Adaptive Support for Collaborative Learning: A Meta-Analysis

Synthesis: Gao, Fischer, Stadler and Chernikova pool 83 experimental and quasi-experimental studies of adaptive instructional support in collaborative learning, reported in 80 articles between 2004 and 2023, and find a medium positive effect (g = .73, 95% CI [.56, .90]). Gains are uneven: large for individual knowledge, medium for group products and processes, small for affect, and absent for cognitive load. The choices separating strong from weak interventions are consistent: cognitive sources, joint-activity targets, system control, and short durations do most of the work. Adaptivity helps when it is built into the task rather than layered on top of it.

Key Findings

1. Adaptive support beats static support, but the average hides a wide range. Across 83 studies the pooled effect was medium and clearly positive, with heterogeneity high enough that the average describes a family of interventions.
2. Adapt to what students know, not how they feel. Cognitive triggers such as prior knowledge, performance, or error patterns produced larger knowledge gains than social or affective triggers, the weakest source.
3. Aim the support at the group's joint work. Support directed at collaborative activities produced the most consistent effects, including on group products and social and affective outcomes; supporting both levels at once added nothing.
4. Let the system or the teacher decide when and how much. Effects were large under system and teacher control. Learner and shared control did not reach significance, which fits evidence that students under-use help.
5. Keep adaptive episodes short. Supports lasting an hour to a day outperformed both single-session supports and implementations spanning a month or more; early support arrives before groups settle, and long-term support competes with students' own regulation.
6. Stacking supports does not compound. Feedback plus scaffolding was the most consistent pairing, but combining cognitive and social-affective sources, or targeting individual and group activity together, yielded smaller social and affective effects.
7. Stronger students gained more, not less. Learners with higher prior knowledge showed a significantly larger effect than low-prior-knowledge learners, and both benefited; adaptivity appears to withdraw redundant help rather than produce an expertise reversal.

How the evidence was assembled

The search covered EBSCO, Web of Science, and Scopus through June 8, 2023, yielding 7,907 records after duplicate removal; full-text screening left 80 articles reporting 83 studies. The studies, published between 2004 and 2023, supplied 551 effect estimates from 6,974 participants (52% female). Effects were computed as Hedges' g, adjusted for pretest differences, and pooled with a random-effects model and robust variance estimation. Three outliers above g > 9 were removed; Egger's test found no asymmetry (t = 1.64, p = .100) and 34 of 37 p-curve results carried evidential value.

Where the gains land

The overall estimate was g = .73, 95% CI [.56, .90], p < .001. Baseline groups reached g = .88, fixed scaffolding .72, no scaffolding .66. Heterogeneity was substantial, Q(550) = 4357.39, p < .001, τ2 = .69, I2 = 93.30%.

Individual cognitive outcomes carried the largest effect (g = .93, 95% CI [.65, 1.21]). Group task performance (g = .62) and the three process measures sat in the medium range: task-related processes g = .50, social interaction quality g = .48, shared regulation g = .54. Affective outcomes were small (g = .28, 95% CI [.01, .54]). Metacognitive outcomes had a point estimate of 1.31 but an interval spanning zero [−.74, 3.36], and cognitive load showed no effect (g = −.06). Cognitive load was unchanged; gains showed up in what groups produced.

Which design choices matter

Cognitive sources produced g = .92 [.65, 1.2] on individual knowledge, against .53 [.22, .83] for social and affective sources; combining the two reached 1.09 [.32, 1.87] but added nothing elsewhere. Cognitive models (g = 1.14 [.56, 1.71]) and data science (g = 1.22 [.69, 1.74]) were the strongest mechanisms for knowledge; human-led adaptation was the only mechanism with a reliable effect on social and affective outcomes (g = .48 [.14, .82]), while combining data science with human judgment fell to .59 [.05, 1.13].

Feedback (g = .96) and feedback plus scaffolding (g = .78) both worked; scaffolding alone reached 1.33 but its interval ran from −1.54 to 4.2 on six effects, so it is not established.

Support aimed at collaborative activities produced g = .72 [.31, 1.14] on group products, .46 [.25, .67] on processes, and .37 [.11, .63] on social and affective outcomes. On knowledge, system control produced 1.05 [.65, 1.45], teacher control .86 [.46, 1.26], and learner control .31 [−.44, 1.05]. Short-term support of an hour to a day reached 1.06 [.48, 1.65] on knowledge and 1.05 [.65, 1.45] on group products; long-term support fell to .74 [.33, 1.14] and .43 [.18, .69]; brief support did not clear zero (.75 [−.25, 1.75]).

Who benefits, and where

Prior knowledge was the only learner characteristic that moderated the effect: learners who already knew the content gained more (g = .74, 95% CI [.42, 1.05]) than learners who did not (g = .63, 95% CI [.46, .8]). Education level and group size did not separate: K–12 (g = .70) and higher education (g = .77) were comparable, as were small (g = .69) and medium groups (g = .65).

Online (g = .77) and hybrid (g = .87) contexts showed large effects; face-to-face was small and nonsignificant (g = .29, 95% CI [−.05, .63]). Computer science (g = 1.13) far outpaced science (g = .42).

What this means for practice

• Instructors: plan adaptive support as a short episode tied to a specific collaborative task, not a standing course feature. Trigger it from what students know and what the group is doing, and do not assume students will ask for help.
• Instructional designers: give the system authority over when, how much, and what kind of support arrives, and aim it at the joint activity. If you add a second support, pair feedback with scaffolding.
• Administrators: start where the environment already produces usable data, which here means online and hybrid courses, and treat face-to-face deployment as a pilot, not a proven gain.

Limitations

• Moderator contrasts compare studies rather than randomize design features, so the patterns are associations that may be confounded by study choice.
• Social and affective outcomes were pooled to keep statistical power, which may hide different patterns for motivation and emotion.
• Emotional and motivational outcomes were mostly self-reported, with few physiological measures, which limits comparability.
• The evidence concentrates on feedback and scaffolding; difficulty progression, rehearsal scheduling, adaptive assessment, and cross-course progression were too rare to test.
• Only English-language studies were included, and group composition, teacher expertise, and task type were not reported consistently enough to model.

Connected Concepts

• Collaborative Learning: the outcome space the synthesis organizes into individual, group, and process dimensions
• Adaptive Learning: the design family, defined here as assessing prerequisites and processes to tailor support
• Scaffolding: one of only two adaptation types with enough studies to estimate
• Feedback: the other adaptation type, and half of the most consistent combination
• Prior Knowledge: the one learner characteristic that moderated effects, in the direction of avoiding an expertise reversal

Connected Articles

• Towards an adaptive AI scaffold for developing student collaborative problem solving: a design study that builds a process-dynamic scaffold of the kind this synthesis says remains rare
• Artificial intelligence agents in computer-supported collaborative learning: a systematic literature review: a review of AI agents in CSCL that reaches the same cognitive-gain, social-affective-uncertainty pattern
• Do intelligent tutoring systems benefit K-12 students? A meta-analysis and evaluation of heterogeneity of treatment effects in the U.S.: a meta-analysis of intelligent tutoring in K-12 that isolates the cognitive outcome this synthesis finds strongest
• The New Kid on the Tutoring Block: Definitions, Developments, and Future Directions for Hybrid Human-AI Tutoring: the hybrid human-AI architecture whose promise the combined-mechanism results did not confirm

Citation

Gao, L., Fischer, F., Stadler, M., & Chernikova, O. (2026). Adaptive support for collaborative learning: A meta-analysis. Review of Educational Research. Advance online publication. https://doi.org/10.3102/00346543261488039