Concept paper · cognition / intelligence / control
Recursive Token Control
A functional account of reflective cognition, cultural algorithms, and variance in human reasoning.
This note proposes a compact interpretation of reflective, “System 2” cognition inside Token Processing Intelligence Theory (TPIT) and Learned Value Transducer (LVT). The central claim is that reflective cognition need not be a separate kind of intelligence. It can be modeled as recursive control over learned token transformations: a system holds intermediate outputs active, feeds selected outputs back through learned transformations, suppresses premature termination, compares alternatives, and changes the control-value assigned to candidate states. Human variation then appears not as a single quantity, but as variance in the depth, stability, selection, correction, and culturally acquired procedures of these recursive loops. The proposal is functional rather than neuroanatomical and is intended to generate testable distinctions, not to rename established constructs.
1. The claim
A useful way to interpret reflective cognition is not “a second mind,” but a mode in which a token-processing system can operate repeatedly on its own intermediate outputs before action terminates the loop.
Dual-process research commonly distinguishes fast, autonomous processing from slower processing that supports hypothetical reasoning and places heavier demands on working memory. The familiar labels “System 1” and “System 2” are convenient shorthand, but they should not be mistaken for two literal anatomical systems. The proposal here treats that distinction as a difference in processing regime. Fast processing produces candidate outputs from learned structure; reflective processing keeps selected outputs available long enough to transform them again.
Type-2-like cognition is recursive token control: the controlled retention, reselection, retransformation, and revaluation of intermediate tokens before a behaviorally consequential output is released.
2. TPIT and LVT
TPIT defines intelligence functionally as token transformation through learned, evolved, or adaptive structure, with outputs altering future state, behavior, or coordination. Tokens need not possess intrinsic meaning; they need only create operational differences that propagate through the system. LVT is a special case in which measurement vectors are transformed through learned structure into control-value vectors that regulate real processes.
This separation is useful for cognition. A thought, number, word, memory, bodily signal, or percept can function as a token without being “true.” Its causal importance depends on what the system does with it. A socially learned narrative such as status, failure, danger, obligation, or identity can therefore be constructed while remaining operationally real: once learned, it can change attention, physiology, choice, and subsequent inputs.
3. Arithmetic as the minimal example
Arithmetic exposes the mechanism clearly. A familiar fact such as 2 + 2 → 4 can be retrieved almost automatically after learning. A larger calculation requires more than possession of elementary transformations. The solver must preserve intermediate tokens, select the next operation, inhibit irrelevant alternatives, detect errors, and feed partial results back into the process. The elementary operations may be highly automatized while their orchestration remains effortful.
On this account, reflective cognition is not defined by a special substance or a special symbol type. It is defined by the organization of feedback. The output of one learned transformation remains available as the input to another, under a control policy that can continue, stop, backtrack, branch, or verify.
4. Human variance is multidimensional
If reflective cognition is recursive control, then variation between people should not collapse into a single “System 2 ability.” Established work already separates components of executive function such as updating, inhibition, and shifting, and links working-memory capacity with executive attention and higher cognition. The present framework translates that heterogeneity into a common functional vocabulary.
Workspace stability is the ability to keep intermediate tokens available without loss or distraction. Recursive depth is how many consequential transformation steps can be sustained before the loop collapses or terminates. Branching width is the ability to maintain competing candidate states. Control policy determines when an intuitive output is accepted versus subjected to further processing. Error correction governs monitoring, backtracking, and revision. Transformation repertoire is the stock of learned operations the system can deploy. These dimensions can vary independently enough that two people with similar elementary knowledge can perform very differently on multistep reasoning.
This also separates capacity from deployment. Cognitive reflection tasks illustrate that an immediately compelling answer can be generated before a person elects to override it. A person may possess the capacity for extended reasoning but not recruit it in a given context; another may have lower raw capacity but a stronger policy of checking, verification, or deliberate delay.
5. Culture supplies executable procedures
Human reflective ability is amplified by cumulative culture. Long multiplication, algebra, formal logic, scientific controls, bookkeeping, programming, and statistical inference are not reconstructed independently by each learner. They are socially transmitted procedures for sequencing transformations. Experimental work on cultural learning shows that populations can preserve and accumulate useful cognitive algorithms that are difficult for isolated individuals to rediscover.
