Tri-System Theory: Three-Level Architectures

Updated 6 July 2026

Tri-System Theory is a framework that partitions system competence into three distinct, interactive components to achieve higher adaptability and performance.
It underpins applications from cognitive modeling to AI and robotics by integrating genetic, adaptation, and neural or evaluative policies.
The theory offers enhanced computational capacity and robust control architectures, addressing limitations inherent in dual-process and single-module systems.

Tri-System Theory denotes a family of three-component or three-level explanatory frameworks in which system competence is attributed not to a single monolithic mechanism, but to the coordinated operation of three distinct subsystems, modes, or traverses. In the literature, the term is used most explicitly for a practopoietic theory of mind that models intelligence as a hierarchy of genetic policy, adaptation policy, and neural policy, yielding a T3-agent whose organization is written $T_{\mathrm{G} \to T_{\mathrm{A} \to T_{\mathrm{N} \to U}}$ (Nikolić, 2015). Closely related triadic formulations appear in general systems theory as the three gestalts of behavior, organization, and substance (Florio, 2014); in LLM control as Fast, Normal, and Slow thinking modes (Li et al., 6 Jun 2025); in long-horizon robotics as Brain, Cerebellum, and Critic (Yi et al., 5 Mar 2026); and in several other technical domains. Taken together, these works do not define a single universal doctrine. They do, however, converge on a common architectural claim: three differentiated components can provide capabilities that are difficult to obtain from a single policy or from a merely dual organization.

1. Conceptual scope and recurrent structure

Across the cited literature, tri-system formulations share a structural motif: the system is partitioned into three qualitatively distinct functions, and overall competence depends on their interaction rather than on any one part in isolation. In the practopoietic formulation, the triad is explicitly adaptive and hierarchical: $T_{\mathrm{G}}$ stores genetic policy, $T_{\mathrm{A}}$ stores adaptation policy, and $T_{\mathrm{N}}$ is the current neural policy that acts on the Umwelt $U$ (Nikolić, 2015). In general systems theory, the triad is conceptual rather than mechanistic: behavior concerns what a system does, organization concerns how it is structured and socially arranged, and substance concerns what it is as a conceptual model with internal representation and identity (Florio, 2014). In DynamicMind, the triad is computational and control-oriented: Fast, Normal, and Slow modes are prompt-conditioned reasoning regimes for one frozen LLM (Li et al., 6 Jun 2025). In robotic manipulation, the triad is hierarchical and real-time: a VLM Brain performs semantic subtask planning, a VLA Cerebellum executes reactive control, and a visual Critic monitors progress and triggers replanning (Yi et al., 5 Mar 2026).

A plausible implication is that “tri-system” functions less as a single theory name than as a recurrent research pattern. The pattern typically introduces an intermediate or meta-level absent from binary decompositions: adaptation between genes and behavior, organization between behavior and substance, a native reasoning mode between fast and slow thinking, or a critic between planner and controller.

Domain	Triadic structure	Core role of the third component
Practopoietic mind	$T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$	Reconfigures the acting policy (Nikolić, 2015)
General systems theory	behavior, organization, substance	Adds conceptual identity beyond behavior and structure (Florio, 2014)
LLM control	Fast, Normal, Slow	Preserves native capability between extremes (Li et al., 6 Jun 2025)
Robotics	Brain, Cerebellum, Critic	Schedules replanning and anomaly handling (Yi et al., 5 Mar 2026)
Training physiology	$CP, W', P_{\max}$	Separates oxidative, glycolytic, and alactic adaptation (Kontro et al., 19 Mar 2025)
Conflict analysis	agreement, disagreement, neutral	Adds a boundary region for non-commitment (Lang, 2019)

2. Practopoietic Tri-System Theory of mind

The most explicit and architecturally developed use of Tri-System Theory is the practopoietic account of mind in which intelligence is organized through three adaptive traverses (Nikolić, 2015). Practopoiesis treats both action in the environment and self-modification as adaptive traverses in a hierarchy of monitor-and-act units. A traverse is the application of knowledge at one level to produce or update knowledge at the next higher level, or to act on the environment. A system with $n$ levels of policy is correspondingly a Tn-system.

