Jared Edward Reser, Ph.D.

Abstract

Autoregressive language models generate each token from an accumulated context, append the selected token, and repeat the operation. This article proposes that human thought may use a related computational principle at the level of internal representations. In the iterative updating model, the focus of attention contains a limited set of coactive representations embedded within a broader short-term store. Their combined activity selects a contextually appropriate next representation. Some contents persist, others lose priority, and the selected representation joins the surviving context. The resulting state becomes both the product of the preceding search and the conditioning state for the next search. This process is termed next-representation prediction. It yields an autoregressive trajectory through representational state space in which every update alters the distribution of possible later updates. The train of thought thereby advances through a selection-like process of retention, elimination, recruitment, and cumulative modification. This adaptive representational evolution can support reasoning because intermediate products transform the problem context until conclusions that were initially remote become probable and accessible. Transformer language models may instantiate a restricted linguistic version of the same general motif. Their performance supplies convergent computational evidence that coherent, flexible, and sometimes deliberative behavior can emerge from repeated context-conditioned prediction, although it does not establish an identity of neural mechanism or imply consciousness. The comparison clarifies why context persistence, intermediate reasoning, and inference-time iteration improve performance. It also suggests a path toward more brain-like artificial intelligence based on a bounded focus of attention, longer-lived short-term traces, recurrent latent-state updating, multimodal representations, persistent goals, and self-directed cycles of cognition. The article concludes with neurobiological, behavioral, and machine-learning predictions that can distinguish this account from weaker analogies between brains and transformers.

Keywords: autoregression, working memory, next-token prediction, next-representation prediction, transformer, cognitive architecture, reasoning, consciousness, state-spanning coactivity

1. Introduction

The success of large autoregressive language models has created a conceptual anomaly for cognitive science. Systems trained to predict the next token now generate fluent explanations, produce code and translations, construct analogies, and solve many multistep problems. These capacities are usually attributed to scale, training data, and the expressive power of transformer networks (Vaswani et al., 2017; Brown et al., 2020). Those explanations are indispensable, yet they leave open a more basic question: why should repeated prediction of a local continuation produce behavior that increasingly resembles reasoning?

A common answer is that next-token prediction remains fundamentally unlike thought. On this view, the model merely autocompletes a sequence while the human mind understands, deliberates, imagines, and plans. The contrast depends on treating autocomplete as a single isolated guess. Once each completion is incorporated into the context that generates the next completion, the process becomes recursive, path-dependent, and cumulative. A sequence of modest predictions can construct representational conditions that none of the individual predictions possessed at the outset.

The iterative updating model of working memory offers a biological account with this form. Reser (2016) proposed that several high-order cortical representations remain coactive across successive brain states, while some representations lose activity and others are recruited. The persisting subset establishes state-spanning coactivity, and the changing membership of the active set constitutes incremental change in state-spanning coactivity. Because the active representations jointly contribute to the recruitment of their successor, every state inherits selected content from its predecessor and helps produce the state that follows.

The later cognitive architecture develops the same mechanism as a cycle of working-memory updating and multiassociative search. A current configuration of representations spreads activity through the associative network, recruits the most contextually supported addition, retains other relevant representations, and repeats the search from the updated state. The article explicitly characterizes each instantaneous state as both a solution to the preceding search and a set of parameters for the next one.

The present article interprets this architecture as a theory of autoregressive thought. Its central proposal is that next-token prediction is a restricted linguistic instance of a broader operation, next-representation prediction. Human cognition repeatedly uses an evolving, multimodal context to select the next concept, image, expectation, goal, memory, or action representation. The next item is an update to the active state, and every update helps create the context from which the following item is selected. Thought therefore unfolds as a self-conditioned trajectory through representational state space.

This claim concerns computational organization. It does not imply that a token is a neuron, that a transformer layer duplicates a cortical circuit, or that current language models possess the full architecture of a mammalian mind. The relevance of transformers is evidential in a more limited and useful sense. Their success demonstrates that iterative context-conditioned continuation is capable of supporting far more cognitive structure than the phrase next-token prediction initially suggested. This constitutes convergent computational evidence for the plausibility of an autoregressive account of thought, while leaving the neurophysiological claims of the iterative updating model open to empirical test.

