A Response to the Commentaries on Riches (2026)

Abstract

I welcome the views of the commentators and am glad that they are likewise enthusiastic about neural network accounts. Nonetheless, I also acknowledge their criticisms, including a lack of attention to processing accounts.

I would like to thank the commentators for their interesting and insightful comments, and I am glad that we are in broad agreement about the potential of neural network models to account for developmental language disorder (DLD). Jones (2026) focused on some gaps in the exemplar theory, for example, the relative underspecification of the process whereby language becomes excessively exemplar-based and whether it is possible to extend the framework to other linguistic phenomena, for example, nonword repetition. Owen Van Horne and Qi (2026) extend the approach in the review article, discussing a wide range of studies based on neural network models that implicate cognitive processes including inhibition and working memory. The comments serve as a reminder of the possible role of alternative mechanisms in DLD (e.g., working memory and inhibition), and also the rapid temporal and acoustic nature of spoken language processing, neither of which I discussed in the initial proposal.

I was particularly struck by Owen Van Horne and Qi’s proposal to manipulate the predictability of the passive by incorporating atypical verbs, for example, atelic predicates. This demonstrates that, with sufficient imagination, one can find numerous ways to manipulate the predictability of the input stimulus, and potentially promote learning, whether or not one agrees with the term “exemplar-busting.”

One theme common across both commentaries was the need to provide a more fleshed-out account of how exemplars are used during language processing. Discussing nonword repetition, Jones (2026) questions whether there is a “problem in the process that operates over existing exemplars [rather] than in the exemplars themselves.” Owen Van Horne and Qi (2026) argue that atypical verbs “serve to illustrate that the past tense marker can be decomposed from the lexical stem.” Their word “illustrate” is suggestive of an explicit metalinguistic process, and this differs slightly from my interpretation of their study, which focuses on the potential of atypical examples to disrupt implicit representations.

Though, in my proposal, I did not address the literature on language processing, I believe that a more competence-based account, whereby individual differences in a fundamental neural mechanism (e.g., myelination or the calculation of precision, or “learning potential”) impact on linguistic representations, has a better chance of explaining the tough-but-fragile hypothesis. Processing accounts are problematic, from this perspective, because there are often compensatory mechanisms. For example, Christiansen and Chater (2016) argue that given the fleeting nature of working memory across all individuals (irrespective of whether they have a language difficulty), language users need to exploit representations in long-term memory early and often via a chunk and parse mechanism. According to this account, working memory difficulties ought not to impact greatly on linguistic processing, at least at the sentence level. By contrast, difficulties with basic neural mechanisms affecting how representations are formed can account well for fragility, because it is hard to envisage how these difficulties may be rectified via a compensatory mechanism.

That said, given the numerous processing accounts of DLD (Ebert & Pham, 2019; Tomas & Vissers, 2019), and the substantial evidence supporting them, it is important to ask to what extent an exemplar-based account of DLD can explain these data. Moreover, as Jones (2026) argues, it is important to explore whether exemplar-related cognitive processes may be at fault. There is clearly substantial work required to truly evaluate the exemplar-based hypothesis of DLD.