In TPIT terms, culture changes the transformation repertoire. In LVT terms, it can also change the control mapping: what is noticed, what counts as an error, which measurements deserve weight, and which output should govern action. Language and institutions therefore do more than store propositions. They can install procedures for operating on representations.
6. Humans and other animals
The framework does not require an absolute human/animal boundary. Nonhuman animals show planning, inhibition, memory, tool use, and other capacities that overlap with components often associated with reflective cognition. The disputed question is the degree and organization of higher-order relational and symbolic processing. A cautious hypothesis is that the human discontinuity, where it exists, lies less in possession of a wholly different cognitive mechanism than in the scale, recursion, symbolic flexibility, and cultural extension of feedback-controlled transformations.
7. From narrative self to control structure
The same account applies to self-narrative. Thoughts are produced from prior tokens: language, memory, social roles, bodily measurements, predictions, and learned valuations. A narrative can then become recursively self-reinforcing when its outputs alter attention and behavior in ways that generate confirming inputs. The claim that a narrative is socially constructed therefore does not imply that it is causally inert. A constructed token pattern can become a stable control structure.
This reframes metacognitive “detachment” without requiring a metaphysical observer. A system can represent one of its own outputs as an object of further processing, reduce its control value, compare it with alternatives, or decline to act on it. The thought remains present; what changes is its position in the control architecture.
8. Testable consequences
The account becomes useful only if it sharpens measurement. It predicts that apparently unitary reasoning performance should decompose into separable limits on retention, recursion, branching, arbitration, and error correction; that training can improve performance by installing new transformation sequences even when elementary operations are unchanged; that expertise should shift formerly recursive procedures toward automatic execution while freeing control capacity for higher-level recursion; and that failures conventionally described as “bias” can arise either from an unsuitable automatic output or from a control policy that terminates processing too early. These are empirical distinctions and should be evaluated against existing cognitive models rather than assumed from the terminology.
9. Scope
This is a functional synthesis, not evidence that TPIT/LVT replace dual-process theory, working-memory models, executive-function theory, or comparative cognition. “Token” is deliberately substrate-neutral, which makes the framework broad but creates a burden: useful applications must specify the tokens, transformations, feedback path, measurement variables, and control effects precisely enough to be falsifiable. The value of the proposal is therefore compression. It offers one language for automatic processing, reflection, cultural learning, self-narrative, and individual variance while preserving the distinctions already measured by cognitive science.
References
- Evans, J. St. B. T., & Stanovich, K. E. (2013). Dual-process theories of higher cognition: Advancing the debate. Perspectives on Psychological Science, 8(3), 223–241. doi:10.1177/1745691612460685
- Baddeley, A. (1992). Working memory. Science, 255(5044), 556–559. doi:10.1126/science.1736359
- Miyake, A., Friedman, N. P., Emerson, M. J., Witzki, A. H., Howerter, A., & Wager, T. D. (2000). The unity and diversity of executive functions and their contributions to complex “frontal lobe” tasks: A latent variable analysis. Cognitive Psychology, 41(1), 49–100. doi:10.1006/cogp.1999.0734
- McCabe, D. P., Roediger, H. L. III, McDaniel, M. A., Balota, D. A., & Hambrick, D. Z. (2010). The relationship between working memory capacity and executive functioning: Evidence for a common executive attention construct. Neuropsychology, 24(2), 222–243. doi:10.1037/a0017619
- Frederick, S. (2005). Cognitive reflection and decision making. Journal of Economic Perspectives, 19(4), 25–42. doi:10.1257/089533005775196732
- Thompson, B., van Opheusden, B., Sumers, T., & Griffiths, T. L. (2022). Complex cognitive algorithms preserved by selective social learning in experimental populations. Science, 376(6588), 95–98. doi:10.1126/science.abn0915
- Boyd, R., Richerson, P. J., & Henrich, J. (2011). The cultural niche: Why social learning is essential for human adaptation. Proceedings of the National Academy of Sciences, 108(Suppl. 2), 10918–10925. doi:10.1073/pnas.1100290108
- Penn, D. C., Holyoak, K. J., & Povinelli, D. J. (2008). Darwin’s mistake: Explaining the discontinuity between human and nonhuman minds. Behavioral and Brain Sciences, 31(2), 109–130. doi:10.1017/S0140525X08003543
Status: conceptual working paper. TPIT and LVT are treated here as proposed functional frameworks. The paper makes no claim that “System 1” and “System 2” are literal neural modules, and no claim that a single mechanism has settled the comparative cognition debate.