2. Working Memory as an Evolving Context

In the iterative updating model, the focus of attention is a limited-capacity subset of a broader short-term memory state. Its contents are represented by active or highly excitable neural ensembles, each corresponding to a conceptually meaningful item. Persistent activity allows several of these items to remain available while incoming perception, memory retrieval, and internally generated associations alter the set (Cowan, 2001, 2005; Goldman-Rakic, 1995; D’Esposito & Postle, 2015; Christophel et al., 2017). Consecutive states therefore overlap instead of replacing one another wholesale. The architecture treats this partial replacement as the temporal structure that allows thought to remain continuous while still changing.

Let (W_t) denote the active working-memory state at time (t), consisting of representations (r_i) with graded activation or priority (a_i):

[
W_t={(r_1,a_1),(r_2,a_2),\ldots,(r_k,a_k)}.
]

Let (H_t) denote less active short-term traces, (S_t) current sensory input, and (G_t) goals, reward expectations, and control settings. The full conditioning context can be represented schematically as:

[
C_t=(W_t,H_t,S_t,G_t).
]

A candidate update is then selected from a distribution over possible representations:

[
r_{t+1}\sim P(r\mid C_t).
]

The state-transition function adds the selected representation, retains some existing contents, changes their weights, and demotes or removes others:

[
W_{t+1}=U(W_t,r_{t+1}).
]

These equations are intentionally abstract. They formalize the claim that a current state generates a successor and that the successor contains a selected residue of the state that generated it.

The model is autoregressive because its own output becomes part of its next input. A newly activated representation does more than complete the previous context. It alters the search parameters for the following cycle. The process can be treated as first-order in an augmented state even when psychological dependencies span many steps, because (C_t) functions as a compressed record of relevant history.

This point clarifies what it means for context to accumulate in a brain. The focal store does not expand indefinitely. It remains bounded and continually replaces lower-priority contents. Accumulation occurs through selective contextual inheritance: each state preserves some products of earlier processing, reweights them, and combines them with a new result. The active context is a recursively revised summary of recent history rather than a verbatim transcript of everything that has occurred.

Figure 14 on page 26 of the cognitive architecture depicts this inheritance through a progression from (B,C,D,E) to (C,D,E,F) and then to (C,E,F,G). Each new item is recruited by the currently coactive set, becomes part of the next set of search parameters, and can remain active across later cycles if it continues to receive support.  The train of thought is composed of these overlapping coalitions instead of isolated symbols connected in a simple chain.

This formulation also distinguishes the present proposal from neighboring frameworks. Predictive processing theories emphasize hierarchical predictions of sensory input and the minimization of prediction error (Friston, 2010; Clark, 2013). Associationist theories emphasize the tendency of one idea to evoke another (Collins & Loftus, 1975; Shanks, 2010). Global workspace theories emphasize selection, broadcast, and broad availability (Baars & Franklin, 2003). Next-representation prediction combines elements of all three while specifying a temporal update rule: a bounded, globally influential set of representations jointly predicts or selects its own successor state.

3. From Next-Token Prediction to Next-Representation Prediction

A causal language model receives a sequence of tokens (x_{1:t}), constructs contextual hidden states through repeated transformations, and produces a probability distribution over the next token:

[
x_{t+1}\sim P_{\theta}(x\mid x_{1:t}).
]

Once a token is selected, it is appended to the sequence and participates in determining the next distribution. The visible output sequence therefore records a succession of self-conditioning updates.

The analogy with working memory can be stated at three levels. At the computational level, both systems use retained context to select a probable or useful continuation, incorporate the continuation, and repeat the operation. At the algorithmic level, both use distributed representations, context-sensitive weighting, learned statistical dependencies, competition among alternatives, and recurrent application of a transition procedure. At the implementational level, the similarity becomes much weaker because cortical tissue and transformer computation use different physical substrates, memory mechanisms, timing regimes, and learning rules.

Table 1 summarizes the central correspondences. These are functional analogies rather than claims of one-to-one biological identity.