Within this framework, a T1-system has a single policy acting on the environment, a T2-system has one policy that acts and a lower policy that learns or changes that policy, and a T3-system has a third and deeper policy that controls the learning mechanisms which in turn control behavior. The biological case is formalized as

$T_{\mathrm{G} \to T_{\mathrm{A} \to T_{\mathrm{N} \to U}}},$

where $T_{\mathrm{G}}$ is the genetic policy, $T_{\mathrm{G}}$ 0 is the adaptation policy, $T_{\mathrm{G}}$ 1 is the neural network policy, and $T_{\mathrm{G}}$ 2 is the Umwelt (Nikolić, 2015). At the species level, the hierarchy can be extended to

$T_{\mathrm{G}}$ 3

with $T_{\mathrm{G}}$ 4 denoting evolutionary policy; in that description, the species is T4 and the individual is T3.

The reinforcement-learning recasting is central. The paper defines a practopoietic hierarchy by the condition that for policy $T_{\mathrm{G}}$ 5, there is a policy $T_{\mathrm{G}}$ 6 whose actions change policy $T_{\mathrm{G}}$ 7, symbolically $T_{\mathrm{G}}$ 8. Traditional reinforcement learning, including TD-learning and Q-learning, is treated as a T2 structure because a learning rule updates the policy that acts on the environment. By contrast, a T3 agent includes a deeper policy shaping the learning rules themselves (Nikolić, 2015).

The paper’s principal argument for the necessity of T3 is a variety argument grounded in Ashby’s law of requisite variety. Treating the brain as if it were a T2 system with frozen learning, the paper estimates roughly $T_{\mathrm{G}}$ 9 synapses, an upper estimate of 4.6 bits per synapse, and a simplified ceiling of about $T_{\mathrm{A}}$ 0 different states (Nikolić, 2015). It then contrasts this with combinatorial demands from language and vision. With an educated native speaker vocabulary of 15,000 words, noun–verb–noun three-word sentences yield about $T_{\mathrm{A}}$ 1 combinations, but adjective–noun–verb–adjective–noun five-word sentences are simplified to about $T_{\mathrm{A}}$ 2 combinations. In vision, assuming 10,000 object categories and working memory capacity around 4 familiar objects gives $T_{\mathrm{A}}$ 3 semantic combinations, and perceptual variation drives the numbers far higher (Nikolić, 2015).

From these estimates, the paper concludes that a T2 agent of human-brain size cannot satisfy real-world requisite variety, whereas a T3 organization gains multiplicative capacity because one policy can rapidly reconfigure another. The illustrative example assigns $T_{\mathrm{A}}$ 4 distinct states to $T_{\mathrm{A}}$ 5 and $T_{\mathrm{A}}$ 6 rapidly changeable states to $T_{\mathrm{A}}$ 7, producing

$T_{\mathrm{A}}$ 8

which the paper describes as seven orders of magnitude beyond the T2 estimate (Nikolić, 2015). On that basis, it argues that only a T3 agent can reach human-level intelligence with feasible resources.

The same source also specifies limitations. The synaptic information estimate, synapse count, idealized memory utilization, and combinatorial approximations are acknowledged as strong assumptions; the mapping from genes to plasticity rules to neural networks is explicitly theoretical; and direct neural evidence for exactly these three levels or for anapoiesis as neural adaptation is not presented in that paper (Nikolić, 2015). Accordingly, Tri-System Theory in this sense is best understood as a theoretical architecture supported by a variety-based argument rather than as a fully established empirical decomposition.

3. Triadic general systems theory: behavior, organization, substance

A broader philosophical precursor appears in the discussion of the three gestalts of general systems theory: behavior, organization, and substance (Florio, 2014). This paper does not use the exact label “Tri-System Theory,” but it explicitly develops a triadic framework in which each gestalt serves as a “systemic touchstone,” that is, a privileged aspect used to classify systems.

