Binz et al.'s seminal paper on “Meta-Learned Models of Cognition” offers a transformative view on cognitive modeling, shifting the traditional paradigm toward a more dynamic and experience-based approach. The authors convincingly argue for the superiority of meta-learned models in acquiring inductive biases from experience, as opposed to the rigid, hand-designed structures of traditional models like cognitive architectures and Bayesian models of cognition. This shift represents not only a theoretical advancement but also a practical one, providing a more realistic and adaptable framework for understanding cognitive processes.

Crucially, the paper's synthesis of meta-learning with rational analysis presents an exciting pathway for constructing Bayes-optimal learning algorithms. This approach resonates strongly with integrative learning theories that we have been working on, suggesting a shared trajectory toward developing learning models that can adapt and thrive amidst complexity. Integrative learning refers to the cognitive process of actively integrating learning materials under the influence of metacognition, resulting in an efficient and profound understanding and mastery of knowledge (Lian, Reference Lian2018). It represents a psychological learning process where metacognition and cognition are highly unified (Yin, Wu, & Lian, Reference Yin, Wu and Lian2020, Reference Yin, Wu and Lian2023). This learning process model encompasses three modes: holistic integration of parts, top-down reasoning, and generalization for in-depth analysis (Rong Lian, personal communication, March 2020).

Firstly, “holistic integration of parts” involves learners first grasping the overall concept of the subject, establishing a comprehensive initial understanding. This is followed by an exploration of specific parts of the material, with each part being connected back and integrated into this broader understanding, thus reinforcing and enriching the overall comprehension. This “whole-part-whole” learning process has been shown to play a positive role not only in animal learning (Yin et al., Reference Yin, Wu and Lian2020, Reference Yin, Wu and Lian2023) but also in human learning processes. In studies with university students learning online network knowledge, it was found that, compared to a non-integrative learning group, the integrative learning group better synthesized and processed fragmented online knowledge, resulting in superior learning performance (Huang, Reference Huang2021). This indicates that individuals, during the learning process, activate metacognition which combines prior knowledge and experience to adjust and optimize new knowledge within working memory, thereby enhancing online learning effectiveness (Mayer, Reference Mayer1997).

Secondly, “top-down reasoning” emphasizes beginning with more broad and generalized high-level concepts and systematically mastering more specific lower-level knowledge points. By understanding and applying higher-level knowledge, they deduce and explain lower-level knowledge, thereby forming a clear, logically structured knowledge framework. Lan (Reference Lan2022) explored the impact of this approach on learning outcomes using a study-recognition paradigm. The results showed that compared to bottom-up learning, top-down reasoning enabled learners to use attention resources more reasonably and effectively, facilitating the relational processing and integration of specific items. Additionally, cognitive semantic processing was smoother, and the integration difficulty between semantics was reduced, significantly enhancing memory effects. Event-related potential (ERP) technology was used to explore the underlying neural mechanisms, revealing that the top-down reasoning group had larger N1 amplitudes and significantly smaller N400 than the bottom-up learning group when learning specific examples. This indicates that top-down reasoning learners more effectively utilized higher-level knowledge, focusing attention resources on more organized processing of specific examples. From the perspective of semantic priming effects, once semantic concepts in memory are activated, their activation can spread to related nodes, increasing their activation level (Meyer & Schvaneveldt, Reference Meyer and Schvaneveldt1971), making the learning of related target stimuli easier (Bueno & Frenck-Mestre, Reference Bueno and Frenck-Mestre2002).

Lastly, “generalization for in-depth analysis” starts with learners forming a general representation of the subject, laying the groundwork for the overall framework. Learners then delve into detailed components, engaging in thorough analysis and research. This in-depth study not only deepens understanding of each part but also enhances and refines the initial general framework, culminating in a multifaceted and detailed overall cognition. Chen (Reference Chen2023) explored the impact of this learning mode on university students’ reading of texts of varying difficulty. Results showed that the “generalization and in-depth analysis” group had significantly higher understanding and memory scores under different text difficulty conditions compared to the control group, and also had a lower rate of knowledge forgetting. Reading is a process involving simultaneous extraction and construction of meaning (García Madruga et al., Reference García Madruga, Elosúa, Gil, Gómez Veiga, Vila, Orjales and Duque2013). In reading, learners must use complex meaning construction processing to form a complete representation, where cognitive control plays a key role in focusing and switching attention, activating and updating representations (Wu, Tian, Chen, Chen, & Wang, Reference Wu, Tian, Chen, Chen and Wang2021). The “generalization for in-depth analysis” mode helps learners construct complete representations more quickly, and through in-depth analysis and summarization of new knowledge, continuously adjust and optimize their meaning representations, thereby promoting more refined processing and encoding of information. This process helps learners more effectively retrieve and remember encoded information.

The three modes of integrative learning—holistic integration of parts, top-down reasoning, and generalization with in-depth analysis—closely align with the computational processes of meta-learned models of cognition. This alignment is pivotal, as the initial encounter with a subject in a holistic, higher-level, or generalized manner is essential for setting effective starting metaparameters that guide the learning process. Such an encounter provides a foundational understanding from which learners can refine their perceptions and strategies in a targeted manner. Within this adaptive framework, learners then engage in a cyclical process of interaction, analysis, and metacognitive adjustment, fine-tuning their approach based on this foundational overview and their evolving comprehension of the subject matter. This methodology not only embodies the adaptability characteristic of meta-learning but also supports real-time adjustments during the learning process. As a result, it leads to learning outcomes that are both precisely tailored to the learner's needs and highly effective. Consequently, emphasizing the value of an initial holistic overview reinforces the importance of integrative learning within the meta-learning paradigm. It empowers learners to dynamically adjust their information processing strategies from the outset, significantly enhancing and adapting their learning experiences to achieve optimal outcomes (Rabinowitz, Reference Rabinowitz2019).

The implications of this area of research are vast, offering new directions for educational practices, artificial intelligence, and cognitive neuroscience. In education, these insights could lead to more personalized and effective learning strategies, tailored to individual (meta-)cognitive patterns. For AI, integrating these models could result in more adaptive and intuitive systems, better mimicking human learning processes. In cognitive neuroscience, this research offers potential for deeper understanding of brain-based learning mechanisms. Altogether, this represents a significant stride in our comprehension and application of cognitive and learning sciences, opening new avenues for exploration and innovation.