Jared Edward Reser, Ph.D.

Abstract

Brains and transformer language models both acquire knowledge by learning how an accumulated context predicts what should come next. In transformers, tokens are converted into context-sensitive representations, related through self-attention and other learned transformations, and used to predict the next token. Prediction error is then propagated through the network so that future instances of similar context produce better predictions. In the iterative updating model of working memory, several representations remain coactive in the focus of attention and short-term store. Their combined spreading activation conducts a multiassociative search for the next representation to enter working memory. The newly activated representation joins the retained context and participates in subsequent searches. This article proposes that repeated multiassociative search gives rise to multiassociative learning, in which the nervous system acquires context-dependent mappings from configurations of representations to useful successor representations. The relevant unit of learning is therefore not merely a pairwise association, (A \rightarrow B), but frequently a higher-order conditional relationship, ({A,C,D,E}\rightarrow B). Both brains and transformers gradually learn which representations are related, how their meanings change in combination, which elements deserve processing priority, and which information should remain available because it is expected to contribute to later predictions. This suggests that working memory is more than a temporary storage mechanism. It is a learned system for constructing and preserving the context from which adaptive updates can be generated. The article distinguishes the fast process of relational search from the slower process of structural learning, compares parameter learning with biological plasticity, and relates in-context learning to temporary cognitive adaptation. It then argues that reasoning and goal pursuit depend on learned trajectories through representational state space. Intermediate updates progressively construct a context in which an initially inaccessible answer, action, or solution becomes probable. The central conclusion is that intelligence depends partly on a system’s learned capacity to curate the context that will generate the right next thought.

Keywords: multiassociative search, multiassociative learning, self-attention, transformer, working memory, context-conditioned learning, neural plasticity, in-context learning, prediction, cognitive architecture

1. Introduction

The rise of transformer language models has shown that repeated next-token prediction can produce unexpectedly broad intellectual capabilities. A transformer is trained on sequences in which the correct continuation is available as a target. It predicts a probability distribution over possible next tokens, receives an error signal when its distribution differs from the observed continuation, and adjusts its parameters so that similar contexts will support better predictions in the future. Across many training examples, the model learns grammatical dependencies, semantic relationships, recurrent patterns, procedures, discourse structures, and regularities reflected in human descriptions of the world.

This learning process is usually discussed separately from biological learning. Transformer models are trained through gradient-based optimization, whereas brains change through synaptic plasticity, neuromodulation, reinforcement, and consolidation. Their representations, data sources, developmental constraints, and physical implementations differ considerably. These differences should not obscure a potentially important computational commonality. Both systems must learn how a distributed configuration of presently relevant information predicts an appropriate successor.

A companion analysis proposed that next-token prediction can be understood as a restricted linguistic instance of next-representation prediction. In the iterative updating model, the current contents of working memory jointly recruit a successor representation. Some previous contents persist, the selected representation joins them, and the revised state generates another update. Thought thereby proceeds autoregressively through partially overlapping states. The current article turns from the production of this sequence to the acquisition of the transition structure that makes the sequence intelligent.

The central proposal is that multiassociative search gives rise to multiassociative learning. Multiassociative search is the momentary process by which several coactive representations jointly contribute to the selection of a successor. Multiassociative learning is the slower process by which repeated configurations, transitions, prediction errors, and outcomes alter the network so that future contexts recruit more useful successors.

The distinction can be expressed simply:

[
\text{Multiassociative search}

\text{using a learned network to select the next representation}
]

[
\text{Multiassociative learning}

\text{changing the network so that later searches improve}
]

The corresponding distinction in transformers is between self-attention during a forward pass and parameter modification during training. Self-attention temporarily determines how contextual elements influence one another. Durable learning occurs when prediction error changes the parameters that will construct attention patterns and contextual representations during later encounters.

This comparison leads to a broader thesis. Brains and transformers do not merely learn which individual elements tend to occur together. They learn how entire configurations imply, modify, disambiguate, and predict one another. They also learn how to assemble a useful context: what should receive priority, what should remain available, what should be ignored, and what successor should be selected from the resulting configuration.

The problem of intelligence can therefore be divided into two related questions:

[
\boxed{\text{Given the present context, what should become active next?}}
]

and:

[
\boxed{\text{Which parts of the present context should remain available to guide later updates?}}
]

A system that learns both operations can construct extended trajectories toward answers, goals, and solutions.

2. Search and Learning Are Coupled but Distinct

Any adaptive predictive system must operate on at least two timescales. On a fast timescale, it uses its present knowledge to interpret an input and generate a response. On a slower timescale, it modifies that knowledge according to the success or failure of its predictions.

Let (C_t) denote the active context of a system at time (t), and let (M_t) denote its enduring learned structure. The fast operation can be written abstractly as:

[
r_{t+1}\sim P_{M_t}(r\mid C_t),
]

where (r_{t+1}) is the selected successor representation. The context is then updated:

[
C_{t+1}=U(C_t,r_{t+1}).
]

The slower learning operation modifies the enduring structure:

[
M_{t+1}=L(M_t,C_t,r_{t+1},O_t),
]

where (O_t) represents an observed continuation, outcome, prediction error, reward, or other evaluative signal.

The fast cycle uses the present network:

[
C_t + M_t \rightarrow r_{t+1}.
]

The slow cycle changes the future network:

[
C_t+r_{t+1}+O_t\rightarrow \Delta M.
]

These cycles form a reciprocal relationship:

[
\text{memory shapes search}
]

[
\text{search produces predictions and actions}
]

[
\text{outcomes evaluate those predictions and actions}
]

[
\text{learning modifies memory}
]

[
\text{modified memory shapes later search}.
]

This distinction is necessary because attention is sometimes described as if it were itself the learning mechanism. Attention determines which information influences a present computation. Learning determines how attention and other computations will be organized in the future.