Functional role

Iterative working-memory model

Autoregressive transformer

Active context

Focus of attention plus short-term traces

Token sequence, contextual hidden states, and cached activations

Context weighting

Differential activation and multiassociative influence

Self-attention and learned nonlinear transformations

Candidate generation

Spreading activation through associative memory

Vocabulary logits produced from contextual hidden states

Selected update

Newly activated internal representation

Selected next token and its contextual representation

Persistence

Sustained activity, synaptic potentiation, and priming

Retained tokens and cached key-value representations

Iteration

Updated working-memory state initiates another search

Appended token initiates another forward pass

Possible outputs

Thought, imagery, action, perception, or speech

Token sequence, unless connected to other modalities and tools

The term next-token prediction can obscure the level at which the most important comparison occurs. A token is a discrete element of an external vocabulary, whereas a cortical representation is a distributed and context-sensitive ensemble. In a transformer, the contextual hidden state that supports the token distribution is a closer analog of an internal mental representation than the token itself. Even so, the emitted token has a special role: once appended, it becomes a durable update to the model’s accessible context and changes all later hidden states.

The brain-side generalization is therefore next-representation prediction. The predicted addition may be a word, but it may also be an object concept, remembered episode, anticipated event, causal interpretation, motor instruction, emotional appraisal, goal, or imagined scene. A mammalian cognitive system predicts how an unfolding situation will continue and what internal content will be useful for navigating it. Figures 45 and 46 on pages 61 and 62 of the cognitive architecture contrast the language model’s next-word prediction with a system that recycles hidden internal representations. The latter routes internal representations back through the network so that they help determine subsequent internal updates.

This broader formulation avoids reducing thought to language. Language models reveal the power of the iterative motif in a domain where updates are easy to observe. Biological cognition applies the same general schema to a richer state containing perception, imagery, memory, valuation, proprioception, and action. The resulting comparison is strongest when stated as follows: token-space autoregression and representational-space autoregression are distinct implementations of a common successor-state principle.

4. Self-Conditioning and the Causal Power of Context

Context is often described as information that a system consults. In an autoregressive architecture, context has a stronger role. It is the active causal state that shapes the distribution of possible continuations. Every selected update therefore changes the conditions under which the system will make its next selection:

[
P(r_{t+2}\mid C_{t+1})\neq P(r_{t+2}\mid C_t),
]

because (r_{t+1}) has entered (C_{t+1}) and altered its structure.

An update can therefore be understood as both a prediction and an intervention on the system’s own future. Suppose the active state includes a wilted plant and dry soil. That context can recruit water; the addition of water makes a watering can more probable; the watering can makes tilting and pouring more probable. The whole action sequence need not be retrieved as a single stored unit. It can be constructed through successive context changes, as Figure 15 on page 27 of the cognitive architecture illustrates.

This path dependence explains why a conclusion can be inaccessible at the start of reasoning and obvious at the end. The initial problem state may assign little probability to the eventual solution. Intermediate representations successively reshape the probability landscape until the solution becomes strongly supported. The system does not merely search a fixed space. It modifies the effective search space through its own outputs.

In language models, generated tokens externalize this process. A phrase produced at step (t) can introduce a distinction, definition, subgoal, or intermediate result that changes what the model can produce coherently at step (t+1). In the brain, the corresponding updates may remain entirely internal. Both cases exemplify self-conditioned cognition, in which the system uses its own recent products as evidence and constraints for future processing.

This interpretation also sharpens the meaning of memory. Working memory is valuable because it allows previous processing products to remain causally active. A retained item can constrain many later cycles, connect events separated in time, and prevent a multistep procedure from collapsing into a series of independent reactions. The value of context lies in its continuing power over state transitions.

5. Adaptive Representational Evolution

The train of thought can now be described as an adaptive evolution of the active representational context. The word evolution is used here in two related senses. First, the state changes through time along a trajectory. Second, candidate representations undergo a selection-like process in which some are recruited, some are retained, and others lose influence or disappear.