The first gestalt, behavior, derives from Rosenblueth, Wiener, and Bigelow. It classifies systems by “the change produced in the surroundings by the object” and by “the examination of the output of the object and of the relations of this output to the input” (Florio, 2014). The behavioral taxonomy distinguishes passive from active behavior, then purposeful from non-purposeful, then teleological from non-teleological, and finally extrapolatory from non-extrapolatory. Extrapolatory teleological systems are linked to proactive behavior and to autonomic MAPE-K loop systems.

The second gestalt, organization, follows Boulding’s “arrangement of levels of theoretical discourse,” including Thermostat, Cell, Plant, Animal, Human Being, Social organization, and Transcendental systems (Florio, 2014). Here the relevant features are not only input–output regularities but also openness, awareness, self-awareness, and especially social composition. Boulding’s social organization is defined as “a set of roles tied together with channels of communication,” thereby making communication structure an explicit dimension of system description.

The third gestalt, substance, is derived from Leibniz. The paper interprets substances as entelechies, conceptual models, algorithms, modules, and “fully interconnected networks of all-open, all-aware active-behaviored ‘nodes’” governed by the Principle of Concomitance (Florio, 2014). Each substance is said to possess a representation-and-reflection mechanism, so that internal models of the world mediate indirect interaction. In computer-science terms, substances are compared to modules, scripts, processes, components, or services; the Characteristica Universalis is treated as a diagrammatic knowledge representation language, and the Calculus Ratiocinator as an evaluation engine. The paper even models Leibniz’s God as a scheduler operating through a procedure called Ultimate Sort, with IntrinsicQuality, ExtrinsicQuality, and Receptivity(W) as core notions (Florio, 2014).

In this triadic setting, behavior is the most external layer, organization introduces hierarchical and social structure, and substance adds conceptual identity, internal representation, compossibility, and selection for existence. The paper explicitly states that substance absorbs the previous two gestalts, since IntrinsicQuality(s) may include “the behavioral class of s” and “the Boulding level of s” (Florio, 2014). This suggests a tri-system conception in which later layers subsume earlier ones rather than merely coexisting with them.

A common misconception would be to treat this triad as a direct three-module engineering architecture. The paper is instead historical, conceptual, and interpretive. Its triadicity is methodological: it offers three lenses for classifying systems, not a single canonical three-block machine.

4. Tri-System architectures in AI and robotics

In recent AI systems, tri-system formulations are operationalized as concrete control architectures. DynamicMind proposes a tri-mode thinking system for zero-shot question answering in LLMs, extending the dual-process framing of fast and slow thinking by adding a Normal mode intended to preserve the model’s intrinsic capabilities (Li et al., 6 Jun 2025). The three modes are defined as Fast Mode $T_{\mathrm{A}}$ 9, Normal Mode $T_{\mathrm{N}}$ 0, and Slow Mode $T_{\mathrm{N}}$ 1, implemented entirely through system prompts $T_{\mathrm{N}}$ 2 on a frozen base model: $T_{\mathrm{N}}$ 3 Fast mode prohibits reasoning and is capped at 128 tokens; Normal mode is the standard instruction-following regime with a 2048-token budget; Slow mode requires decomposition, step verification, and long chain-of-thought with a 4096-token budget (Li et al., 6 Jun 2025).

