PaperlayerView exact revision on arXiv

Source-linked AI summary

LectūraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching

Jaward Sesay, Yue Yu, Siwei Dong, Börje F. Karlsson

arXiv:2606.16428v3cs.CLcs.AIcs.HC

TL;DR

Personalized learning systems often adapt recommended content without modeling how instruction is delivered through multimodal, embodied teaching. LectūraAgents addresses this gap with hierarchical multi-agent preparation and delivery, reporting improvements in lecture quality, personalization, assessment, and embodied teaching over baseline frameworks.

  • Problem

    Existing personalized learning systems typically adapt recommended content rather than how instruction is delivered through multimodal and embodied teaching.

  • Method

    LectūraAgents uses a hierarchical ProfessorAgent-led team for personalized lecture preparation and embodied delivery, with TASA aligning teaching actions to speech and content.

  • Results

    LectūraAgents showed substantial improvements over baseline frameworks in lecture content quality, personalization, assessment quality, and embodied teaching performance.

  • Takeaways & Limitations

    LectūraAgents provides a pedagogically grounded framework for adaptive personalized learning experiences spanning lecture preparation and embodied delivery.

  • Takeaways & Limitations

    Its action–speech alignment relies heavily on offline heuristics and limited teaching actions, potentially constraining instructional richness and robustness across slide layouts.

Abstract

from arXiv · show

Effective personalized AI-assisted learning demands systems that can not only generate accurate learner-specific educational materials, but also dynamically adapt their instruction to diverse learners. However, existing educational agents have primarily focused on lecture content automation and simulations, which often fall short of modelling multimodal and embodied instructional methods tailored for the individual learner. To this end, we propose LectūraAgents - a multi-agent framework that enables personalized learning through end-to-end adaptive embodied teaching. At its core, LectūraAgents mirrors a professor-student relationship, in which a ProfessorAgent leads a collaborative team of specialized subordinate agents through research, planning, review, and embodied delivery of lecture contents that adapt to a learner's needs. The framework offers three main contributions: (1) a hierarchical multi-agent architecture for end-to-end personalized learning; (2) an adaptive embodied teaching mechanism, wherein the ProfessorAgent executes visible and pedagogically motivated teaching actions (e.g., handwrite, highlight, underline, etc.) over contents in a teaching environment; and (3) a Teaching Action-Speech Alignment (TASA) algorithm that employs salience-based heuristics and temporal semantic segmentation to generate coherent teaching action sequences aligned with learner profiles. We evaluate LectūraAgents on diverse courses at high school, undergraduate, and graduate levels using sample-specific rubric-based analysis; with generated lecture materials and teaching actions assessed and validated by expert educators. Experimental results show consistent gains in lecture content quality, embodied teaching quality, assessment, and personalization over existing approaches, positioning LectūraAgents as a pedagogically well-grounded framework for personalized learning at scale.

1 Introduction

LectūraAgents addresses the gap between adapting instructional recommendations and adapting how content is delivered by providing hierarchical, end-to-end personalized lecture generation and embodied delivery. It coordinates specialized agents across preparation and teaching sessions and evaluates personalization across diverse educational levels.

  • Motivation: Existing personalized learning solutions typically adapt what is recommended, rather than how instructional content is delivered.
  • Related Work: Many related educational-agent frameworks remain limited to virtual simulations, teacher–student dialogue, or workflows that generate personalized learning materials.
  • Contributions: LectūraAgents proposes a hierarchical multi-agent framework for end-to-end personalized lecture generation and embodied lecture delivery.
  • Architecture: A ProfessorAgent coordinates specialized validator and executor agents across Lecture Preparation and Lecture Delivery sessions to plan, research, generate, evaluate, and teach lecture artifacts.
  • Embodied Teaching: The framework includes a Teaching Action-Speech Alignment algorithm and uses visible, pedagogically motivated teaching actions to adapt instruction to individual learners.These actions are enacted during the teaching session using lecture artifacts prepared earlier.
  • Evaluation: Evaluations across high school, undergraduate, and graduate courses assessed lecture quality, teaching quality, and personalization, finding high-quality artifacts and coherent embodied teaching action sequences.