In a transformer, attention weights are generated anew from the current token representations. They are transient features of the forward pass. The durable knowledge lies in the learned parameters used to construct queries, keys, values, feed-forward transformations, and output probabilities. The original transformer architecture formalized this process through learned query, key, and value projections combined with scaled dot-product attention. (arXiv)

In the brain model, spreading activation is likewise an online process. It uses existing synaptic structure to determine which inactive representations receive enough convergent support to become active. Durable learning occurs when coactivity, prediction error, reward, novelty, or other signals alter synaptic relationships and later transition probabilities.

The source architecture already articulates this progression. It proposes that every multiassociative search can expose previously separate representations to one another, modify associative relations, and improve subsequent processing. The paper summarizes the relation directly: multiassociative searching gives rise to multiassociative learning.  Repeated coactivity is then expected to strengthen statistical codependencies that persist as semantic and procedural knowledge.

3. Multiassociative Search in the Iterative Updating Model

The iterative updating model begins with a limited-capacity focus of attention embedded within a broader short-term store. The focus contains several coactive representations, while less active or activity-silent traces remain temporarily available outside it. The contents of these stores influence processing because their underlying neural populations continue to transmit activation through cortical and subcortical networks.

Suppose the focus of attention contains:

[
W_t={B,C,D,E}.
]

Each representation is individually associated with a large number of possible memories and responses. The important claim is that these active representations search cooperatively. Their combined excitatory and inhibitory effects converge on candidate additions. A representation (F) may receive enough collective support to become active even when no single item would have recruited it independently:

[
P(F\mid B,C,D,E)
\gg
P(F\mid B),P(F\mid C),P(F\mid D),P(F\mid E).
]

Once (F) becomes active, it joins representations that remain from the preceding state:

[
W_{t+1}={C,D,E,F}.
]

This revised configuration then performs another search:

[
{C,D,E,F}\rightarrow G.
]

The architecture calls this operation multiassociative search. Its 2016 precursor used the term polyassociativity to describe coactive assemblies pooling their activity and jointly determining which ensemble becomes active next.  The later paper describes active items as spreading activation through the same global network and converging in parallel on the next addition to working memory.

The process is more sophisticated than a simple chain of associations:

[
A\rightarrow B\rightarrow C\rightarrow D.
]

It is instead a sequence of configuration-dependent transitions:

[
{A,B,C,D}\rightarrow E
]

[
{B,C,D,E}\rightarrow F
]

[
{C,D,E,F}\rightarrow G.
]

Every newly selected item changes the set of conditions responsible for the next selection. The system therefore traverses a probability landscape whose local structure is continually altered by its own updates.

The active representations are also not treated as immutable symbols. An ensemble is described as a fuzzy, distributed set whose exact composition can vary across activations. Its effective meaning changes with the other representations active at the same time.  The multiassociative search algorithm further proposes that new additions redistribute activation among existing items and alter the composition of the assemblies constituting them.

This means that the system does not simply retrieve a successor from fixed meanings. It first constructs context-dependent versions of its present representations, then uses those contextualized representations to select the update.

4. Relational Integration in Transformer Self-Attention

A transformer begins with token embeddings and positional information. At each attention layer, a token position is mapped into query, key, and value vectors. For a particular attention head (h) and layer (\ell), the relation between positions (i) and (j) can be represented as:

[
A_{ij}^{(\ell,h)}

\operatorname{softmax}{j}
\left(
\frac{q_i^{(\ell,h)}{k_j^{(\ell,h)}}^\top}
{\sqrt{d_k}}
+
m{ij}
\right),
]

where (m_{ij}) imposes the causal restriction that a decoder position cannot use future tokens. The resulting representation at position (i) includes a weighted combination of value vectors:

[
z_i^{(\ell,h)}

\sum_{j\leq i}
A_{ij}^{(\ell,h)}v_j^{(\ell,h)}.
]

In a dense causal transformer, each position can therefore evaluate its relationship to every causally available earlier position. Multihead attention performs this operation through several learned projection spaces, while successive layers repeatedly contextualize the representations. The final hidden state contributes to the model’s probability distribution over the next token. (arXiv)

It is slightly misleading to say that some tokens are simply “in attention” while others are excluded. Attention is ordinarily graded, query-dependent, head-dependent, and layer-dependent. A token can strongly affect one computation while contributing very little to another. The same token can participate through different relational features in different heads.

The relationship detected by attention need not be simple semantic similarity. A contextual position may be useful because it supplies:

• the referent of a pronoun,
• the subject of a verb,
• the beginning of a repeated pattern,
• the premise of an argument,
• the definition of a term,
• the condition of a rule,
• the location of a previously introduced object,
• the value assigned to a variable,
• the goal of a procedure,
• or a contrast that changes the interpretation of another item.

Self-attention therefore performs a form of learned relational integration. Several contextual positions contribute unequally to a transformed representation, and that representation helps determine the next-token distribution.

The resemblance to multiassociative search lies in this many-to-one organization. In both cases, several active contextual elements jointly contribute to a successor computation. In both cases, the significance of each element depends on the other elements present. In both cases, learned relationships determine which portions of the context exert the strongest influence.

The iterative updating architecture explicitly draws this comparison. It proposes that a newly activated working-memory item redistributes activation among the representations already present, changing how strongly each contributes to the next search. It then compares this operation with query-key similarity in self-attention, where contextual tokens receive different weights when determining the next-token probability distribution.

A careful formulation is therefore:

Transformer self-attention and cortical multiassociative spreading activation are distinct mechanisms for context-dependent, many-to-one relational integration.

This does not imply that cortical cells calculate dot products or that attention heads correspond to individual brain regions. It identifies a shared computational problem and a partially similar solution.