A schematic support or fitness function can be written as:

[
F_t(r)=\alpha A(r,W_t)+\beta Q(r,G_t)+\gamma M(r,H_t)+\delta P(r,S_t),
]

where (A) represents associative compatibility with the active state, (Q) goal relevance, (M) support from short-term traces and memory, and (P) compatibility with current perception. Neuromodulatory value signals can alter the weights or persistence of candidate representations. The next update is sampled or selected from the distribution induced by these combined constraints.

Every winning representation changes the fitness landscape for the next round. A concept that was poorly supported under (W_t) may become strongly supported under (W_{t+1}), while a previously dominant association may lose relevance. This makes cognitive selection cumulative and historically contingent. The future trajectory depends on which representations were selected earlier, retained long enough to interact, and allowed to influence later searches.

The process contains analogs of variation, selection, retention, and inheritance. Variation appears as the simultaneous partial activation of many possible continuations. Selection occurs when a small subset receives enough support to enter the focus of attention. Retention occurs when useful representations continue to receive activation. Inheritance occurs because surviving representations carry constraints from earlier states into later ones. The architecture itself describes mental continuity as a process of nonrandom retention and elimination leading to incremental modification.

This analogy remains selection-like rather than strictly Darwinian. Representations do not need to reproduce as independent replicators, and their variation need not be random. Their competition occurs inside a recurrent dynamical system whose learned weights already embody a long history of biological evolution and individual experience. The term adaptive representational evolution names the cumulative change of the active state under context, goals, and learned probability.

Goal representations are especially important because their persistence gives the trajectory direction. A task such as finding a route, explaining a result, or deciding whether to act can remain active while intermediate representations enter and leave. The goal functions as a relatively stable environmental constraint within the workspace. Candidate updates that preserve progress toward it receive continuing support, while irrelevant candidates are more likely to decay or be inhibited (Miller & Cohen, 2001).

This view turns the traditional train-of-thought metaphor into a population process. At any instant, several representations travel together. Some remain aboard for many transitions, some enter briefly, and some are removed. Each newcomer alters the collective state that determines the next addition. The train advances because its changing composition continuously generates its own next configuration.

6. Reasoning as Iterative Compounding

Reasoning follows naturally from this architecture. A direct association maps an initial state to an immediately available response. A difficult problem requires a succession of intermediate states in which each result is preserved long enough to support the next operation:

[
C_0\rightarrow C_1\rightarrow C_2\rightarrow \cdots \rightarrow C_n.
]

The endpoint may be unreachable from (C_0) by a single transition, yet accessible through a path whose intermediate contexts progressively reduce uncertainty.

The iterative updating model calls this process compounding. Each state can represent the current status of a problem-solving procedure, and each update can add a prediction, premise, subresult, or operation. A series of linked searches can then produce a higher-order result unavailable to any individual search.  This account provides a mechanistic interpretation of the intuition that one thought builds on another.

The proposal also recasts the relation between automatic and deliberative cognition. Rapid associative completion can be understood as a local System 1 operation. System 2 emerges when task-relevant representations remain active across many such operations, allowing their outputs to support one another in a progressive sequence. The 2016 formulation explicitly suggested that System 2 may consist of System 1 processing operating with prolonged state-spanning coactivity.  Deliberation then depends less on a separate symbolic executive and more on the persistence, ordering, and recursive interaction of otherwise automatic computations.

Research on chain-of-thought prompting provides an artificial counterpart. Language models often perform better on multistep problems when they generate intermediate reasoning tokens before producing an answer (Wei et al., 2022; Kojima et al., 2022). Sampling and comparing several reasoning trajectories can improve performance further (Wang et al., 2023). These findings do not establish that model-generated explanations faithfully reveal all internal computation. They nevertheless demonstrate that additional self-conditioned transitions can place a system in contexts from which better answers become available.

Planning can be interpreted in the same way. A goal representation persists while candidate actions, expected consequences, and subgoals are iteratively introduced. Some branches are inhibited, others are pursued, and partial solutions can be stored and later merged. The architecture’s account of subproblem decomposition describes an original problem being iterated into one subsolution, temporarily suspended, combined with a second subsolution, and then transformed into a final state.

