A design framework for LLM-powered educational agents that prioritize durable learning over short-term task performance. Introduced by Khosravi et al. (2026), AI learning companions are defined as adaptive, pedagogically informed agents integrated into learning environments — distinct from both task-oriented LLMs and simple prompted tutors.

AI Learning Companions Framework

Definition

A design framework for LLM-powered educational agents that prioritize durable learning over short-term task performance. Introduced by Khosravi et al. (2026), AI learning companions are defined as adaptive, pedagogically informed agents integrated into learning environments — distinct from both task-oriented LLMs and simple prompted tutors.

The Learning-Performance Paradox

LLMs demonstrably improve task outputs (writing quality, code correctness, analysis speed), but this can create a paradox: students who produce better work with AI may learn less. This mirrors and extends the LLM Fallacy Misattribution phenomenon where users misattribute AI-assisted outputs to their own competence. The framework addresses the question posed by Transfer Of Learning: do AI-assisted gains persist when the tool is removed?

Three Foundations

1. Pedagogical Foundation — grounded in learning science; how students learn with AI, not just from it

2. Adaptive Foundation — AI learns about the student over time, connecting to LLM Student Modeling Memory and longitudinal personalization

3. Responsible Design Foundation — transparency, accountability, inclusivity, security; aligns with AI Tutor Safety Harms harm taxonomy

Case Studies

Validated across five diverse educational contexts, levels, and tool designs, revealing both promise and current limitations. The framework calls for a deliberate shift from task-optimized LLMs toward companions that foster durable understanding, metacognitive growth, and learner agency — connecting directly to Self Regulated Learning and Metacognition.

Connected Concepts

Self Regulated Learning

Metacognition

Connected Articles

LLM Fallacy Misattribution

Transfer Of Learning

LLM Student Modeling Memory

AI Tutor Safety Harms

Citation

Khosravi, H., Gasevic, D., Sadiq, S., Yan, L., Lodge, J., Tangen, J., Denny, P., & DiCerbo, K. (2026). Building AI Companions that Prioritise Learning over Performance