2 Related Work

Related work spans cognitive foundations of personalized learning, general-purpose and educational multi-agent frameworks, and embodied teaching systems. However, existing approaches do not yet provide coherent end-to-end personalized, adaptive, and pedagogically informed embodied instruction.

  • Personalized Learning Foundations: Personalized learning builds on memory theories that motivate learner-centred pathways, individualized pacing, and instruction adapted to how learners process and retain information.Atkinson and Shiffrin’s encoding-and-rehearsal model and Cowan’s account of memory capacities provide cited cognitive foundations.
  • Educational Multi-Agent Frameworks: LLM-agent research established planning, tool use, task decomposition, and multi-step coordination, enabling educational multi-agent frameworks.Examples include EduAgent’s cognitive-science-based student personas, Agent4Edu’s memory-based learner simulations, and EducationQ’s teacher-student-evaluator interactions.
  • Embodied Teaching: Embodied teaching combines verbal instruction with spatial actions such as writing, highlighting, underlining, and pointing to guide attention, reduce cognitive load, and support concept formation.Earlier systems including AutoTutor demonstrated conversational scaffolding through animated pedagogical agents.
  • Limitations of Existing Approaches: Existing action-based systems, including PASS, emphasize instructional cues but fall short of coherent end-to-end personalized, adaptive, and pedagogically informed embodied instruction.PASS automated slide and speech generation from word documents, illustrating progress toward action-based instructional delivery without resolving the broader integration gap.

3 Lect¯uraAgents

LectūraAgents is an end-to-end hierarchical multi-agent framework that combines planning, research, pedagogical embodiment, personalization, and continual learning. It separates lecture preparation from adaptive delivery, where the ProfessorAgent coordinates specialized agents and executes motivated teaching actions while speaking.

  • Architecture: LectūraAgents integrates planning, research, and pedagogical embodiment in an extensible hierarchical architecture supporting personalization and continual learning.The framework is organized into four interconnected modules spanning lecture preparation and delivery.
  • Multi-agent Collaboration: The ProfessorAgent coordinates validator and executor agents through ranked collaboration for planning, research, lecture-artifact creation, review, and delivery.The team includes LecturePlanner, ResearchAgent, SlideAgent, ScriptAgent, SpeechAgent, and TasaAgent, organized across three ranks.
  • Lecture Delivery: Lecture delivery supports Teach Mode for generating new personalized lectures and Study Mode for real-time question answering over learner-uploaded materials.The ProfessorAgent acts as an embodied instructor, performing semantically bounded, visually interpretable, pedagogically motivated operations over teaching-environment contents while speaking.
  • Lecture Preparation: Lecture preparation produces personalized plans, slides, scripts, speech, teaching actions, notes, and assessments conditioned on learner profiles, preferences, and usage history.The preparation session begins from a lecture prompt and learner profile, with optional syllabus or reference materials and configurable academic level, language, persona, and slide count.
  • Alignment: The TasaAgent aligns teaching actions with slide contents and speech by temporally segmenting content and scripts and applying salience-based heuristics.Supported actions include Rough Notation, such as highlighting, underlining, circling, and boxing, and Handwriting actions for writing key points while speaking.

4 Experiments

LectūraAgents is evaluated end-to-end across personalized lecture generation and embodied teaching, using expert rubric-based assessments of content quality, personalization, assessment, and teaching actions. Across model, baseline, and learner-group comparisons, the framework demonstrates strong overall performance and consistent advantages in personalization and post-learning outcomes.

  • Evaluation Design: The experiments assess lecture content quality, teaching quality, assessment, and personalization through quantitative and qualitative evaluations.The evaluation addresses both personalized lecture generation and embodied teaching capabilities.
  • Evaluation Design: The study evaluates 280 personalized lectures generated under seven frontier models, with 40 lectures per model spanning high school, undergraduate, master’s, and PhD profiles.Lectures use the same prompts, learner profiles, and Kokoro TTS model to support fair comparison.
  • Model Results: Gemini 3 Pro ranks first overall with an AAR of 80.4%, supported by LCQ (80.2%), PQ (83.3%), AQ (81.6%), and TAQ (76.5%).GPT-5.1 remains competitive with an AAR of 78.8%, particularly in AQ (82.3%) and PQ (80.5%).
  • Embodied Teaching: Embodied teaching produces generally accurate and coherent action sequences, with strong performance in spatial accuracy, handwriting, rough notation, and embodied teaching.The results indicate that the framework reliably converts generated lecture materials into visible instructional actions.
  • Comparative Results: LectūraAgents scores higher than baseline systems across LCQ, PQ, and AQ, with the largest difference in personalization quality.It also achieves the strongest post-learning assessment performance across learner groups and higher perceived understanding, assessment readiness, future learning support, and overall learning experience.