5. Transformer Training as Context-Conditioned Learning

During autoregressive training, a transformer is given a token sequence:

[
x_1,x_2,\ldots,x_t,x_{t+1}.
]

The model uses the preceding context to predict the next token:

[
P_\theta(x_{t+1}\mid x_{\leq t}).
]

The observed token (x^*_{t+1}) supplies the training target. A common loss is the negative log probability assigned to that target:

[
\mathcal{L}_t

-\log P_\theta(x^*{t+1}\mid x{\leq t}).
]

Backpropagation estimates how the model’s parameters contributed to the loss, and an optimizer adjusts them:

[
\theta_{n+1}

\theta_n-\eta\nabla_\theta\mathcal{L}.
]

The network is therefore repeatedly asked to answer:

Given this configuration of context, what continuation should be expected?

Across training, it becomes better at constructing hidden representations and relational weightings that support accurate answers.

The learned structure goes beyond a record of token adjacency. A token’s contribution is conditioned on position, surrounding tokens, and representations constructed in prior layers. The network learns statistical dependencies distributed across varying distances and levels of abstraction. The same token can therefore make different contributions in different contexts.

For example, bank participates in one relational configuration in:

The canoe drifted toward the bank.

and another in:

The company applied to the bank for a loan.

The model does not need two entirely separate vocabulary items. It learns to construct different contextualized hidden states from the surrounding sequence. Those states then support different continuations.

The training objective supplies pressure to learn any regularity that reduces uncertainty about the next token. Some of these regularities are linguistic, while others concern the world described by language. Text records relations involving objects, agents, causes, temporal sequences, tools, intentions, institutions, physical events, and social practices. A model trained on sufficiently varied language can acquire partial internal structure corresponding to these latent regularities because doing so improves prediction.

This does not mean that all world knowledge can be recovered from text or that linguistic learning is equivalent to embodied development. It means that next-token prediction supplies a general error signal capable of shaping representations of any pattern that reliably influences linguistic continuation.

6. From Pairwise Association to Multiassociative Learning

Classical association is often represented as a relation between two elements:

[
A\rightarrow B.
]

This form is useful but insufficient for explaining many context-sensitive transitions. A representation may imply one successor in one configuration and a different successor in another:

[
{A,C,D}\rightarrow B
]

[
{A,E,F}\rightarrow G.
]

The meaning and predictive significance of (A) are therefore conditional on the other active elements.

Consider the word current. In one context:

[
{\text{current},\text{river},\text{water},\text{flow}}
]

it refers to directional movement of water. In another:

[
{\text{current},\text{wire},\text{voltage},\text{circuit}}
]

it refers to electrical charge flow. In a third:

[
{\text{current},\text{news},\text{recent},\text{events}}
]

it indicates temporal recency.

A capable system does not merely learn three independent pairwise links. It learns how each configuration changes the effective interpretation of the shared element and how the resulting configuration constrains later representations.

This can be described as configuration-to-successor learning:

[
\mathcal{C}\rightarrow r,
]

where (\mathcal{C}) is a structured set of contextual representations and (r) is a contextually appropriate successor.

Multiassociative learning is the biological version proposed here. When several representations become coactive and jointly recruit another representation, the transition creates an opportunity for context-dependent plasticity. If the configuration and result recur, or if the transition receives reinforcing feedback, the network becomes more likely to reproduce the appropriate successor under similar conditions.

A simple conceptual learning rule would be:

[
\Delta M

\Phi(W_t,r_{t+1},\delta_t,\rho_t),
]

where:

• (W_t) is the configuration of active representations,
• (r_{t+1}) is the recruited update,
• (\delta_t) is a prediction-error signal,
• (\rho_t) represents reward, salience, novelty, or behavioral relevance,
• and (\Phi) denotes the plasticity process.

The function need not strengthen every pair of coactive items equally. It may preferentially alter connections that contributed to the transition, were temporally eligible, participated in a bound representation, or were tagged by reinforcement. The resulting association remains dependent on the broader context.

The 2016 account anticipated this form of configuration-sensitive learning. It argued that two ensembles repeatedly active together would become more likely to recruit one another, especially in the presence of the other contextual ensembles with which they had historically coactivated.  The later architecture similarly proposes that coactive configurations form statistical codependencies and that new learning retunes relationships among groups of items rather than simply replacing old memories.

This distinction is important because much of human knowledge is conditional. A behavior that is adaptive under one set of circumstances may be harmful under another. A word, object, person, rule, or action has no single consequence independent of context. Multiassociative learning allows the nervous system to preserve this conditionality.

7. Learning Contextual Meaning

A further implication is that learning alters not only which representation comes next, but also what the currently active representations mean.

Suppose a system contains a representation (r_i). Its context-independent identity may be relatively stable, but its effective state at time (t) is produced through interaction with the surrounding context:

[
\tilde{r}_{i,t}

f(r_i,C_t).
]

The contextualized representation (\tilde{r}_{i,t}) can differ substantially across situations. In a transformer, this transformation is produced through repeated attention, feed-forward processing, residual integration, and normalization. In the proposed brain model, it may arise through changes in the neural assemblies participating in an ensemble, recurrent interactions among coactive ensembles, top-down feedback, and current sensory input.

The system consequently learns two related mappings:

[
(r_i,C_t)\rightarrow \tilde{r}_{i,t}
]

and:

[
{\tilde{r}{1,t},\tilde{r}{2,t},\ldots}
\rightarrow r_{t+1}.
]

The first mapping contextualizes the present. The second predicts a successor from the contextualized present.

This cycle helps explain how distributed statistical learning can produce behavior that appears symbolic. At the psychological level, a person may seem to apply a rule such as:

When condition (A), goal (B), and constraint (C) are present, perform operation (D).

At the neural level, the rule may be implemented by a learned transition:

[
P(D\mid A,B,C)\gg P(D\mid A)
]

without a separate symbolic sentence encoding the rule.

Transformers provide an engineering demonstration that distributed networks can learn such conditional behavior. Their internal vectors are subsymbolic, yet their outputs can express categories, procedures, grammatical rules, and structured arguments. The existence of these abilities does not establish how brains implement the same functions, but it supports the broader claim that context-conditioned statistical learning can give rise to apparently rule-governed cognition.