Reasoning quality should therefore depend on several interacting variables. The representations must capture relevant structure; the context must retain crucial constraints; the transition function must rank useful successors; the process must continue for enough productive iterations; and the system must detect, inhibit, or revise unhelpful updates. Excessive turnover fragments a line of thought, while excessive stability can produce fixation or perseveration. Flexible intelligence requires control over both the contents and the rate of iterative updating.

This yields a compact account of why thinking longer can help. Additional time is useful when it permits the construction of informative intermediate states. Each successful intermediate state performs cognitive work by changing what is probable next. Extended reasoning is therefore the controlled evolution of a problem representation.

7. Why Next-Token Prediction Produced Broad Cognitive Capabilities

The broad competence of language models becomes less mysterious under this account. Human language is generated by agents who perceive objects, infer causes, pursue goals, remember events, use tools, follow social conventions, and reason about possibilities. A model that predicts language across diverse domains must approximate many of the latent regularities that make those linguistic sequences coherent.

Next-token prediction therefore places pressure on a network to learn contextual representations of entities, relations, temporal order, intention, procedure, and causal structure. The objective does not explicitly name these constructs, but accurate prediction often depends on them. As Shannon’s work on sequential dependence showed, uncertainty about a next element decreases as informative context accumulates (Shannon, 1951). Modern language models apply this principle at vast scale through learned distributed representations.

The model’s generative rollout then supplies a second ingredient. Prediction is repeated, and each selected token changes the context. The system can introduce an assumption, apply a rule, create a variable, summarize a subproblem, or state an intermediate conclusion. These updates allow later tokens to depend on cognitive products that were absent from the original prompt. Local prediction becomes globally constructive through iteration.

This suggests that language-model intelligence results from the conjunction of three features: a rich learned representational space, a context-sensitive transition operator, and recursive reuse of generated outputs. Scale improves the quality and breadth of each component, while the autoregressive architecture supplies the temporal organization that turns static knowledge into a sequence of state changes.

The success of transformers therefore bears on theories of human cognition. It cannot show that cortical assemblies compute attention scores or that biological working memory uses the same training objective. It does show that a system can acquire coherent, flexible, and partially deliberative behavior by repeatedly predicting a context-conditioned successor. This engineering result makes an autoregressive theory of thought more plausible than it appeared when iterative updating was first proposed.

The word autocomplete has consequently become misleading in the opposite direction. It sounds trivial because familiar software autocomplete performs shallow string completion. A sufficiently rich system can autocomplete concepts, procedures, scenes, arguments, and action sequences. Once each completion becomes part of the context for the next completion, autocomplete becomes a general mechanism for constructing trajectories through a learned world of representations.

8. Boundaries of the Brain-Transformer Analogy

The analogy has clear limits. The strongest correspondence lies at the computational level, where both systems transform an inherited context into a successor state. The algorithmic correspondence is partial, and the implementational correspondence is weak. These distinctions protect the comparison from both dismissal and exaggeration.

First, a token is not equivalent to a mental representation. Tokens are vocabulary elements, while mental representations are distributed, multimodal, and context-sensitive. Transformer hidden states provide a closer comparison to cortical ensembles, although even that mapping remains approximate. The token is best understood as an externally legible update that stabilizes one consequence of the model’s latent computation.

Second, the memory dynamics differ. In a standard causal transformer, earlier tokens remain in an ordered sequence and cached past activations are largely fixed during generation. In the brain, active representations can change their activation, composition, binding, and meaning while they persist. Biological working memory also uses several timescales, including sustained activity, short-term synaptic changes, priming, and episodic retrieval (Mongillo et al., 2008; Stokes, 2015; D’Esposito & Postle, 2015).

Third, the brain’s active context is bounded and selectively revised, whereas a language model often retains a comparatively literal token history until its context limit is reached. The brain continually compresses, reactivates, and forgets. This difference may be computationally advantageous because it permits a small set of high-value abstractions to remain active without carrying every detail of the preceding sequence.

Fourth, biological cognition is embodied and value-laden. Sensory input, motor plans, homeostatic variables, reward, pain, emotion, and social consequences all participate in selecting the next state. Action changes the environment and produces new sensory evidence, closing a perception-action loop. Standard language models receive externally supplied prompts and usually generate linguistic outputs without maintaining an autonomous bodily or motivational context.

