Abstract
Understanding the cognitive architectures that enable learners to process feedback and self-regulate is fundamental to fostering intelligent second language (L2) learning, particularly as AI-based pedagogies become prevalent. Direct written corrective feedback, while common in beginner L2 writing, often triggers only surface-level processing. This study investigates whether written languaging (WL), a metacognitive activity where learners explain language problems in writing, can deepen feedback processing, thereby activating a more robust cognitive architecture for error correction. Using a within-subjects crossover design with 15 beginner Chinese-as-a-second-language (CSL) learners from a UK secondary school, we compared the immediate and transfer effects of direct feedback with WL versus direct feedback only. Results showed that the WL condition significantly reduced errors per 100 characters (Z = −2.556, p = .011, r = 0.66), with 86.7% of participants showing improvement, indicating enhanced cognitive regulation at the local level. However, the effect was hierarchical, with no significant impact on clause-level accuracy. General Certificate of Secondary Education (GCSE) writing scores improved significantly from pre-to-post-test (Z = −3.342, p = .001, r = 0.89), demonstrating transfer to subsequent performance. Qualitative analyses revealed that learners’ attention focused predominantly on characters (48.4%) and vocabulary (29.8%), with engagement moderated by proficiency and motivation. This study provides empirical evidence for a cognitive architecture of feedback processing, wherein WL functions as a metacognitive amplifier. This architecture is operationalised as a three-stage processing sequence noticing, hypothesis-testing, and metalinguistic reflection, with writing accuracy and WL texts serving as observable proxies for the underlying cognitive processes. Specifically, our findings reveal that WL’s facilitative effects are hierarchical, strongest at the level of local error reduction and not yet extending to clause-level syntactic accuracy, and are moderated by individual differences in proficiency and motivation. These empirical insights offer a cognitive-psychological foundation that could inform the future design of adaptive feedback systems; however, as this study did not involve any AI system, these implications are theoretical and await empirical validation in AI-mediated learning environments.