typically depends on Individual cognitive foundations, though structured scaffolding with fading and verified transfer can invert that sequence in specific cases. The framework makes “fluent incompetence”—producing sophisticated outputs without genuine understanding—a central pedagogical risk (Ayodele et al., 31 Jan 2026).
Co-writing research proposes a different two-dimensional taxonomy. CoCo Matrix adapts Flower and Hayes’ cognitive process theory of writing and maps human–agent collaboration by entropy and information gain, using the standard definitions
This yields four quadrants: high entropy–high information gain, high entropy–low information gain, low entropy–high information gain, and low entropy–low information gain. In a review of 34 published systems, low entropy and high information gain was identified as under-explored but promising for settings in which the agent supports divergent planning while the human performs focused translation (Wan et al., 2024).
Professional metacognition has also been formalized at the level of whole learning scenarios. A six-node open-systems model with Environment, Input, Processes, Structures, Output, and Feedback generates 216 mathematically possible scenarios, reduced by four sequential filters to 24 priority scenarios distributed across novice, developing, and expert/adaptive tiers. The taxonomy emphasizes internal monitoring–control relations, cross-cluster entry and exit routes, and feedback topology, and it identifies multiple viable developmental trajectories rather than a single path from novice to expert (Gibson et al., 22 May 2026).
Cognitive taxonomy can also be used to organize failure modes. CCS-7 classifies seven vulnerabilities in LLMs—Authority Hallucination, Context Poisoning, Goal Misalignment Loops, Identity / Role Confusion, Memory / Source Interference, Cognitive-Load Overflow, and Attention Hijacking—and evaluates TFVA-style guardrails across 12,180 trials on seven model architectures. The paper reports that Identity Confusion is nearly fully mitigated across architectures, while Source Interference is markedly backfire-prone; for Mistral, false claim adoption rose from 0.40 to 0.88 under the tested guardrail condition (Aydin, 9 Aug 2025). In software development, a separate study organized 90 biases into 15 categories for developer–LLM interaction and found that 48.8% of observed programmer actions were biased, with developer–LLM interactions accounting for 56.4% of those biased actions (Zhou et al., 12 Jan 2026).
6. Evaluation challenges, controversies, and research directions
One recent development is the shift from taxonomies of cognition to taxonomies of explanatory intent. RF×G organizes saliency explanations along two axes—Reference-Frame, distinguishing pointwise from contrastive explanations, and Granularity, ranging from class-level to group-level semantics. It introduces four faithfulness metrics, CCS, CGC, PGS, and CGS, and evaluates ten saliency methods across four architectures and three datasets. The reported pattern is that IIA performs best across the new metrics, while group-level explanation is generally easier than fine-grained class-contrastive explanation (Elisha et al., 17 Nov 2025). This indicates that cognitive taxonomy can structure not only tasks or minds, but also the questions an explanation is supposed to answer.
Across the literature, several limitations recur. Some studies rely on a single observational case, as in the GIRA software-design analysis (Kumar et al., 2010). Others do not report inter-annotator agreement or detailed annotation rubrics, including the Bloom-level LLM probing study and the 844-question RBT classification study (Raimondi et al., 19 Feb 2026, Laddha et al., 2021). Prompting-based evaluation frameworks may lack statistical significance testing or depend on fragile prompt selectors (Budagam et al., 2024). Benchmarks with GPT-based scoring or strong curricular locality raise transfer questions, as in Chinese K–12 evaluation (Ma et al., 29 Nov 2025). Domain-specific taxonomies for command-line tasks or professional metacognition are powerful within their target settings, but their generalizability must be established empirically (Alonso-Carracedo et al., 30 Jun 2026, Gibson et al., 22 May 2026).
Proposed future directions therefore converge on several themes. Mechanistic work calls for causal interventions, alternative taxonomies such as SOLO or Anderson–Krathwohl, sequence-level representations, multilingual extensions, and cross-model transfer (Raimondi et al., 19 Feb 2026). Reasoning research argues for interactive benchmarks and stronger emphasis on meta-cognitive controls rather than easily counted sequential behaviors (Kargupta et al., 20 Nov 2025). AGI measurement work emphasizes private held-out tasks, broad human baselines, and uncertainty-aware cognitive profiles rather than a single scalar capability score (Burnell et al., 27 May 2026). Taken together, these developments suggest that cognitive taxonomy has become a general research instrument for specifying what kind of cognition is under study, at what level of abstraction, under which operational constraints, and for which evaluative purpose.