DynamicMind’s control principle is Thinking Density,

$T_{\mathrm{N}}$ 4

with $T_{\mathrm{N}}$ 5 in TMC construction. A lightweight Mind Router implemented as DeBERTaV3-base predicts the best mode from question text alone, trained on the Thinking Mode Capacity dataset of about 39K questions (Li et al., 6 Jun 2025). On Llama-3.1-8B-Instruct, the average Thinking Density across math, commonsense, MMLU, and ScienceQA is reported as 0.27 for CoT, 0.29 for PBC, 0.25 for TALE-EP, and 1.33 for DynamicMind; on MMLU, DynamicMind obtains 52.48% at 34.17 tokens versus CoT at 52.02% and 290.51 tokens (Li et al., 6 Jun 2025). The paper also reports that on GSM8K around 80.52% of questions are better solved in the natural unaugmented mode than in explicitly enforced fast or slow modes, which is the empirical basis for treating Normal as a genuinely distinct regime rather than a midpoint (Li et al., 6 Jun 2025).

A second operationalization appears in long-horizon manipulation. Critic in the Loop introduces a bionic Tri-System VLA architecture comprising a VLM Brain for semantic subtask generation, a VLA Cerebellum for fast action chunks, and a visual Critic for progress estimation, anomaly detection, and control routing (Yi et al., 5 Mar 2026). The planner and controller are formalized as

$T_{\mathrm{N}}$ 6

while the critic outputs

$T_{\mathrm{N}}$ 7

with $T_{\mathrm{N}}$ 8 and $T_{\mathrm{N}}$ 9 either a progress token or the anomaly token $U$ 0 (Yi et al., 5 Mar 2026). Control authority is transferred by an event-driven policy: success is detected when $U$ 1, accident when $U$ 2, and stagnation when $U$ 3, with $U$ 4 (Yi et al., 5 Mar 2026). The critic is always active, but only the Brain replans on trigger events; otherwise the Cerebellum sustains closed-loop execution at about 20 Hz.

Empirically, the robotic tri-system is evaluated on Arrange the Tableware and Tidy up the Desk. On Arrange the Tableware, the tri-system reaches 10/10 in Ordered, 9/10 in Scattered, 7/10 in Left cup, and 7/10 in Fallen, compared with lower results for single-system and dual-system baselines (Yi et al., 5 Mar 2026). These results are presented as evidence that a third evaluative subsystem can improve robustness under OOD perturbations, long-horizon dependencies, and failure recovery.

Both systems articulate the same general point in different technical idioms. A dual architecture of planner and controller, or of fast and slow reasoning, is argued to be insufficient because it lacks either a native intermediate regime or an always-on evaluative scheduler. The triadic addition is not decorative; it changes the control topology.

5. Other domain-specific tri-system formulations

Outside cognition and embodied AI, tri-system structures appear in several technically unrelated fields.

In training physiology, the three-dimensional impulse-response model replaces a one-dimensional Banister-style load model with three parallel fitness–fatigue channels corresponding to the alactic, lactic, and aerobic energy systems (Kontro et al., 19 Mar 2025). Performance is represented through the three-parameter critical power model, whose parameters are $U$ 5, $U$ 6, and $U$ 7. The maximal power-duration relation is written

$U$ 8

and power above $U$ 9 is decomposed into oxidative, glycolytic, and alactic contributions via $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 0, $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 1, and $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 2 (Kontro et al., 19 Mar 2025). System-specific training loads $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 3, $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 4, and $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 5 then drive separate impulse-response chains for each energy system. The paper is explicit that this is a conceptual and methodological proposal rather than a full validation study.

In gravity, tri-gravity is a tri-system theory in the literal sense of three interacting spin-2 fields $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 6 analyzed in the ADM Hamiltonian formalism (Molaee et al., 2019). On a special physical branch of the constraint surface, the theory retains one common diffeomorphism invariance, gains enough second-class constraints to remove the Boulware–Deser ghost, and propagates exactly one massless plus two massive spin-2 fields. The paper’s DOF count yields 12 configuration-space degrees of freedom, matching $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 7 (Molaee et al., 2019). Here “tri-system” does not mean cognitive levels or modes; it means three coupled metric sectors.

