1 Introduction

The current dilemmas facing Chinese L2 writing assessment include low efficiency in teacher evaluation, lack of standards for student self- and peer-assessment, and machine scoring's emphasis on form over content; introduces the new possibilities generative AI offers for writing feedback and its potential in facilitating learners' noticing of gaps between interlanguage and target language, while noting AI's limitations such as over-correction, excessive cognitive load, and repetitive, tedious feedback; proposes that teachers need to critically evaluate AI feedback and develop integration strategies, ultimately focusing on intermediate-to-advanced transitional stage Chinese learners to explore how teachers can combine AI feedback to select evaluation foci and cultivate learners' reflective abilities, thereby improving the efficiency of written language knowledge internalization.

2 Teacher-Student Collaborative Assessment and Selection of Evaluation Foci

Discusses the information overload problem caused by AI feedback and its impact on learners' uptake efficiency, proposing teacher-student collaborative assessment with "teacher-led, student-centered" approach as a solution, emphasizing "selective evaluation" to focus on key issues. The selection of evaluation foci should follow four principles: typicality, teachability, gradualness, and systematicity, with reference to Sun Shuguang's (2020) three-level system of vocabulary, grammar, and structure in English writing instruction, as well as Cui Xiliang's (2024) four categories of typical problems for Chinese L2 learners: linguistic expression, thinking patterns, register identification, and discourse construction. This study targets second-year Chinese language majors preparing for HSK Level 6, identifying 10 categories of evaluation foci based on four writing assignments at the beginning of the semester, divided into four levels: core clauses, complex sentences, compound sentences, and beyond-sentence level, respectively addressing vocabulary selection and sentence completion issues, predicate structures and condensed structures, logical relations and cohesion in complex sentences, as well as discourse coherence, register differences, and genre differences. The induction of evaluation foci includes two stages: AI-assisted revision and annotating revision rationales: using the "Wenxin Yiyan" tool to generate sentence-aligned revised text through specific prompts, then analyzing revision rationales sentence by sentence and annotating by level, ultimately forming samples for classroom explanation and practice.

3 Construction of the Evaluation Focus System

Constructs a four-level system for generative AI-assisted Chinese writing assessment, aiming to help students analyze the rationales behind AI revisions and establish a reflective framework. The core clause level focuses on two categories of evaluation foci: temporal expressions and subjectivity expressions. Temporal expressions address learners' "incomplete" errors due to lack of understanding of "sentence-completing elements," emphasizing the distinction between "realis sentences" and "irrealis sentences" and achieving predicate temporality, noun referentiality, and adjective gradability through explicit linguistic elements. Subjectivity expressions focus on confusion among modal verbs and absence of subjective linguistic forms, guiding students to distinguish three types of modality and analyze differences between objective expression and subjective judgment. The complex sentence level addresses the requirement for learners above Level 6 to "use complex sentence patterns," selecting four situations as evaluation foci: predicate elements serving as subjects/objects, predicate elements serving as modifiers, two events condensed into causative structures, and two events condensed into verb-complement structures, cultivating students' awareness of analyzing event relations and selecting appropriate complex sentence patterns. The compound sentence level focuses on expression of logical relations in complex sentences, particularly addressing the generalization of causal conjunctions in adverbial complex sentences, the disunity between intended and formalized contrastive relations, and the misuse of "并" (bìng) and semantic relation mismatches in coordinate complex sentences. The beyond-sentence level focuses on discourse cohesion and coherence, selecting two categories of evaluation foci: the discourse functions of "这" (zhè)/"那" (nà) and personal pronouns, and the extraction of topic sentences and transitional sentences. The former emphasizes the functions of demonstrative pronouns in summarizing preceding text and introducing following text, as well as appropriate collocation of personal pronouns with ellipsis; the latter utilizes AI's ability to supplement topic and transitional sentences to help students clarify discourse structure and master macro-structures and common expressions of different text types.

4 Concluding Remarks

Explores the transformation of teachers' roles under generative AI assistance, shifting from heavy correction to in-depth interaction, proposing the evaluation focus system as scaffolding instructional support, enhancing students' willingness to analyze AI revisions through gradual training, improving participation in teacher-student collaborative assessment and course satisfaction, and noting that future work will continue observing learners' acceptance of AI feedback and examining the sustainable improvement effects and transferability of this model on Chinese writing ability and language literacy.