5 Limitations and Future Work

LectūraAgents has limitations in teaching action–speech alignment, embodied-instruction coverage, and computational efficiency. Future work should address heuristic alignment, expand supported teaching actions, and reduce orchestration overhead.

  • Limitations: The teaching action–speech alignment module relies heavily on offline heuristics, limiting its robustness across diverse slide layouts.The passage identifies heuristic dependence as a current limitation that may constrain alignment robustness.
  • Limitations: The module supports a limited set of teaching actions, which may constrain the richness of embodied instruction.Examples of teaching actions are not specified in the passage beyond the broader framework context.
  • Limitations: Multi-agent orchestration can introduce latency and compute overhead.The passage lists orchestration cost as an additional limitation of the framework.

6 Conclusion · Appendix A

LectūraAgents is introduced as a hierarchical multi-agent framework for end-to-end adaptive, personalized AI-assisted learning, organized around a ProfessorAgent-led professor-student relationship. Its personalized and embodied capabilities, including TASA, were evaluated across educational levels and with real students, showing substantial improvements over baseline frameworks.

  • 6 Conclusion: LectūraAgents provides end-to-end adaptive and personalized AI-assisted learning experiences.The framework is hierarchical and designed to address personalization challenges in AI-assisted learning.
  • 6 Conclusion: A ProfessorAgent leads specialized subordinate agents through research, planning, evaluation, and embodied instructional delivery.This organization frames the system as a collaborative professor-student relationship.
  • 6 Conclusion: The framework adapts instructional contents to diverse students through personalized and embodied capabilities.The conclusion identifies these capabilities as supporting enhanced learning and study experiences.
  • 6 Conclusion: TASA is highlighted as a capability supporting LectūraAgents’ personalized and embodied learning experience.The passage gives TASA as an example of the framework’s personalized and embodied capabilities.
  • 6 Conclusion: The evaluation included pedagogical experiments with frontier models across high school, undergraduate, and graduate-level topics.The experiments assessed the framework across multiple educational levels.
  • 6 Conclusion: A separate efficacy study evaluated LectūraAgents with real students.The conclusion describes this as one of the two main experiments.
  • 6 Conclusion: Experimental results showed substantial improvements over baseline frameworks in lecture content quality, personalization, and assessment quality.The supplied passage reports these improvement areas but does not provide numerical values.

A.1 Lect¯uraAgents: Detailed Architecture … A.1.8 Slide Content Block Types

LectūraAgents uses a modular, hierarchical multi-agent architecture for coordinated lecture generation and delivery. Its lifecycle, collaboration mechanisms, tools, memory, model interface, and structured slide blocks support adaptive, personalized, and consistently rendered instruction.

  • A.1.1 Core Modules and Components: Four core modules support agent coordination, LLM integration, teaching-action alignment, memory management, and content rendering across lecture generation and delivery.The modular design enables extension and maintenance of individual components.
  • A.1.2 Agent Hierarchy and Roles: Agents are organized into three ranks, with ProfessorAgent coordinating and validating, Rank 2 agents managing execution and validation, and Rank 3 agents performing specialized tasks.Responsibilities and tool access are defined according to each agent’s role.
  • A.1.3 Agent States and Lifecycle: A state machine moves agents from IDLE through acknowledgment and execution to completion, failure, or revision, supporting tracking, error handling, and feedback-driven revision.Higher-ranking agents can provide feedback that enables lower-ranking agents to revise their work.
  • A.1.4 Multi-agent Collaboration: Agents collaborate sequentially for dependent tasks and in parallel for independent tasks, while SwarmOfRanks coordinates activities across hierarchical ranks.These patterns orchestrate the multi-stage lecture generation workflow.
  • A.1.5 Tools and Capabilities: Modular reusable tools provide capabilities including web search, file parsing, text-to-speech synthesis, and code execution through clear interfaces.The tools abstract API interactions, file processing, and multimedia generation.
  • A.1.6 Adaptive Memory: A three-layer memory architecture combines short-term session interactions, long-term learner data, and dynamic learning-pattern adaptation through a unified adaptive-memory interface.Agents can access relevant learner context efficiently across these memory types.
  • A.1.7 LLMs: A unified API supports multiple frontier models from leading LLM providers, enabling model switching by task requirements, cost considerations, and performance needs.Provider-specific implementations handle authentication, API communication, and response formatting without requiring agent code changes.
  • A.1.8 Slide Content Block Types: Slides support multiple pedagogically targeted content blocks, including definitions, equations, examples, steps, and questions, which are automatically rendered with consistent styling and formatting.The block types provide structured presentation for core concepts and learner engagement.