8. Learning What Should Remain Active

The system must learn more than which successor follows a context. It must also learn how to preserve the information that will make later predictions possible.

Working memory is capacity-limited. A brain cannot maintain every recently active representation at equal priority. Some items must remain strongly active, some can be demoted to a less active short-term state, and others must be allowed to decay.

This creates a retention problem:

[
\text{Which representation is likely to remain useful?}
]

Let (r_i) be a representation presently available to working memory. A retention policy can be written abstractly as:

[
k_{i,t}

\pi_{\text{keep}}(r_i\mid C_t,G_t),
]

where (G_t) includes current goals and (k_{i,t}) specifies the degree to which the representation should remain available.

The value of retaining an item depends partly on its expected future contribution:

[
V_{\text{retain}}(r_i)

\mathbb{E}
\left[
\sum_{n=1}^{N}
\gamma^n
U(r_i,C_{t+n},G_t)
\right],
]

where (U) denotes the item’s usefulness to later processing. This equation is a computational abstraction rather than a claim that neurons explicitly calculate a discounted value function.

The important point is prospective:

A representation should remain active when the system has learned that it is likely to help interpret, predict, or control what happens later.

This interpretation changes the meaning of attention. Attention is not solely a judgment about present importance. It is also a prediction about future processing value.

The 2016 model proposed that prefrontal neurons become tuned through experience to determine which aspects of a situation should be maintained for their processing utility in the immediate future. Sentence comprehension was offered as an example. Relevant representations must be identified and preserved long enough for later portions of the sentence to be interpreted correctly.

This proposal is compatible with influential accounts of prefrontal control. Miller and Cohen argued that actively maintained prefrontal patterns representing goals and means can bias processing pathways throughout the brain. (Annual Reviews) O’Reilly and Frank developed a computational model in which prefrontal and basal-ganglia systems learn strategic, task-appropriate control over working memory through reinforcement-sensitive gating. (PubMed)

The transformer analogue is partial. A standard causal transformer ordinarily retains earlier tokens until they leave the context window. It does not literally decide which tokens continue to exist. It does, however, learn which earlier positions should exert effective influence on current processing. Attention heads and hidden transformations can strongly prioritize some contextual information while making other information functionally negligible.

The contrast can be expressed as follows:

[
\text{Transformer: learns what retained context to use}
]

[
\text{Brain: learns what to use and what to continue retaining}.
]

This difference suggests an avenue for artificial cognitive architecture. A system with an explicit, limited working-memory store could learn to preserve, summarize, demote, or discard information according to its expected value for future updates.

9. Three Interdependent Learning Problems

The preceding analysis suggests that an intelligent system must solve at least three learning problems.

9.1 Learning Useful Representations

The system must discover recurrent patterns worth representing. These may include objects, actions, agents, categories, locations, causes, rules, goals, and relations. In transformers, these structures are distributed across learned embeddings, attention projections, feed-forward networks, and higher-order activation patterns. In brains, they are thought to depend on distributed neural assemblies, ensembles, and interactions among specialized cortical and subcortical systems.

Representation formation determines what elements are available for thought.

9.2 Learning What to Maintain

The system must estimate which representations will remain relevant. This determines what information survives long enough to interact with later input, later memories, and intermediate results.

Retention determines the temporal reach of context.

9.3 Learning What to Recruit Next

The system must use the maintained configuration to select a contextually appropriate successor:

[
r_{t+1}
\sim
\pi_{\text{update}}(W_t,H_t,S_t,G_t).
]

Here (W_t) is the focus of attention, (H_t) is the broader short-term store, (S_t) is current sensory input, and (G_t) contains goals and motivational variables.

Successor selection determines the direction in which thought moves.

These operations are mutually dependent. Poor representations cannot support precise retention or prediction. Poor retention deprives later searches of essential constraints. Poor update selection sends the cognitive trajectory into irrelevant regions of representational space.

A compact formulation is:

[
\boxed{
\text{Intelligence requires learning what to represent, what to retain, and what to recruit next.}
}
]

10. Developmental Learning in Brains and Transformers

Both brains and transformers begin as relatively poor predictors and improve through repeated exposure. The similarity is functional rather than ontogenetic. A transformer’s sequence of optimization steps is not a childhood, and biological maturation includes genetic, anatomical, hormonal, sensorimotor, and social processes absent from ordinary model training. Nevertheless, both systems gradually organize distributed networks around recurring conditional structure.

A developing brain receives continuous multimodal input. It learns from perception, movement, bodily consequences, reward, pain, social interaction, exploration, error, and internally generated activity. The learner helps determine its own future data by acting on the world.

A transformer is commonly trained on externally assembled sequences. It receives the observed next token as a target and adjusts its parameters according to prediction error. The network gradually acquires relational structure because any dependable regularity that constrains linguistic continuation can reduce its loss.

The parallel lies in cumulative conditional learning:

[
\text{experience}
\rightarrow
\text{prediction}
\rightarrow
\text{error or outcome}
\rightarrow
\text{network modification}
\rightarrow
\text{improved future prediction}.
]

Over time, both systems piece together regularities that were not available from any single example. An isolated sentence, perception, or event may support only a weak inference. Recurrence across many contexts allows stable abstractions to emerge.

The source architecture proposes a developmental sequence in which learning initially emphasizes events close together in time. As working-memory persistence increases, the system becomes capable of relating events separated by longer delays and constructing higher-order dependencies.  This is presented as an architectural hypothesis rather than an established account of neurodevelopment.

