Revisiting pedagogical agents in the era of large language models: from social agency to social interdependence

Abstract

This presentation revisits pedagogical agents in the era of large language models. Social Agency Theory (SAT) long explained engagement through social cues such as face, voice, gesture, and politeness. Fluent LLM agents change the design question: learners must question, evaluate, and influence the next step. We therefore propose Social Interdependence Theory (SIT) as a more suitable lens—shifting attention from “Does it feel social?” to “How is reasoning organised across the partnership?”

Key Topics Covered

• Theoretical Gap in SAT: Why social cues alone are insufficient when collaboration, not appearance, drives learning
• What LLMs Change: Semantic generation, contextual memory, and adaptive scaffolding beyond scripted agents
• Social Interdependence Theory: Positive interdependence, promotive interaction, and individual accountability as design heuristics
• Design Implications: Productive friction, shared artefacts, and co-regulation that keep learners active
• Evaluation and Challenges: Measuring joint activity, verification burden, memory/profiling, and the illusion of collaboration

Research Methodology

This talk synthesises pedagogical-agent research and SIT into a design-oriented theoretical framework. Drawing on the paper of the same title, we treat SIT as a heuristic for structuring human–LLM learning: the agent contributes breadth and feedback; the learner contributes context, judgement, and verification. Design moves and evaluation criteria are derived from this pivot rather than from cue-based engagement alone.

Key Findings

• From cues to shared responsibility: Fluent answer delivery can look collaborative while still producing passive offloading
• SAT → SIT pivot: Changes the agent’s role, the learner’s role, and the mechanism of learning
• Three design moves: Make progress contingent on learner reasoning; centre shared artefacts rather than chat; co-regulate planning, monitoring, and reflection
• Partnership evaluation: Look for substantive reasoning, negotiation turns, revision, transfer, and independent continuation

Implications for Education

Designing LLM-powered pedagogical agents for blended learning should prioritise reciprocal influence over social polish. Productive friction, visible co-construction, and co-regulation help protect learner agency while still leveraging model breadth. Success is a structured partnership that leaves the learner more capable.

Future Directions

Future work should implement these patterns with guardrails—signalling uncertainty, making memory consent transparent, and testing whether learner input actually changes the trajectory—while evaluating process evidence through discourse analysis, interaction coding, process tracing, and logs.

Conclusion

Revisiting pedagogical agents in the LLM era requires moving from social agency to social interdependence. The central design question is no longer whether an agent feels social, but how reasoning and responsibility are organised across the human–AI partnership.