A.2 More on Evaluation Methodology · A.2.1 Overview

LectūraAgents is evaluated with rubric-based pedagogical and comparative assessment, using expert-validated scores for generated learning and teaching artifacts. The methodology examines personalized lecture-content generation and embodied use of those materials across diverse learner profiles.

  • A.2.1 Overview: The evaluation uses a rubric-based methodology for pedagogical and comparative assessment.Generated learning and teaching artifacts are scored and validated by expert educators.
  • A.2.1 Overview: Expert educators score and validate the generated learning and teaching artifacts.The validation covers both learning materials and embodied teaching artifacts.
  • A.2.1 Overview: The methodology evaluates LectūraAgents’ ability to generate high-quality personalized lecture content.This capability is examined for diverse learner profiles.
  • A.2.1 Overview: The evaluation covers personalized lecture-content generation for diverse learner profiles.Learner diversity is an explicit condition of the evaluation.
  • A.2.1 Overview: The methodology evaluates the framework’s ability to use generated materials during embodied teaching.This tests how generated materials function in the teaching process.
  • A.2.1 Overview: LectūraAgents is assessed on two core capabilities: personalized lecture-content generation and embodied teaching with those materials.These capabilities define the central scope of the evaluation.

A.2.2 Lect¯uraAgents’ Pedagogical Evaluation Under Frontier Models · A.2.3 Rating

The pedagogical evaluation spans 280 lectures across seven models, academic levels, learner profiles, and subject areas, using rubric-based metrics for content, assessment, and delivery. Ratings convert weighted boolean rubric outcomes into metric-level and overall Average Achieved Ratings, with satisfied and failed criteria receiving graded rewards or penalties.

  • A.2.2 Lect¯uraAgents’ Pedagogical Evaluation Under Frontier Models: Lecture content was evaluated through Cognitive Load, Syllabus Coverage, Instruction-following, Adaptive Emphasis, Preference Alignment, Engagement, Motivation, and Tone/Style metrics.These metrics assess content alignment, topic coverage, instruction adherence, learner adaptation, preferences, engagement, motivation, and language appropriateness.
  • A.2.2 Lect¯uraAgents’ Pedagogical Evaluation Under Frontier Models: Assessment Quality was measured by Concept Coverage, Cognitive Appropriateness, Answer Validity, and Rationale criteria across assessment materials and learner-profile inputs.The criteria examine topic coverage, difficulty and learner alignment, solution accuracy, and explanation quality.
  • A.2.2 Lect¯uraAgents’ Pedagogical Evaluation Under Frontier Models: Lecture Delivery Evaluation assessed Temporal Alignment, Accurate Handwriting Action, Accurate Rough Notation, and Spatial Accuracy.These criteria verify action-speech timing, handwriting correctness, annotation correctness and timing, and annotation precision.
  • A.2.3 Rating: Each rubric criterion was evaluated as boolean, then weighted and averaged to produce Average Achieved Ratings at metric and overall levels.The overall score for each lecture under a model is computed as the weighted average of passed rubric criteria.
  • A.2.3 Rating: Satisfied criteria contributed +5, +3, or +1 according to whether the behavior was highly desirable, desirable and important, or nice-to-have.The rating weights were selected from {-5, -3, -1, 0, +1, +3, +5}.
  • A.2.3 Rating: Unsatisfied criteria were treated as failures with scores of 0, -1, -3, or -5, representing no credit through critical failure.The penalties distinguish lowest-severity, minor, moderate, and critical failures.