The idea has a useful analogue in machine learning. Early in training, a model may first capture local and frequent regularities. As its internal representations improve, it becomes increasingly capable of exploiting longer dependencies and more abstract configurations. Mechanistic work on induction heads has reported that attention circuits capable of matching and continuing repeated patterns can emerge during training and coincide with improvements in in-context prediction, although the authors describe the evidence as stronger for small attention-only models than for large models with multilayer perceptrons. (arXiv)

The deeper commonality is that neither learner receives a finished ontology. Categories and relationships must be inferred gradually from statistical structure. The network learns what tends to remain invariant, what changes with context, what predicts later events, and what information should receive priority when similar situations recur.

11. Prediction Error, Reward, and Credit Assignment

A learner must determine not only that an outcome occurred, but which portions of the preceding computation contributed to it. This is the problem of credit assignment.

In a transformer, backpropagation assigns gradients to parameters throughout the computational graph. The gradient indicates how small parameter changes would have altered the loss. Repeated optimization distributes knowledge across the network so that relevant contextual relationships become more likely to influence future predictions.

Biological credit assignment is constrained by local connectivity, timing, cell physiology, neuromodulation, and multiple forms of plasticity. Brains are not known to implement ordinary backpropagation through the exact operations used by transformers. The computational problem is nevertheless comparable:

Which active relationships contributed to a successful or unsuccessful prediction, and how should the system change?

A generic biological formulation might be:

[
\Delta w_{ij}

\eta
\cdot
e_{ij,t}
\cdot
\delta_t,
]

where (e_{ij,t}) is an eligibility factor reflecting recent participation of a connection and (\delta_t) is a teaching or prediction-error signal. The equation is deliberately general and can encompass several possible plasticity mechanisms.

Dopaminergic responses are widely associated with errors in reward prediction, making them plausible contributors to the reinforcement of transitions that lead toward valuable outcomes. (Science) The iterative updating architecture further proposes that dopamine-like modulation can increase the persistence of representations related to novelty, opportunity, threat, or unresolved goals, allowing them to influence a longer series of searches.

This creates a link between learning and working-memory maintenance. Reward does not merely reinforce an overt action. It may also reinforce the internal conditions that preceded the action:

• which features were attended,
• which representations were maintained,
• which memory was retrieved,
• which interpretation was selected,
• and which intermediate update redirected the trajectory.

A complete theory of multiassociative learning therefore requires multiassociative credit assignment. The system must determine which subset of a large active configuration was responsible for an adaptive transition.

12. Why Co-Occurrence Alone Is Insufficient

Two items can be simultaneously active without having a meaningful relationship. A person can shift abruptly between unrelated topics, causing remnants of the first thought to overlap briefly with the second. Similarly, tokens can appear in the same document for incidental reasons.

If all coactive elements became strongly associated, memory would accumulate large numbers of spurious relationships. Intelligent learning therefore requires mechanisms that distinguish informative coactivity from accidental proximity.

In transformer training, co-occurrence affects learning only through its contribution to prediction error. Two tokens are not connected by a simple direct rule merely because they appeared in the same sequence. Their influence is distributed through many parameters and shaped by whether using their relationship improves predictions across examples.

The brain likely uses several filters:

• temporal coordination,
• focused attention,
• neural binding,
• behavioral relevance,
• novelty,
• reward,
• prediction error,
• repeated recurrence,
• and biochemical eligibility or tagging.

The 2016 paper explicitly recognizes the problem. It observes that unrelated thoughts can occur close together and argues that the brain must prevent all temporally adjacent representations from wiring together indiscriminately. Dopaminergic processes, binding, and molecular tagging are suggested as possible safeguards.

Multiassociative learning should therefore be defined as more than learning from coactivation:

Multiassociative learning is the selective modification of context-dependent relationships among coactive representations according to their predictive, causal, or behavioral value.

This definition preserves the role of association while avoiding a simple contiguity account.

13. In-Context Learning and Temporary Cognitive Adaptation

Transformers exhibit two importantly different forms of adaptation.

13.1 Parameter Learning

During pretraining or fine-tuning, the model’s parameters change. The resulting adaptation remains after the training example leaves the context. This is durable learning.

13.2 In-Context Learning

During inference, instructions and examples within the prompt can alter the model’s behavior without changing its underlying parameters. GPT-3 demonstrated task adaptation from descriptions and examples supplied entirely through text, without gradient updates during the evaluated task. (arXiv)

In-context learning is therefore not learning in the same durable sense as parameter modification. It is a temporary reorganization of behavior produced by an active context. The prompt establishes a task, mapping, convention, or local regularity, and the model uses its preexisting parameters to adapt its current predictions.

A functional brain comparison is:

[
\text{in-context learning}
\approx
\text{temporary working-memory adaptation}
]

[
\text{parameter learning}
\approx
\text{long-term synaptic learning}.
]

A person can follow a newly stated rule before it has been consolidated into long-term memory. The instructions remain active or readily retrievable, and their presence changes how existing knowledge is applied. When the temporary context disappears, performance may decline.

The brain also contains intermediate forms of retention. Information can leave the focus of attention while remaining temporarily potentiated or primed for reactivation. The cognitive architecture associates focal maintenance with sustained firing and the broader short-term store with temporary synaptic potentiation.  Synaptic theories of working memory similarly propose that short-lived changes in synaptic efficacy can preserve information without continuous elevated firing. (Science)

This suggests a three-level comparison:

Timescale

Transformer

Brain

Immediate processing

Current attention and hidden activations

Current neural activity and focal attention

Temporary adaptation

Prompt-conditioned behavior

Working memory, priming, temporary potentiation

Durable learning

Parameter updates

Long-term synaptic and systems-level plasticity

The mapping is approximate, but it clarifies why temporary context can produce apparently learned behavior without changing permanent knowledge.

14. Learning to Construct a Predictive Context

The strongest claim of this article is that learning concerns not only associations and predictions, but the construction of the context from which prediction occurs.