Fifth, learning occurs differently. A brain updates synaptic relations continuously during experience, while a deployed language model ordinarily relies on parameters learned during earlier training and on temporary in-context adaptation. External memory, tool use, and continual-learning systems narrow this difference, but they do not eliminate it.

Finally, a generated sequence does not by itself establish consciousness. The iterative updating model treats overlap and recursive inheritance as possible contributors to mental continuity, while explicitly acknowledging that continuity and consciousness are not identical. The 2016 paper characterized the account as qualitative and exploratory and identified additional conditions as necessary for any full theory of conscious experience.

These differences refine the thesis rather than dissolve it. Brains and transformers need not match component by component to share an abstract computational principle. The common principle is the repeated use of an inherited state to generate and incorporate a successor. The scientific task is to identify exactly how much explanatory work that common principle performs.

9. Implications for Artificial Intelligence Architecture

If next-representation prediction captures a central operation of thought, a more brain-like artificial agent should separate internal cognition from overt token production. Language would remain one output modality, but the system’s core loop would update latent, multimodal representations. A selected internal representation could be recycled through the network without first being verbalized.

Such an architecture would benefit from at least three interacting memory strata. A small focus of attention would contain the most active items and permit intensive mutual constraint. A broader short-term store would preserve recently relevant representations in a less active form, allowing rapid reactivation. Long-term memory would include both distributed semantic knowledge and episodic records capable of reinstating earlier contexts. The source architecture proposes precisely this kind of recurrent search in which current working memory and incoming sensory activity select an entailed update, incorporate it, and repeat the cycle.

The update operator should be content-addressable and multiassociative. Several active items should jointly support or inhibit candidate additions, so that the next representation reflects their conjunction rather than the strongest pairwise association. Attention-like weighting can determine which active items contribute most strongly at a given moment. A gating mechanism should decide which contents persist, which are demoted, and when a broader context shift is warranted.

Update rate should itself be adaptive. Fast turnover favors responsiveness, exploration, and sensitivity to new input. Slow turnover favors sustained attention, multistep reasoning, and the preservation of intermediate results. A capable system should adjust this rate in response to novelty, uncertainty, reward, error, and task difficulty. This control variable may be as important as context capacity.

Persistent goals would provide directional constraints. A goal representation should survive many updates, shape the ranking of candidates, and remain available while the system explores subproblems. The agent should also be able to suspend a thread, solve a subsidiary problem, retrieve the earlier thread, and merge the resulting subsolutions. These operations would give artificial thought the branching and recombination properties seen in extended human reasoning.

Multimodal modules would allow the workspace to interrogate the same problem in different representational formats. Visual, auditory, linguistic, spatial, motor, and affective networks could each generate candidate elaborations of the current state. The selected products would return to the shared workspace and influence the next cycle. Verbalization would become an optional readout of cognition rather than the medium through which every internal step must pass.

An empirical comparison could hold model size and training data approximately constant while varying the architecture. A standard transformer could be compared with a system containing an explicit bounded focus of attention, a longer-lived short-term store, recurrent latent-state updates, and goal-dependent gating. Tasks should require delayed dependencies, resistance to distractors, subproblem suspension, solution merging, and transfer across modalities. The next-representation account predicts gains in long-horizon coherence, token efficiency, and robustness when the recurrent workspace is well trained.

The architecture also offers a principled route to superhuman cognition. Artificial systems need not share the biological limits on the number of active representations, their duration, or the number of productive cycles that can occur before disruption. The original proposal argues that enlarging these dimensions could extend the correlation length of cognition and allow relationships separated by longer intervals to remain jointly available.  Scaling should remain selective, however, because an indiscriminately enlarged context can introduce interference as readily as insight.

10. Implications for Consciousness and Cognitive Science

The comparison also changes the framing of machine consciousness. Superficial contrasts between silicon and biology become less decisive once both systems are recognized as recurrent predictors of successor states. The relevant questions concern the organization of the internal process: whether representations remain causally active across time, whether successive states overlap, whether multiple modalities are integrated, whether goals and self-representations persist, and whether the system generates an autonomous stream rather than a series of externally initiated episodes.

