Late maturation of semantic control promotes conceptual development

Reviewer #1 (Public review):

Summary:

This study was motivated by the general claim that delayed development of cognitive control can be beneficial for learning, and investigated this claim in the specific domain of conceptual development. A comprehensive set of computational model simulations showed that delaying the onset of semantic control produces faster learning with only minimal effects on conceptual abstraction. The simulations also showed that control was most effective at intermediate levels between modality-specific "spokes" and the multimodal "hub". A meta-analysis of developmental data was consistent with the claim of delayed onset of semantic control: young children show substantially better semantic knowledge than the ability to constrain that knowledge to a specific task at hand.

Strengths:

The computational modelling is based on a very well-established model of semantic cognition, which means that the simulations allow exploring the specific issues under investigation here in the context of a model that accounts for a very large set of semantic cognition phenomena. The simulations are comprehensive - manipulating different parameters of the model provides important insights into how (and why) it works.

In addition to simulations exploring delayed maturation, there is an exploration of where semantic control is most effective, yielding the interesting result that control is most effective when it targets intermediate levels of semantic processing. To my knowledge, this is a novel finding and a concrete prediction for future testing.

The meta-analysis is designed in a very clever way that allows extracting evidence of semantic control from a large body of prior work. The results are quite clear and compelling in showing that semantic knowledge is acquired before children are able to use task demands to constrain the use of that knowledge.

Weaknesses:

Computational models of cognition inherently require simplification in order to focus on the mechanisms under investigation. However, it is also important to keep these simplifications in mind because they limit the generality of the inferences that can be made from the simulation results. Two aspects are important in this context:

(1) The multimodal structure was orthogonal to the surface similarity structure of the concepts to be learned. It is certainly true that multimodal structure does not perfectly mirror surface similarity, but closely related things tend to be perceptually similar. There are exceptions (whales, penguins, etc.), but they are *exceptional*, not typical. It may be that the somewhat extreme dissociation of multimodal and surface similarity structures creates demands that are not faced in natural conceptual development.

(2) Much of the benefit of delayed semantic control seems to be because the model is not penalised for activating task-irrelevant features. This blurs the distinction between being aware of a feature and making a response based on that feature. A full model that also includes a response layer could become a lot more complicated and more difficult to understand, so maybe there is an advantage to using a simpler architecture.

In addition, there is a bit of a misalignment between the model simulations and the meta-analysis. In the model, there are distinct modality-specific "spokes" and control is required in order to focus on modality/spoke in a task-appropriate way. The meta-analysis does not compare a task-defined selection of a modality; it compares the selection of taxonomic vs thematic relations, both of which are multimodal. One way to resolve this is to say that taxonomic and thematic relations are also represented in distinct sub-systems of semantic knowledge and semantic control is needed to select between them in a task-appropriate way.

This is particularly relevant to the inference at the bottom of p. 38: "taxonomic and thematic relationships ...[are]... both being encoded within the same system of representation", which seems in direct contradiction to the present results, or at least to the logic of combining these simulations with this meta-analysis. The simulations are based on semantic control being used to select/constrain the correct distinct sub-system (modality-specific spoke); the meta-analysis is based on semantic control being used to select/constrain the correct relationship type. If these two things are analogous in some way, then the relationship type has to be something like a distinct sub-system.

Reviewer #2 (Public review):

Summary:

This paper investigates the idea that the protracted maturation of the prefrontal cortex - often viewed as a developmental limitation - may actually confer advantages for conceptual learning in children. The authors focus on semantic control processes, which govern the context-sensitive application of conceptual knowledge, and are closely associated with late-developing regions of the prefrontal cortex.

Drawing on a computational model, the paper formally tests whether delayed maturation of semantic control promotes the acquisition of conceptual knowledge. The simulations demonstrate that when semantic control and anatomical connectivity mature later, conceptual learning is accelerated without compromising the integrity of the learned representations. Notably, the benefit is most apparent when control connections target intermediate layers in the computational model, suggesting a nuanced interplay between control processes and the underlying conceptual network.

To validate these computational insights in a human developmental context, the authors conduct a meta-analysis of the classic triadic matching task - a paradigm where participants decide which of two choices best matches a reference concept based on either taxonomic or thematic relations. Critically, when these relations conflict, semantic control is required to select the context-appropriate match. Results indicate that context-sensitive semantic control develops more slowly than basic conceptual knowledge, showing marked improvements between 3 and 6 years of age.