A.2.4 Expert Recruitment and Evaluation Procedure

The evaluation used a purposively sampled panel of five experienced educators from diverse teaching contexts and disciplines. Before assessment, experts attended an online workshop reviewing the evaluation dimensions, criteria, and weighting scheme.

  • Expert Recruitment: Five expert educators were recruited through purposive sampling based on teaching, curriculum development, and educational assessment experience.The panel included secondary-school teachers and university instructors, each with at least five years of teaching experience.
  • Expert Recruitment: The panel represented STEM, social science, and humanities disciplines across secondary-school and university teaching contexts.This composition provided expertise spanning multiple educational levels and subject areas.
  • Evaluation Procedure: Before evaluation, the experts participated in an online workshop reviewing the evaluation dimensions, criteria, and weighting scheme.

A.2.5 Comparative Evaluation of Lect¯uraAgents with Related Frameworks · Appendix B

LectūraAgents was comparatively evaluated against Instructional Agents, GenMentor, and Google’s Learn Your Way using learner-profiled lecture generation. The comparison covered lectures across educational levels and included topic and profile documentation.

  • A.2.5 Comparative Evaluation of Lect¯uraAgents with Related Frameworks: The comparison included two multi-agent frameworks—Instructional Agents and GenMentor—and Google’s Learn Your Way system.
  • A.2.5 Comparative Evaluation of Lect¯uraAgents with Related Frameworks: Researchers generated 20 lectures using the released code of Instructional Agents and GenMentor.
  • A.2.5 Comparative Evaluation of Related Frameworks: The 20 lectures comprised 5 lectures for each educational level and spanned 10 learner profiles.
  • A.2.5 Comparative Evaluation of Lect¯uraAgents with Related Frameworks: The same lectures were generated with LectūraAgents for performance comparison against the related frameworks.
  • A.2.5 Comparative Evaluation of Lect¯uraAgents with Related Frameworks: Table A11 summarizes the lecture topics and learner profiles generated for each framework or system.
  • A.2.5 Comparative Evaluation of Lect¯uraAgents with Related Frameworks: Example topics included World War II’s causes and consequences and an introduction to large language models.
  • A.2.5 Comparative Evaluation of Lect¯uraAgents with Related Frameworks: The examples represented a 12th-grade history learner preferring timeline-based explanations and an undergraduate computer science learner interested in artificial intelligence.

B.1 Code and Data · B.1.1 Installation and Usage

The study provides its supporting data through a Hugging Face repository and makes the code available upon reasonable request. Installation requires API-key configuration and package setup, while users can generate and deliver lectures through either a browser-based frontend or terminal commands.

  • B.1 Code and Data: Supporting data is available in the LectūraAgents Hugging Face repository, while the code can be obtained from the corresponding author upon reasonable request.The documentation directs users to follow the installation instructions or README to get started.
  • B.1.1 Installation and Usage: Users must add an LLM API key and, optionally, a SerpApi key to the parent-directory .env file before installation.Supported LLM providers include OpenAI, Anthropic, Gemini, and Deepseek; SerpApi is recommended to help reduce hallucination.
  • B.1.1 Installation and Usage: From the parent directory, users install the required packages with pip3 install -r requirements.txt.This is the documented package-installation command.
  • B.1.1 Installation and Usage: The frontend opens the teaching environment at http://127.0.0.1:8080/, where users can try existing lectures or generate new ones through chat or prompt panes.Generated lecture materials appear below the slide as they are produced, and the initial interface is illustrated in Figure 13.
  • B.1.1 Installation and Usage: During lecture generation, the group chat session displays the full process unfolding in real time.The documented view is presented as the Swarm-of-Ranks group chat in Figure 14.
  • B.1.1 Installation and Usage: After generation, the teaching environment automatically updates with the slide deck and controls for Next, Play, Previous, Restart, Temporal Segmentation, and Chat.The completed teaching-and-learning view is shown in Figure 15.
  • B.1.1 Installation and Usage: Terminal-based generation uses lecture_prep.py with options for the lecture title, description, learner profile, slide count, academic level, handwriting mode, slide-image mode, syllabus, materials, and output directory.The documented command also supports instructor voice, LLM, and research settings, while lecture_delivery.py loads the generated lecture into the teaching environment.
Loading 2606.16428v3…