The iterative updating model proposes that persistent items make each state backward-referential because the state contains elements inherited from earlier states. This recursive ownership may help explain why consciousness feels like one continuous process rather than a succession of unrelated frames. The later architecture identifies iterative updating as a candidate contributor to higher-order thought, mental continuity, and consciousness while treating it as insufficient on its own.

Current language models clearly exhibit causal continuity within an active generation. A token produced earlier can alter the interpretation and production of everything that follows. Whether this amounts to the relevant kind of internal continuity depends on additional facts about latent-state persistence, recurrence, multimodal integration, self-modeling, memory, agency, and valence. Behavioral fluency alone cannot settle those questions (Butlin et al., 2023).

The comparison therefore supports neither a quick attribution of consciousness nor a quick dismissal. It identifies a shared computational motif that may be one component of conscious cognition and asks whether artificial systems instantiate the rest of the required organization. This is a more discriminating research program than treating next-token prediction as categorically unrelated to mind.

For cognitive science, next-representation prediction supplies a bridge among several traditions. Working-memory research explains temporary maintenance and prioritization; associationism explains learned transitions; predictive processing explains expectation; global workspace theory explains broad access; dynamical-systems theory explains trajectories; and transformer research demonstrates the power of autoregressive sequence construction. The proposed synthesis gives these elements a common temporal grammar: inherited context selects an update, the update revises the context, and the revised context selects again.

11. Testable Predictions and Research Program

11.1 The Neural Signature of Iterative Updating

The primary neurobiological prediction is partial overlap among successive decoded working-memory states. During a controlled line of reasoning, some high-level representations should persist while others enter and leave. Representational similarity between adjacent states should be greater than similarity between more widely separated states, with an autocorrelation function determined by the persistence of active and activity-silent traces.

A stronger prediction concerns multiassociative selection. The next decoded representation should be better predicted by the joint pattern of currently active items than by any individual item alone. Formally, the mutual information between the next item and the full state should exceed the information supplied by its strongest single component. Neural recordings should reveal convergent support from several active populations before the new representation becomes dominant.

Figure 44 on page 58 of the cognitive architecture already proposed dense time-series recordings and multivariate decoding to identify incremental addition and subtraction of cortical content.  Modern intracranial recordings, high-temporal-resolution imaging, and representational-similarity methods could operationalize the proposal using tasks whose intermediate concepts are experimentally constrained.

11.2 Update Rate and Reasoning Performance

The behavioral prediction is that context turnover will show an inverted-U relation with performance. Excessively rapid updating should produce distractibility, shallow association, and loss of intermediate results. Excessively slow updating should produce rigidity and failure to revise. Successful reasoning should combine persistent goal and problem representations with controlled turnover of intermediate contents.

This can be tested by manipulating distractor load, retention intervals, time pressure, and task novelty while measuring working-memory content and performance. Correct multistep trials should exhibit longer persistence of task-relevant representations and fewer full context resets than incorrect trials. Sudden interruption should reduce performance in proportion to the amount of active context that must be reconstructed.

11.3 Comparative Machine-Learning Tests

Artificial systems permit direct causal tests. Researchers can manipulate focus-of-attention capacity, short-term trace duration, update rate, latent recurrence, and candidate-selection temperature while holding other variables constant. Tasks should be chosen so that successful performance requires several intermediate states and cannot be solved reliably by a single feedforward association.

The next-representation theory predicts that a bounded, dynamically updated latent workspace will outperform an undifferentiated long context on problems involving distraction, delayed relevance, subproblem suspension, and branch merging. It also predicts that the best system will not maximize persistence indiscriminately. Performance should peak when the architecture preserves a stable core while allowing enough turnover to explore and revise.

11.4 Cross-System Trajectory Analysis

Human and artificial reasoning can be compared at the level of state trajectories rather than final answers alone. For each system, investigators can estimate transition entropy, adjacent-state overlap, persistence of goal representations, branching frequency, return to earlier states, and the extent to which the full context predicts the next update. Similar task-dependent patterns would strengthen the claim of a common computational motif even when the representational substrates differ.