Overall, the paper argues that the delayed development of prefrontal cortex-based control processes allows for a period of less constrained learning, ultimately enhancing conceptual acquisition. The findings challenge the traditional view of late PFC maturation as solely disadvantageous and instead position it as an adaptive feature for building robust conceptual frameworks in early childhood.

Strengths:

(1) Novel Theoretical Contribution
The paper offers a compelling, counterintuitive argument that a developmental lag in the maturation of control processes might be beneficial for semantic learning. This stands in contrast to the conventional framing of late prefrontal cortex (PFC) development as purely disadvantageous (e.g., a "necessary but unfortunate" constraint).

(2) Well-Grounded Computational Approach
The authors propose a neural network model that is both theoretically driven (hub-and-spoke framework) and systematically tested under various conditions (different timelines for control onset, and different connectivity patterns). Their simulations replicate and extend previous findings about how insulating the multimodal hub from direct control inputs helps preserve abstract conceptual representations.

(3) Neuro-anatomical basis
The paper connects its computational claims to empirical neuroanatomy, particularly the lack of direct structural connectivity between ventral ATL (the "hub") and the PFC in humans. This lends biological plausibility to the argument that control signals likely reach the ATL via intermediate regions (e.g., posterior temporal cortex).

(4) Meta-Analysis of Triadic Match-to-Sample
The authors leverage decades of developmental data on conceptual matching tasks, reframing them in terms of semantic control vs. semantic representation. Their analysis nicely illustrates that children can identify semantic relationships (taxonomic or thematic) at age 2 if the task does not require them to select between conflicting semantic relations. In contrast, the ability to choose a task-relevant relation only emerges more robustly in 3-6 years. This developmental pattern aligns with the computational model's predictions.

Weaknesses:

The contribution of the paper might be considered rather specialist, and might not appeal to a broad public, which should be typical of a generalist journal. Moreover, the scope of the model is fairly narrow - its relatively small, controlled training environment raises questions about scalability to more naturalistic, high-dimensional data. Finally, the meta-analysis does not test directly the model predictions in terms of specific outcomes of the task, error patterns, or model fit, but only the developmental pattern which was an already observed phenomenon that in part motivated the hypothesis and the model itself.

Author response:

On the control of taxonomic versus thematic information. Both reviewers had questions about the relationship between the focus of the meta-analysis, the control of responses based on taxonomic versus thematic relationships, and the simulation. Both the model and the meta-analysis focus on the same mechanism, the controlled selection of task-appropriate features. In the case of the meta-analysis, this was the features and associations needed to identify the taxonomic or thematic relationships. As reviewer 1 notes, one possibility is that these kinds of structures are represented in distinct cortical regions. For instance, Mirman, Schwartz and colleagues have suggested that temporoparietal regions may preferentially support thematic knowledge while temporal regions may preferentially support taxonomic knowledge. Alternatively, they may be supported by different features instantiated within the same regions. However, whether taxonomic and thematic relationships require access to features in different regions or not, is not crucial to the conclusions of this paper. The simulations used here happen to select features based on their inclusion in a particular sensory modality, yet they could learn to select any combination of features. Indeed, prior simulations using the Jackson et al., (2021) model show that the functional impact on learning of “deep” conceptual representations (together with controlled behaviours) is the same regardless of whether the potentiated features are localised within one spoke or distributed across spokes. Thus, the key results regarding the acquisition of semantic knowledge before the maturation of control in the current work should hold regardless of whether knowledge of taxonomic and thematic relations is localised to different anatomical regions.