Suppose a problem initially produces the state:

[
W_0={\text{goal},\text{known facts},\text{constraints}}.
]

The final solution may have very low probability under that initial state:

[
P(\text{solution}\mid W_0)\approx 0.
]

The system selects an intermediate representation:

[
W_0\rightarrow r_1.
]

The update changes the context:

[
W_1=U(W_0,r_1).
]

A second intermediate representation becomes available:

[
W_1\rightarrow r_2.
]

After several transitions:

[
W_0\rightarrow W_1\rightarrow W_2\rightarrow\cdots\rightarrow W_n,
]

the solution may become strongly supported:

[
P(\text{solution}\mid W_n)
\gg
P(\text{solution}\mid W_0).
]

The system has not simply retrieved the answer. It has constructed the context in which the answer can be selected.

This gives intermediate thought a precise computational role. Each useful intermediate representation alters the conditions governing future search. It may introduce a distinction, recover a missing fact, create a subgoal, resolve an ambiguity, eliminate an alternative, or preserve an intermediate result.

The iterative updating model describes reasoning in these terms. It proposes that each state can encode the current status of a problem and that successive associative predictions can produce a chain of intermediate states progressing toward a terminal solution.

Learning must therefore extend beyond local correctness. The system must learn which intermediate updates tend to create productive later contexts.

15. From Immediate Prediction to Goal-Directed Updating

A distinction is needed between the most probable immediate continuation and the update most useful for reaching a distant goal.

A standard pretrained language model is directly trained to predict the observed next token:

[
\max_\theta
P_\theta(x_{t+1}\mid x_{\leq t}).
]

A goal-directed cognitive system must sometimes select an update according to its longer-term utility:

[
r_{t+1}

\arg\max_r
\mathbb{E}
\left[
R_{t+1}
+
\gamma R_{t+2}
+\cdots+
\gamma^n R_{t+n}
\mid C_t,r
\right].
]

The update that produces the best eventual result may not be the most immediately familiar association. Deliberation may require suppressing a prepotent continuation, preserving an unresolved question, exploring an alternative, or selecting a step whose significance becomes apparent only later.

Language-model pretraining nevertheless provides indirect pressure toward useful trajectories. Explanations, derivations, programs, narratives, and arguments contain sequences in which intermediate elements support later conclusions. Predicting such text requires learning regularities of successful progression. Post-training, reinforcement, process feedback, tool use, and search can provide more direct pressure toward long-horizon outcomes.

Biological cognition receives stronger intrinsic goal signals. Needs, rewards, threats, social consequences, and persistent objectives can remain active while intermediate representations change. Miller and Cohen’s theory describes maintained goal representations as bias signals that direct the flow of activity toward task-appropriate mappings. (Annual Reviews) In the iterative updating model, the most persistent representations can function as the enduring topic or objective around which other contextual elements are revised.

This suggests a hierarchy of update policies:

[
\text{next probable token}
]

[
\text{next contextually relevant representation}
]

[
\text{next useful representation}
]

[
\text{next representation expected to advance a persistent goal}.
]

The final level most closely approximates deliberative thought.

16. Learning Trajectories Through Representational Space

A learned network can be represented as a probability landscape over possible cognitive states. The active context occupies a region of this landscape, and multiassociative search determines likely transitions to neighboring regions.

Learning changes the landscape. Repeatedly successful transitions become easier to traverse, while unsuccessful or contradictory transitions become less accessible. A familiar procedure can eventually become a steep and reliable route:

[
W_0\rightarrow W_1\rightarrow W_2\rightarrow W_{\text{goal}}.
]

An unfamiliar problem presents a flatter, more uncertain landscape with many competing paths.

A solution can be treated as an attractor state, meaning a configuration toward which several different starting contexts can converge. The cognitive architecture describes iterative thought as tending toward stable combinations representing reliable beliefs, truths, or solutions.

This landscape metaphor unifies the present concepts:

• Long-term learning shapes the landscape.
• Attention weights the currently relevant dimensions.
• Working memory preserves the coordinates needed to remain on a trajectory.
• Multiassociative search selects the next movement.
• Iterative updating carries the system through the landscape.
• Reward and error reshape the routes taken in the future.

Intelligence then depends partly on the ability to enter and sustain productive trajectories.

17. Three Nested Timescales of Intelligence

Brains and transformers can both be analyzed at three nested temporal levels.

17.1 Relational Integration Within an Update

At the shortest level, several contextual elements interact to construct a current representation. In a transformer, multiple attention heads and feed-forward operations transform token representations. In the brain, recurrent and spreading neural activity may integrate coactive representations across cortical and subcortical systems.

17.2 Iterative Updating Across a Sequence

The product of one computation becomes part of the input to the next:

[
C_t\rightarrow r_{t+1}\rightarrow C_{t+1}.
]

This produces a sequence of tokens in a language model and a sequence of internal representations in the proposed brain architecture.

17.3 Structural Learning Across Experiences

Repeated sequences and their outcomes alter the network:

[
M_t\rightarrow M_{t+1}.
]

This determines how later contexts will be interpreted, maintained, and extended.

The full cycle can be summarized as:

[
\boxed{
\text{integrate context}
\rightarrow
\text{select an update}
\rightarrow
\text{revise context}
\rightarrow
\text{evaluate the transition}
\rightarrow
\text{modify the network}
}
]

The inner cycles produce cognition. The outer cycle teaches the system how to produce better cognition.

18. A Unified Context-Conditioned Learning Algorithm

The common computational structure can be represented as a seven-stage process.

