Unpacking Interaction Profiles and Strategies in Human-AI Collaborative Problem Solving: A Cognitive Distribution and Regulation Perspective

Synthesis: This empirical study by Hao, Liu, Fan, Long, Yu, Chen & Zhang (Tsinghua University) adopts an integrated distributed cognition and co-AI Regulation in Education perspective to examine how college students collaborate with AI agents on complex Problem Solving tasks. Through dialogue coding and cluster analysis of 173 students, it identifies three collaborative problem-solving modes — Delegated Reasoning (DR), Concerted Interpretation (CI), and Delegated Elaboration (DE) — and shows that the DR mode achieves the highest task performance but the lowest regulatory engagement, while the CI mode engages learners most deeply in self-regulation. The work exposes a core tension between the efficiency of the distributed human-AI system and the depth of learners' cognitive and regulatory involvement.

From distributed cognition to human-AI collaboration

The paper extends distributed-cognition theory (Hutchins) beyond human teams into the human-AI context. Rather than treating students' use of generative AI as a monolithic behavior, it asks how cognitive work is distributed and regulated between human and AI collaborators turn-by-turn. This connects directly to the knowledge base's core theme of Human AI Collaboration and to Collaborative Learning research on AI as a partner or mediator. Because generative AI is highly reactive to user input, the study argues, the quality of collaboration depends on how students allocate and offload cognition and whether they maintain metacognitive oversight of the process.

Three modes of human-AI collaboration

Using an agglomerative hierarchical clustering of the proportion of SEDA-adapted utterance codes (Cohen's κ = 0.81), the authors derived three profiles:

• Delegated Reasoning (DR) — students outsource the reasoning process to the AI (highest Reasoning Invitation). The AI does the heavy cognitive lifting; the human verifies the final product.
• Concerted Interpretation (CI) — students expand on ideas and construct arguments themselves (highest Elaboration, Reasoning, Agreement), treating the AI as a dialogic partner rather than an answer source.
• Delegated Elaboration (DE) — students prompt the AI to elaborate, explain, or give examples (highest Elaboration Invitation and Team Dynamics), an exploratory middle ground.

Epistemic Network Analysis showed distinct structural logics: in DR and DE, Metacognition co-occurred with invitation-related actions (delegation and offloading), whereas in CI, metacognition was tied to human-initiated reasoning and elaboration — a difference the authors interpret as the difference between the student as "dispatcher" and the student as co-constructor.

Efficiency versus regulatory engagement

The results reveal a striking trade-off. The DR group achieved the highest task performance (significantly outperforming CI) and showed the highest semantic similarity between human and AI discourse — an interaction profile that aligns with the classical distributed-cognition ideal of efficient, low-loss information transfer. Yet DR also reported the lowest self-regulation, while the CI group reported significantly greater use of self-regulation strategies. No group differences emerged for co-regulation or socially-shared regulation, suggesting that current generative AI acts as a reactive responder rather than a proactive regulatory partner, leaving the regulatory burden on the individual learner.

The authors caution that efficient semantic coupling does not equal deep learning. Over-delegating higher-order reasoning, explanation, and integration to AI may trap learners in a "cognitive comfort zone" and incur a cognitive debt — obtaining a correct product at the cost of the self-constructive process of understanding. This has direct implications for Learning Analytics (where semantic cohesion is often treated as a marker of good collaboration) and for the design of AI-empowered tools in higher education.

Design implications

Drawing on the tension between system efficiency and regulatory engagement, the authors propose four design principles for educational AI: prioritize dialogic tension over seamless efficiency (e.g., strategic decoupling, counterexamples); enable proactive co-regulation (AI that monitors for premature consensus and prompts reflection); embed metacognitive diagnostics and Scaffolding to keep the "mind-in-the-loop"; and cultivate a synergistic co-constructive ecology where AI manages an inner loop of questioning while instructors steer an outer loop of value and strategy.

Key Findings

• Cluster analysis of 173 college students' human-AI dialogues identified three collaborative problem-solving modes: Delegated Reasoning (DR, n=78), Concerted Interpretation (CI, n=55), and Delegated Elaboration (DE, n=40).
• The DR group achieved the highest task performance, significantly outperforming the CI group (H = 9.437, p = .009).
• Semantic similarity between human and AI discourse was highest in DR; CI was significantly lower (b = −0.086), with no significant DE-vs-DR difference.
• The CI group reported significantly greater self-regulation than DR (H = 7.06, p = .029); no differences emerged for co-regulation or socially-shared regulation, as generative AI acted reactively rather than as a proactive regulatory partner.
• Epistemic Network Analysis revealed that DR/DE deploy delegation/offloading logics (metacognition tied to invitations), while CI reflects dialogic co-construction (metacognition tied to human reasoning and elaboration).
• The study surfaces a tension between system efficiency and regulatory depth — efficient semantic coupling does not equal deep learning, and over-delegation may incur "cognitive debt."

What this means for practice

• Instructional designers. Build in dialogic tension rather than frictionless efficiency — strategic decoupling, counterexamples and pauses for reflection — because the Delegated Reasoning profile (n = 78) produced the highest task performance (H = 9.437, p = .009) but the lowest self-regulation.
• Instructors. Require students to restate or critique AI reasoning in their own words so Metacognition stays with the learner; the AI behaved as a reactive responder, leaving co-regulation and socially-shared regulation unchanged across all three profiles.
• Researchers. Instrument for regulatory depth as well as task accuracy, since semantic similarity between human and AI discourse was highest in Delegated Reasoning yet did not index deeper learning.

Limitations

• The analysis is cross-sectional and single-session: 213 students recruited from universities in Beijing completed one 45-minute human-AI problem-solving task, with 173 retained after screening for missing dialogue records or incomplete questionnaires.
• Self-regulation, co-regulation and socially-shared regulation were measured by post-task self-report questionnaire rather than behavioral or physiological trace data — the gap the authors' own call for multimodal fusion (eye-tracking, affective computing) concedes.
• The three profiles were derived by agglomerative clustering of dialogue codes (Cohen's κ = 0.81); clustering is descriptive, so the design cannot show that a profile causes the performance or regulation differences.
• Generalizability is bounded by the task — an AGI-themed argumentation problem on the MAIC platform, done by volunteers compensated with 200 RMB.

Citation

Hao, Z., Liu, X., Fan, J., Long, Y., Yu, J., Chen, W., & Zhang, Y. (2026). Unpacking Interaction Profiles and Strategies in Human-AI Collaborative Problem Solving: A Cognitive Distribution and Regulation Perspective.