Balancing Teacher and Student Agency: Co-Orchestration Tool Design Supporting Real-Time Dynamic Pairing

Synthesis. Yang et al. (2026) use participatory speed dating with 17 teachers and 13 students to map how control and Learner Agency should be distributed across the three stages of dynamic pairing — before, during, and after pairing students between individual and Collaborative Learning modes. They frame findings within a hybrid-control design space and recommend structured teacher guidance early, with progressively increasing student autonomy as activities unfold. Neither teachers nor students wanted full control: both groups favored shared decision-making, converging on a "ready to collaborate" mechanism for timing and on teacher-led pairing with limited student role/partner input. The result is an actionable principle for human–AI co-orchestration tools in K-12 classrooms.

Key Findings

1. Both stakeholders rejected extreme control allocations: full teacher control risks eroding Student Engagement and Self-Regulated Learning, while full student control risks chaotic classroom dynamics, unproductive pairings, and added teacher workload — validating a Human AI Collaboration middle ground.
2. For collaboration timing, both teachers and students ranked "When students are ready" (a "Ready to collaborate" button) highest, converging on a shared-control mechanism that preserves teacher oversight while respecting student readiness and Self-Directed Learning.
3. For partner/role assignment, students preferred structured teacher guidance ("Teacher decides" ranked highest), while teachers favored a balanced "Teacher assigns, student chooses role" — both groups converged on shared decision-making that pairs teacher authority with limited student input.
4. For collaboration content, teachers favored "Teacher proposes, students decide" while students favored "Students propose, teacher decides"; both groups ranked full "Student choice" least preferred, citing immaturity and the risk of avoiding productive challenge.
5. For productive collaboration, teachers favored automated supervision ("System detects keywords") while students favored "Get help from NPC" automated assistance — both embraced Human-in-the-Loop support to monitor collaboration quality and reduce instructor burden.
6. Teacher preferences were consistently more unanimous than student preferences (higher Kendall's W across challenges), reflecting a more stable conception of Teaching and classroom management goals.

The Tension in Real-Time Dynamic Pairing

The study addresses a core tension in AI-augmented classrooms: how to balance teacher orchestration with student Learner Agency during dynamic transitions between individual and Collaborative Learning work. In the study context, individual learning means solving mathematics problems inside an Intelligent Tutoring system, while collaborative learning takes the form of peer tutoring within the system, where one student assumes the Solver role and the other acts as Tutor by providing hints and corrective Feedback. Real-time Learning Analytics — tracking mastery via Bayesian Knowledge Tracing and metacognitive states — feed a teacher-facing orchestration tool that suggests when to pair, unpair, and reassign partners.

Prior research establishes the pedagogical value of combining individual and collaborative modes, with students making fewer errors and requesting fewer hints when instruction integrates both. Yet existing orchestration frameworks focus on teacher–AI or system-level distributions of control, leaving a gap in understanding how Stakeholders perceive hybrid control across the different stages of a co-orchestration tool. The authors frame this through the theoretical hybrid-control lens of Eshel and Kohavi, which treats teacher and student control not as a zero-sum tradeoff but as two independent dimensions — student control and opportunities for Self-Directed Learning — that both groups can exercise simultaneously. This connects Scaffolding theory to social-organizational support, extending it from individual learning support to classroom-level coordination.

Study Approach: Participatory Speed Dating

The researchers used participatory speed dating (PSD), a design method that combines storyboarding with rapid, structured evaluation to gather feedback from multiple Stakeholders early in the ideation phase. Across two phases — idea co-generation with 5 participants and idea evaluation with 25 — 74 design ideas were generated and narrowed to 34 representative solutions spanning varied levels of hybrid control. A mixed-method analysis triangulated quantitative ranking data (means, standard deviations, and Kendall's W agreement coefficients) with qualitative reasoning from affinity diagramming of participants' verbalized justifications.

The design process was organized around the three stages of the pairing process and seven "How Might We" challenges: before pairing (preparing for collaboration, choosing the right start time), during pairing (assigning roles and partners, selecting content, ensuring productivity), and after pairing (smooth transitions, evaluating collaboration to inform future pairings). This three-stage design space offers a structured lens for evaluating other orchestration tools, complementing The Missing Evaluation Axis: What 10,000 Student Submissions Reveal About AI Tutor Effectiveness's focus on individual tutor performance with classroom-level coordination metrics.

Preferences by Stage

Before pairing. Students valued quizzes, personality tests, and direct ways to share preferences with teachers to help the system understand their strengths and improve pairing — prioritizing a "good partner" over knowing the rules. Teachers, by contrast, preferred data-driven methods such as pairing rules and academic quizzes, distrusting personality tests as scientifically weak ("more noise than useful information"). Teachers leaned on experience and classroom management skills for factors like personality compatibility and social-emotional needs rather than standardized assessments.

During pairing. For timing, both groups favored "When students are ready," with students appreciating autonomy at their own pace and teachers valuing a signal they could act on. For partners and roles, students supported teacher-led control (citing teachers' better knowledge of skills and history), while teachers favored giving students some role choice to boost Motivation and engagement. For content, both groups converged on hybrid options — "Teacher proposes, students decide" (teachers) and "Students propose, teacher decides" (students) — and both ranked full student choice lowest out of concern that students might avoid challenge or lack maturity. For productivity, automated supports were favored by both, differing on whether assistance should supervise (teachers) or assist (students).

After pairing. The two challenges in this stage address ensuring smooth transitions back to individual learning and evaluating collaboration outcomes to improve future pairings — underscoring that orchestration is an ongoing, iterative loop rather than a one-time event.

What this means for practice

• Instructors. Phase the handover of control instead of fixing it once: keep structured teacher guidance through preparation and initiation, then cede decision-making to students as the activity unfolds and learners demonstrate readiness and self-regulation.
• Instructors. Gate the shift to collaboration on a readiness signal rather than the schedule — both teachers and students ranked "When students are ready" highest of the timing options, a mechanism that preserves teacher oversight while respecting student pacing.
• Instructors. Assign partners yourself but leave role choice open: students ranked "Teacher decides" highest for partner assignment, while teachers favored "Teacher assigns, student chooses role," giving students input where it costs the pairing quality nothing.
• Instructors. Keep content selection hybrid — propose the material and let students decide, or take their proposal and decide yourself — since both groups ranked full student choice lowest, citing immaturity and the risk of avoiding productive challenge.
• Instructors. Automate supervision of collaboration quality so pairing does not add to your workload: teachers favored system detection of keywords and students favored an in-system help option for getting unstuck.

Limitations

• The sample is small and self-selected: 30 participants (17 teachers and 13 students) recruited through network connections, snowballing and social media, with Phase 1 running 5 co-generation sessions and Phase 2 running 25 evaluation sessions.
• The study captures stated preferences rather than classroom agency in action: participants ranked 34 storyboard design ideas, 4 to 7 per challenge, in one-hour Zoom sessions and explained their reasoning, so no pairing decision was observed or measured in a live classroom.
• The design space is knowingly incomplete: the authors state that certain regions, such as the challenge with full student control, remain unmapped.
• The context is narrow — K-12 mathematics, pairing students inside an intelligent tutoring system that tracks mastery with Bayesian knowledge tracing — so the rankings may not transfer to other subjects, ages, or tool designs.

Citation

Kexin Bella Yang, Menghan Liu, Liyi Xu, Nikol Rummel, Vincent Aleven (2026). Balancing Teacher and Student Agency: Co-Orchestration Tool Design Supporting Real-Time Dynamic Pairing. Accepted at CSCW 2026, to appear in PACM HCI.