Stage 1: Activate a Context

Information from perception, memory, prior outputs, goals, or instructions becomes available:

[
C_t={r_1,r_2,\ldots,r_n}.
]

Stage 2: Contextualize the Components

Each representation is transformed by its relationships to the others:

[
\tilde{r}_{i,t}=f(r_i,C_t).
]

Stage 3: Assign Processing Priority

The system determines which contextualized representations should contribute most strongly:

[
a_{i,t}=\pi_{\text{priority}}(\tilde{r}_{i,t},C_t,G_t).
]

Stage 4: Select a Successor

The weighted context generates a probability distribution over possible updates:

[
r_{t+1}
\sim
P(r\mid \tilde{C}_t,G_t,M_t).
]

Stage 5: Revise the Context

The new representation is added while some earlier contents persist and others lose priority:

[
C_{t+1}=U(C_t,r_{t+1}).
]

Stage 6: Evaluate the Transition

The system receives an observed continuation, reward, error, or measure of goal progress:

[
O_t=\operatorname{Evaluate}(C_t,r_{t+1}).
]

Stage 7: Modify Future Processing

The enduring network changes:

[
M_{t+1}=L(M_t,C_t,r_{t+1},O_t).
]

This algorithm is substrate-neutral. Transformers and brains instantiate its stages through different mechanisms. The comparison concerns the organizational pattern rather than an identity of implementation.

19. Major Similarities

The strongest similarities are computational.

Distributed Representation

Both systems represent knowledge across networks rather than storing every concept or rule in a single isolated location.

Context Dependence

The effective meaning and influence of an element depend on the other elements active with it.

Many-to-One Integration

Several contextual representations jointly contribute to the selection of a successor.

Prediction-Based Organization

The system is organized around reducing uncertainty about what comes next.

Recursive Updating

Each selected output changes the context used for later selection.

Experience-Dependent Transition Structure

Repeated examples and outcomes alter the probability of later transitions.

Temporary and Durable Adaptation

Both systems exhibit transient changes in processing state and slower changes in enduring structure.

These similarities support the proposal that brains and transformers solve a common abstract problem:

They must learn how distributed configurations of currently relevant information predict appropriate successor states.

20. Major Differences and Limits of the Comparison

The mechanisms remain substantially different.

20.1 Backpropagation and Biological Plasticity

Transformers are commonly trained by global gradient-based optimization. Biological plasticity is locally constrained and depends on cellular, synaptic, neuromodulatory, and systems-level processes. The existence of a shared credit-assignment problem does not establish a shared solution.

20.2 Token Sequences and Multimodal Experience

Transformers typically learn from discrete tokenized sequences. Brains learn from continuous vision, audition, touch, proprioception, interoception, action, reward, and social interaction.

20.3 Fixed Context Availability and Selective Maintenance

A standard transformer keeps tokens available within its context window, although their effective influence varies. A brain actively maintains, demotes, reactivates, and forgets representations across several timescales.

20.4 External Training Objective and Biological Value

A transformer receives an engineered loss function. A brain’s learning is constrained by survival, bodily regulation, reward, punishment, attachment, curiosity, and social consequences.

20.5 Offline Training and Continuous Learning

Many deployed transformers do not modify their core parameters during ordinary interaction. Brains continue learning throughout life.

20.6 Linguistic Output and Internal Cognitive Updating

A language model ordinarily expresses each autoregressive update as a token. The brain’s updates can remain internal and take visual, spatial, emotional, episodic, procedural, or motor forms.

20.7 Attention Weights and Neural Influence

Attention scores are mathematically explicit coefficients within a model layer. Neural relevance is distributed across firing rates, timing, connectivity, oscillations, synaptic states, neuromodulation, and broader circuit dynamics.

These differences mean that self-attention should not be identified literally with multiassociative spreading activation. The more defensible conclusion is that they provide distinct implementations of context-sensitive relational selection.

The source neurobiological model is also explicitly exploratory. Its 2016 presentation describes unverified assumptions and calls for direct experimental tests of incremental changes in coactive neural populations.  The transformer comparison strengthens the computational plausibility of the model but does not verify its proposed neural mechanism.

21. Testable Predictions for Neuroscience

The theory generates several empirical predictions.

21.1 Configurations Should Predict Successors Better Than Individual Items

The next decoded working-memory representation should be better predicted from the joint configuration of active contents than from any single content alone:

[
I(r_{t+1};W_t)

\max_i I(r_{t+1};r_{i,t}),
]

where (I) denotes mutual information.

This could be tested using tasks in which several clues jointly determine a unique concept while each clue alone remains ambiguous.

21.2 Retention Should Reflect Expected Future Relevance

Representations that will become useful later in a task should be maintained or remain more readily reactivatable than equally salient representations with no later relevance. This effect should be enhanced by explicit goals and prior learning.

21.3 Learning Should Be Configuration-Specific

Training a transition:

[
{A,B,C}\rightarrow D
]

should increase the probability of recruiting (D) most strongly when the full or functionally equivalent configuration is reinstated. Pairwise presentation of (A) or (B) alone should produce a weaker effect.

21.4 Reward Should Alter Internal Retention and Search

Reinforcement should change not only overt response probabilities, but also which representations are maintained and which internal successors are recruited during later trials.

21.5 Errors Should Follow Contextual Failures

Reasoning errors should often be traceable to one of three failures:

1. a relevant representation was not activated,
2. it was activated but lost before it became useful,
3. an inappropriate intermediate update redirected the trajectory.

21.6 Development Should Expand Capturable Dependencies

As working-memory control matures, learners should become better at discovering relationships separated by longer delays, while still using shorter-range regularities as a scaffold.

The cognitive architecture previously proposed time-series recordings and multivariate decoding to identify partial addition and subtraction of working-memory contents.  The present theory adds a further requirement: researchers should test whether the retained configuration predicts the identity of the next recruited representation.

22. Testable Predictions for Artificial Intelligence

Artificial systems allow direct manipulation of the proposed variables.

22.1 A Bounded Latent Focus May Improve Context Management

A model with a small, dynamically updated latent focus of attention may outperform an undifferentiated long context on tasks requiring resistance to distraction, maintenance of a goal, and preservation of intermediate results.

22.2 Learned Retention Should Improve Long-Horizon Reasoning

A system trained to estimate the future usefulness of representations should preserve task-relevant variables more reliably than a system using recency-based retention alone.