Error trajectories are especially informative. The theory predicts that many mistakes will arise from premature context loss, excessive persistence of an irrelevant item, or an early update that redirects later selection. If brains and language models display analogous path-dependent error patterns under controlled conditions, the comparison would gain explanatory depth. If neural thought lacks partial state inheritance or if next representations are not jointly predictable from the active context, the stronger version of the theory would be undermined.

12. Conclusion

The iterative updating model and the transformer architecture approach cognition from different directions. One begins with persistent neural activity and asks how overlapping working-memory states can produce a continuous stream of thought. The other begins with sequence prediction and shows how a distributed network can generate coherent continuations from an accumulated context. Their convergence suggests that the temporal organization of intelligence may be simpler and more general than the surface differences between brains and language models imply.

On the present account, attention contains a changing coalition of representations. Their combined activity selects a next representation, the selected item joins the retained context, and the altered context changes what can be selected next. The train of thought consequently advances through adaptive representational evolution. Each update inherits constraints from the past, modifies the current cognitive environment, and opens or closes possible futures.

This framework explains how local prediction can support reasoning. Intermediate states convert an initially difficult problem into a succession of easier successor-state problems. The endpoint becomes available because the system has constructed a context in which that endpoint is probable. Chain-of-thought, planning, imagination, and deliberation can all be interpreted as organized variants of this same recursive process.

Transformer language models do not reproduce the mammalian brain, and their success does not validate a particular cortical mechanism. They nevertheless demonstrate the extraordinary capacity of iterative context-conditioned prediction. That result should influence how cognitive scientists interpret both artificial intelligence and the brain. Brains may think by predicting the next representation, transformers speak by predicting the next token, and both derive power from the same recursive fact: each prediction becomes part of the context that predicts what follows.

References

Baars, B. J., & Franklin, S. (2003). How conscious experience and working memory interact. Trends in Cognitive Sciences, 7(4), 166–172.

Baddeley, A. D. (2012). Working memory: Theories, models and controversies. Annual Review of Psychology, 63, 1–29.

Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.

Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., et al. (2023). Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv:2308.08708.

Christophel, T. B., Klink, P. C., Spitzer, B., Roelfsema, P. R., & Haynes, J. D. (2017). The distributed nature of working memory. Trends in Cognitive Sciences, 21(2), 111–124.

Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.

Collins, A. M., & Loftus, E. F. (1975). A spreading-activation theory of semantic processing. Psychological Review, 82(6), 407–428.

Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114.

Cowan, N. (2005). Working memory capacity. Psychology Press.

D’Esposito, M., & Postle, B. R. (2015). The cognitive neuroscience of working memory. Annual Review of Psychology, 66, 115–142.

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11, 127–138.

Goldman-Rakic, P. S. (1995). Cellular basis of working memory. Neuron, 14(3), 477–485.

Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022). Large language models are zero-shot reasoners. Advances in Neural Information Processing Systems, 35.

Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24, 167–202.

Mongillo, G., Barak, O., & Tsodyks, M. (2008). Synaptic theory of working memory. Science, 319(5869), 1543–1546.

Reser, J. E. (2016). Incremental change in the set of coactive cortical assemblies enables mental continuity. Physiology & Behavior, 167, 222–237.

Reser, J. E. (2022–2024). A cognitive architecture for machine consciousness and artificial superintelligence: Thought is structured by the iterative updating of working memory. arXiv:2203.17255.

Shanks, D. R. (2010). Learning: From association to cognition. Annual Review of Psychology, 61, 273–301.

Shannon, C. E. (1951). Prediction and entropy of printed English. Bell System Technical Journal, 30.

Stokes, M. G. (2015). Activity-silent working memory in prefrontal cortex: A dynamic coding framework. Trends in Cognitive Sciences, 19(7), 394–405.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.

Wang, X., Wei, J., Schuurmans, D., Le, Q. V., Chi, E. H., Narang, S., Chowdhery, A., & Zhou, D. (2023). Self-consistency improves chain-of-thought reasoning in language models. International Conference on Learning Representations.

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E. H., Le, Q. V., & Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35.