In decision theory and conflict analysis, three-way decisions-based conflict analysis models partition universes into agreement, disagreement, and neutral regions (Lang, 2019). At the agent level, for a strategy $T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 8,

$T_{\mathrm{G}}, T_{\mathrm{A}}, T_{\mathrm{N}}$ 9

are defined from acceptance and rejection evaluations $CP, W', P_{\max}$ 0 and $CP, W', P_{\max}$ 1. At the issue level, analogous definitions use $CP, W', P_{\max}$ 2 and $CP, W', P_{\max}$ 3 for an agent group $CP, W', P_{\max}$ 4 (Lang, 2019). The third region is not an intermediate mechanism but a boundary or non-commitment region, and the tri-system character lies in three-way partitioning under uncertainty.

In quantum mechanics, tri-supersymmetry furnishes yet another meaning. Finite-gap, parity-even periodic Schrödinger operators admit a triplet of nontrivial commuting integrals $CP, W', P_{\max}$ 5, with three possible $CP, W', P_{\max}$ 6-gradings $CP, W', P_{\max}$ 7, $CP, W', P_{\max}$ 8, and $CP, W', P_{\max}$ 9 (Correa et al., 2008). For an $n$ 0-gap system, the paper derives $n$ 1 nontrivial isospectral super-extensions, and conjectures that finite-gap systems with antiperiodic singlets admit self-isospectral tri-supersymmetric extensions with partner potential shifted by a half-period (Correa et al., 2008). Although the context is unrelated to the cognitive or systems-theoretic literature, it exemplifies the broader mathematical use of triadic organization.

6. Interpretation, limitations, and recurrent debates

The central interpretive issue is whether Tri-System Theory should be regarded as a single transdisciplinary theory or as a recurring family of three-part formalisms. The literature supports the latter reading more directly than the former. The practopoietic T3 theory, the three gestalts of systems theory, the tri-mode LLM controller, and the robotic Brain–Cerebellum–Critic architecture all employ triadic decompositions, but their primitives, objectives, and evidential bases differ substantially [(Nikolić, 2015); (Florio, 2014); (Li et al., 6 Jun 2025); (Yi et al., 5 Mar 2026)].

A second recurrent debate concerns whether the third component is genuinely necessary or merely a refinement of a dual model. The practopoietic paper argues that T2 architectures are fundamentally limited by variety and that multiplicative variety requires a third adaptive level (Nikolić, 2015). DynamicMind argues that dual-process framing misses the LLM’s pretraining-shaped Normal mode and therefore overwrites a useful native regime (Li et al., 6 Jun 2025). Critic in the Loop argues that a planner–controller pair is insufficient because robust execution requires a continuously active evaluator that decides when planning should be re-engaged (Yi et al., 5 Mar 2026). These are domain-specific arguments, but they all contest the adequacy of a two-part decomposition.

The limitations are equally domain-specific. The practopoietic variety argument depends on strong upper-bound assumptions about synaptic information, synapse count, and combinatorial estimates (Nikolić, 2015). DynamicMind’s router is model-specific to some degree, its tri-mode discretization may be coarse, and TMC is built from one base LLM and limited domains (Li et al., 6 Jun 2025). Critic in the Loop still bottlenecks on VLM reasoning in some OOD arm-selection cases and requires demonstration data including failure-recovery cases (Yi et al., 5 Mar 2026). The three-dimensional training model explicitly lacks empirical estimation of system-specific $n$ 2 and $n$ 3 parameters and assumes independent adaptation channels (Kontro et al., 19 Mar 2025). In general systems theory, the triad of behavior, organization, and substance is conceptually rich but only partially operationalized (Florio, 2014).

A final misconception is to equate all tri-system theories with “System 1, System 2, System 3” psychology. Some works do borrow cognitive metaphors, but the actual triads vary: genes–adaptation–neural state, behavior–organization–substance, fast–normal–slow, brain–cerebellum–critic, $n$ 4– $n$ 5– $n$ 6, three metrics in tri-gravity, or agreement–disagreement–neutral. The shared principle is not a fixed triplet of named modules, but a recurrent thesis that three differentiated components can resolve limitations of one-level or two-level models.