On model size and scalability. Both reviewers noted the relatively small size of the model and wondered about implications for ecological validity of the simulations and scalability to larger, noisier, and potentially more systematically structured training environments. We agree this is an important direction for future research, but one that faces two nontrivial challenges. First, reviewer 1 notes that, whereas our model environment employs orthogonal structures across spokes and for the cross-modal features, perceptual structure may be better-aligned with conceptual structure for real-world experience. While we appreciate the intuition, its validity depends to a key extent on how visual information about objects is encoded. Conceptual structure is certainly not apparent, for instance, in the distance between bitmap images of objects, nor the overlap of simple feature-extraction algorithms (such as edge detection or Fourier decomposition, etc). Even in this age of deep vision models, it remains unclear how the visual system extracts and discerns perceptual similarity from retinal input (see e.g. Mukherjee & Rogers, 2025). Most successful contemporary models train neural networks to assign visual images to semantic categories, suggesting that the visual features the model learns, and thus the perceptual similarities it represents, depend on learning to generate semantic information. Therefore, it is not clear whether the similarity that people perceive amongst instances of the same class is natively apparent in the bottom-up visual input, or whether it depends on semantic/cross-modal learning and representation. It should also be noted that within our training environment, there are features in each modality that are predictive of features in other modalities, as well as some that are only predictive of features within this modality. Thus, the full cross-modality conceptual structure is not orthogonal to the information available in each sensory domain, instead there is a relationship between surface and multimodal similarity in the dataset as in the real-world environment. In general, one virtue of the small-scale modelling endeavour in the current work is that we can be very explicit about the nature of the structure apparent within and across spokes.

The second non-trivial issue concerns the nature of the mechanisms that allow for context-sensitive responding in large-scale language/vision models such as GPT 4. Such models are trained on web-scale language and vision and provide a means of simulating controlled behaviour with realistic stimuli, so might seem to provide a means of assessing scalability of current neuro-cognitive models. Large language/vision models rely, however, on transformer architectures whose relationship to hypothesized mechanisms of control in the mind and brain is unclear. In transformers, context-sensitive responding depends upon “attention” mechanisms that are fully distributed and integrated throughout the entire system—there is no distinction between control, representation, and short-term memory in the architecture. As a consequence, it is very difficult to understand why a model behaves the way it does, or to relate patterns of behaviour to hypothesised mechanisms in the human mind/brain. Yet transformers are currently the only models capable of exhibiting context-sensitive patterns of responding based on both language and vision. Scaling up neuro-cognitive models will require developing alternative architectures that preserve the critical hypothesised distinctions between representation and control while retaining the ability of transformers to learn from large-scale ecologically realistic corpora of language and images. In the meantime, small-scale simulations like those reported here provide some critical insights into aspects of architecture and maturation that may aid in this endeavour.

On including a response layer. Reviewer 1 notes that our model does not separately simulate response-generation and the selective activation of relevant feature representations. We agree that there are interesting questions about how feature-potentiation and response-generation relate to one another, and that incorporating response selection in the current model would significantly complicate the analysis. The general idea that control potentiates/suppresses task-relevant feature representations in addition to simply promoting the correct response derives from classic work by Martin and others (e.g., Martin et al., 1995) showing that, for instance, regions involved in colour perception activate more strongly in tasks requiring retrieval of colour than tasks involving retrieval of action and vice versa—results consistent with the model training/testing procedure in the current work. In general, it may be counterproductive to become aware of aspects of a concept that would be irrelevant, or even actively unhelpful in making a response, suggesting guided activation is a necessary precursor to response selection (Botvinick & Cohen, 2014). Here, we focus on this important feature potentiation step.

On the novelty of the meta-analysis. Reviewer 2 suggests the results of the meta-analysis were already known and provided motivation for the simulation. However, an important contribution of the current work is the observation that, in fact, there is little prior work on the development of semantic control. The widely known developmental delay in domain-general executive control, which did indeed motivate the study, is exclusively based on tasks requiring very different forms of executive control. Many of these involve no meaningful stimuli or require the child to completely inhibit a practiced response and generate an opposite or completely arbitrary responses, instead of requiring the child to use context to select among two or more meaningful behaviours that are equally valid in different contexts (see the introduction to Part 2). This observation, coupled with recent evidence that semantic control relies on dedicated and partially non-overlapping neural systems to executive function, illustrates the utility of the current meta-analysis: delineating the developmental trajectory of semantic control requires a task in which control is applied to the context-appropriate retrieval and manipulation of semantic knowledge, such as the triadic matching task. Moreover, the results show that semantic control, while arising later than semantic representation, nevertheless begins to mature earlier (around 2.5 years) than typical estimations of domain-general executive control (around 4). Thus, the meta-analysis contributes to our understanding of cognitive development while also testing a key prediction of the model.