22.3 Configuration Learning Should Exceed Pairwise Learning

Models explicitly optimized to preserve higher-order contextual interactions should perform better on ambiguity, conditional rules, compositional reasoning, and situations in which no individual cue is sufficient.

22.4 Internal Updates Should Reduce Dependence on Verbalized Reasoning

A recurrent architecture that feeds latent representations back into a workspace should solve some problems with fewer overt tokens than a system required to express every intermediate state linguistically.

22.5 Outcome Feedback Should Reshape Context Curation

Training on final-answer accuracy alone should alter some retention strategies, but process-level feedback about useful intermediate states should produce more reliable context management.

22.6 Trajectory Metrics Should Predict Success

Useful measurements would include:

• persistence of goal representations,
• overlap between adjacent states,
• rate of context turnover,
• entropy of candidate updates,
• frequency of full context resets,
• return to suspended subproblems,
• and the amount of joint context needed to predict the next state.

These metrics could be compared across artificial systems and, where decoding permits, across brains.

23. Implications for Cognitive Architecture

The comparison suggests that extending a transformer into a more brain-like cognitive agent requires more than increasing context length. The system would need to learn how to manage several forms of temporary representation.

A possible architecture would contain:

1. A small focus of attention for intensive interaction among currently prioritized representations.
2. A broader short-term store containing recently relevant but less active information.
3. A long-term associative network capable of content-addressable retrieval.
4. A recurrent next-representation predictor that selects internal updates without requiring immediate linguistic output.
5. A learned retention policy that preserves representations according to expected future utility.
6. Persistent goal and value representations that bias a long series of updates.
7. Multimodal modules capable of contributing visual, auditory, spatial, motor, and semantic candidates.
8. Outcome-sensitive plasticity that changes both successor selection and context-maintenance policies.
9. Thread suspension and reinstatement so that subproblems can be solved and later merged.
10. Continuous learning from the consequences of perception, thought, and action.

The original architecture already proposes that a transformer’s attention and context windows may serve as rough analogues of focal and broader short-term storage, while emphasizing that a small brain-like focus could improve coherence.  It also compares transformer prediction with multiassociative search and argues for recycling internal hidden representations rather than only output tokens.

The present analysis adds that the architecture must learn its own context-management policy. An intelligent workspace cannot rely solely on a fixed capacity or a hand-coded eviction rule. It must learn which representations will help generate valuable future updates.

24. Implications for Understanding Intelligence

The proposed synthesis yields a compact account of intellectual development.

An intelligent system gradually learns:

1. what recurring structures deserve representation,
2. which elements tend to occur together,
3. how their meanings change in combination,
4. which configurations predict which continuations,
5. which information should remain active,
6. which updates advance a goal,
7. and which trajectories lead to reliable solutions.

The first four construct a model of conditional structure. The final three turn that model into a controlled thought process.

This reframes intelligence as more than possession of knowledge. A system may contain the information needed to solve a problem but fail because the correct information is not activated together, is not maintained long enough, or does not recruit the necessary intermediate update.

Accordingly:

[
\text{knowledge}
\neq
\text{effective thought}.
]

Effective thought additionally requires:

[
\text{knowledge}
+
\text{context construction}
+
\text{retention}
+
\text{successor selection}
+
\text{iterative progression}.
]

A concise approximation would be:

[
\text{Intelligence}
\approx
f(
\text{representation quality},
\text{context quality},
\text{retention quality},
\text{transition quality},
\text{feedback},
\text{productive iterations}
).
]

This explains why working-memory capacity and attentional control are so important even when long-term knowledge is extensive. Intelligence depends on placing the right knowledge into causal contact at the right time.

25. Conclusion

Brains and transformers learn by repeatedly confronting the same general problem: given a structured context, what should happen next? Their implementations differ, but both systems gradually acquire distributed mappings between configurations of information and appropriate continuations.

In transformers, self-attention and related network operations construct a context-sensitive representation from tokens. The model predicts the next token, receives error feedback, and changes its parameters so that later contexts are interpreted more effectively. In the iterative updating model, representations in working memory jointly spread activation and recruit a next representation. Coactivity, prediction error, reinforcement, and plasticity then alter the network so that future searches become more adaptive.

The resulting learning is multiassociative because the relevant unit is often a configuration rather than a pair. An element can imply different successors under different conditions, and its effective meaning can change according to the representations active around it. Both brains and transformers must therefore learn conditional relationships among distributed contextual states.

The comparison also reveals that an intelligent learner must acquire a policy for managing context. It must learn what should be represented, what should receive priority, what should remain available, and what successor should be selected. A representation is worth maintaining when it is expected to contribute to a later interpretation, prediction, action, or solution.

Reasoning follows when this process is iterated. An initial problem state generates an intermediate update. That update changes the context, allowing another update to become accessible. Across a learned trajectory, the system constructs a state in which an initially improbable answer becomes probable. The final solution is selected because earlier updates created the conditions needed to select it.

Multiassociative search explains how a present context generates its next representation. Multiassociative learning explains how experience changes the network so that later contexts generate better representations. Thought traverses the learned probability landscape, while learning continually reshapes the landscape through which thought travels.

The central conclusion can therefore be stated plainly:

Intelligence is partly the learned ability to construct and preserve the context that will generate the right next thought.

References

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

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.

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.

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.

Fuster, J. M. (2009). Cortex and memory: Emergence of a new paradigm. Journal of Cognitive Neuroscience, 21(11), 2047–2072.

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

Hebb, D. O. (1949). The organization of behavior. Wiley.

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.

Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., et al. (2022). In-context learning and induction heads. arXiv:2209.11895.

O’Reilly, R. C., & Frank, M. J. (2006). Making working memory work: A computational model of learning in the prefrontal cortex and basal ganglia. Neural Computation, 18(2), 283–328.

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.

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323, 533–536.

Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599.

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.