TL;DR
AdaPT is an LLM-powered multi-agent system designed to help teachers adapt existing high-quality lesson plans for under-resourced or diverse classrooms. By shifting the AI paradigm from "generation from scratch" to "context-aware transformation," AdaPT aligns with actual teaching workflows, significantly reducing preparation workload while promoting educational equity.
What Problem Does This Paper Address?
Educational inequality often means that high-quality instructional materials designed for well-resourced urban schools fail to meet the needs of students in rural or lower-performing classrooms. To bridge this gap, teachers must practice differentiated instruction—tailoring lessons to local contexts and varying student readiness.
However, doing this manually creates a massive workload. While generative AI tools show promise in education, existing systems primarily focus on generating lesson plans from scratch. In reality, teachers rarely do this; they prefer to adapt and refine existing, trusted "standard" plans. Furthermore, teachers often lack structured, actionable student profile data, making it difficult to translate student differences into targeted lesson modifications.
Methods
The authors introduce AdaPT (Adaptive Lesson Plan Transformer), a human-AI collaborative system that reframes lesson preparation into a structured transformation process.
Multi-Agent Backend
AdaPT utilizes a backend powered by two specialized large language model (LLM) agents:
- LessonPlan Agent: Responsible for parsing uploaded lesson plans into a standardized schema (Curriculum Standards, Learning Objectives, Learning Profile, and Learning Tasks). It uses Retrieval-Augmented Generation (RAG) mapped against a tree-structured national curriculum knowledge graph to ensure pedagogical alignment. During transformation, it recalibrates objectives and iteratively adjusts tasks.
- LearningProfile Analysis Agent: Takes natural language descriptions of a class provided by the teacher (along with any available LMS data) and synthesizes them into a structured, schema-validated JSON profile. This profile tracks knowledge readiness, learning preferences (e.g., analytical vs. creative), student interests, and contextual factors (e.g., rural vs. urban, hardware availability).
Interactive User Interface
The frontend provides a three-panel workspace:
- Profile Panel: Allows teachers to view and manually adjust AI-summarized student characteristics using sliders and dropdowns.
- Lesson Plan Workspace: Displays the lesson in editable, semantically grouped blocks where teachers can trigger localized AI revisions.
- Review Window: Provides explicit Traces and Reasons—visual connecting lines that explain why the AI modified a specific task (e.g., lowering programming difficulty because the student profile indicates weak debugging skills).
Datasets
The system does not train a new model from scratch but builds on top of general-purpose LLMs. For evaluation, the authors utilized real-world standardized K-12 Information Technology (IT) lesson plans co-developed by a school-level preparation team.
Key Insights and Ideas
- Workflow Alignment is King: The primary reason teachers reject conventional AI lesson planners is that they force creation from scratch. AdaPT succeeds by mimicking the natural human workflow of modification and editing.
- Context Over Raw Difficulty: Simply telling an AI to "make a task easier" yields poor results. AdaPT achieves better engagement by swapping out abstract or urban-centric examples for culturally relevant ones (e.g., replacing urban transportation examples with local agricultural scenarios for rural students).
- Traceability Breeds Trust: Showing the exact rationale behind an AI-generated change via visual anchors drastically lowers the teacher's cognitive burden when reviewing and validating automated suggestions.
Key Results
The system was evaluated via a user study with 12 participants (9 in-service teachers and 3 domain experts) and compared against a baseline of directly prompting ChatGPT.
- High User Ratings: On a 7-point Likert scale (where 4 is the neutral baseline representing standard ChatGPT use), AdaPT scored exceptionally high across all measured dimensions:
- Completeness of core components: 6.83
- Logical coherence across sections: 6.58
- Formatting consistency: 6.33
- Responsiveness to profile differences: 6.17
- Drastic Time Savings: In real-world pilot testing, teachers using AdaPT completed localized lesson plan adaptations within 15–20 minutes, compared to the 1 hour+ typically required for manual adaptation.
Limitations and Future Work
- Sensitivity to Difficulty Gradients: LLMs still struggle to grasp fine-grained task difficulty. They can easily distinguish between basic and advanced tasks but falter when creating smooth, multi-tiered learning progressions.
- Token Overhead and Cost: Generating highly contextualized and precise adaptations requires long, detailed prompts, resulting in high computational and token costs.
- Cross-Domain Dependencies: The current student profiling model treats subjects in isolation. Future work aims to incorporate cross-domain analytics, recognizing that a student's struggle with math word problems might stem from reading comprehension deficiencies rather than mathematical ability.
Summary updated