Learners valued adaptive scaffolds when suggestions fit their writing strategy

Li and colleagues investigate a subtle problem in adaptive learning: a system can change its support, yet a learner may not experience that support as adaptive. Their ICQE 2023 paper uses Epistemic Network Analysis to model connections among reported writing strategies, perceptions of adaptivity, and essay performance after a two-hour source-based academic-writing task.

The broader task involved 253 university learners in a 16-week academic-writing course for non-native English speakers. Ninety-four learners received adaptive scaffolding, and 22 of them volunteered for follow-up interviews. The abstract calls the interviewees undergraduates, while the Methods section describes the course participants as graduate learners, so the safest public description is simply university learners. During the task, students read materials on artificial intelligence, differentiation, and scaffolding, then wrote a 300 to 400 word essay. Five prompts appeared at planned times. A rule-based mechanism withheld a suggestion if the relevant learner behavior had already been detected.

Mandarin interviews were transcribed, divided by conversational turn, and consensus-coded by two native-Mandarin researchers. The analysis used four strategy codes plus two perception codes: perceived as adaptive and not perceived as adaptive. ENA treated each participant as a unit and the whole interview as a stanza. That choice allowed strategy and perception statements made far apart in the interview to connect, but it also removed temporal locality. The overall network used an optimized unit-circle layout; the performance comparison used means rotation, network subtraction, and a Mann-Whitney U test.

The strongest connection to perceived adaptivity was a read-first-then-write strategy. The strongest connection to non-adaptivity was writing while reading and taking notes. High and low performers differed on the first ENA dimension in the reported test, with U equal to 98, p below .01, and r equal to .63. The second dimension did not differ. Learners who reported reading first, then writing, and who perceived the support as adaptive tended to perform better. However, some learners perceived the scaffolds as adaptive and still obtained lower scores.

The result is relational, not causal. The ENA corpus contains only 22 self-selected interviewees from the adaptive condition, strategies were self-reported rather than reconstructed from trace data, and the whole-interview stanza can connect ideas separated by many minutes. The paper therefore does not show that one strategy causes higher performance or that perceived adaptivity makes a scaffold effective. It shows how ENA can expose alignment and mismatch among learner strategy, system support, and outcome, generating sharper questions for larger, multimodal studies.