Learning Context: A Unified Framework and Roadmap for Context-Aware AI in Education

Synthesis: Liu, Bradford, Hatchett, Diaz, Luzi, Wang, Basu Mallick, and Baraniuk (Rice University, OpenStax, SafeInsights, Adobe Research) introduce a unified Learning Context (LC) framework to move AI-based education from "context-blind mimicry" toward a principled, holistic understanding of the learner. This white paper provides a multidisciplinary roadmap for making teaching and learning systems context-aware by encoding cognitive, affective, and sociocultural factors over the short, medium, and long term. It proposes operationalizing LC theory into an interoperable computational data structure, leveraging the Model Context Protocol (MCP) to enable a wide range of AI tools to share and act on learner context.

Key Findings

• A unified Learning Context (LC) framework. The paper proposes encoding cognitive, affective, and sociocultural factors about the learner across short-, medium-, and long-term timescales — moving beyond single-feature or single-moment personalization.
• From context-blind to context-aware AI. The framework targets the transition of AI-based education from "context-blind mimicry" to a holistic, principled understanding of the learner.
• An interoperable computational data structure. A central proposal is operationalizing LC theory into a shared, interoperable data structure for learner context.
• Leveraging the Model Context Protocol (MCP). By building on MCP, a wide range of AI tools could share and act on learner context, enabling cross-tool context awareness.
• A multidisciplinary roadmap. The white paper lays out concrete steps across cognitive, affective, and sociocultural dimensions for making teaching and learning systems context-aware.

What this means for practice

• Designers. Build a holistic model of the learner rather than a narrow, decontextualized one: encode cognitive, affective, and sociocultural factors across short-, medium-, and long-term timescales instead of single features or single moments.
• Designers. Adopt the Model Context Protocol as the interoperability layer so that a range of adaptive and other AI tools can share and act on the same learner context, making context-aware personalization practical across education platforms rather than locked inside one product.
• Designers. Warm-start the system with slow-to-surface signals: dialogue recovery reached 91.4% for misconceptions but only 68.6% for conscientiousness and 60% for language proficiency, so capture those traits in the context data layer rather than waiting for a new learner to reveal them.
• Designers. Treat privacy and performance as a managed tradeoff: establish the upper performance bound unconstrained by privacy protections first, then quantify how much learning utility each privacy-preserving mechanism costs before deploying.
• Researchers. Control for raw domain mastery before attributing any gain to context awareness, using the roadmap's proposed baselines (Humanity's Last Exam, TutorBench, SocraticBench, LLF Bench) to verify tutoring moves, dialogic skill, and feedback-driven updating separately.

Limitations

• This is a white paper and roadmap, not a deployed system: its four vignettes are demonstrations and simulated experiments, so no learning gains from an operational context-aware tutor are reported.
• The quantitative evidence comes from simulated rather than real learners — Vignette 2 used a single synthetic student profile ("Maya") and Vignette 4 ran a pilot of 35 simulated interactions with the GPT-5.1 model on generated dialogues.
• Vignette 4's pilot exposes an observability gap in the framework's own data: misconception recovery reached 91.4% and anxiety 100%, but conscientiousness recovered at only 68.6% and language proficiency at 60%, so key intended context signals are not reliably observable from dialogue alone.
• The stated deployment thresholds (Cohen's d ≥ 0.20 on primary learning outcomes and ≥ 95% accuracy in classifying learner states against ground-truth measures) are proposed go/no-go criteria deferred to future efficacy trials; neither had been met or tested when the paper was written.

Connected Concepts

• AI in Education — the umbrella field
• Learner Modeling and Adaptive Instruction — the learner-modeling foundation
• Personalized Learning — the goal context-aware AI serves
• Adaptive Learning — a downstream application of learner context
• Human AI Collaboration — context-aware AI as a collaborative partner
• Learning Analytics — the data layer for learner context
• Cognitive Offloading — a risk context-aware design should mitigate
• Motivation — an affective factor encoded in LC
• Self-Regulated Learning — long-term learner context for regulation

Connected Articles

• A systematic mapping review at the intersection of artificial intelligence and self-regulated learning — AI and self-regulated learning mapping
• Fostering self-regulated learning through adaptive learning technology: A differentiated perspective on the role of feedback — Adaptive learning and feedback for SRL

Citation

Liu, N., Bradford, B., Hatchett, J., Diaz, G., Luzi, L., Wang, Z., Basu Mallick, D., & Baraniuk, R. (2026). Learning context: A unified framework and roadmap for context-aware AI in education.