Paper deep dive

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

Jaward Sesay, Yue Yu, Siwei Dong, Yemin Shi, Guangyao Chen, Börje F. Karlsson

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 101

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 99%

Last extracted: 6/20/2026, 8:20:59 AM

Summary

LectūraAgents is a hierarchical multi-agent framework designed for adaptive, personalized, and embodied AI-assisted learning. It utilizes a professor-student relationship model where a central ProfessorAgent coordinates a team of specialized agents (validator and executor agents) to manage the entire lecture lifecycle, from preparation to delivery. The framework features a Teaching Action-Speech Alignment (TASA) algorithm to synchronize verbal instruction with visual teaching actions (like highlighting or handwriting) and supports two modes: 'Teach Mode' for generating new personalized lectures and 'Study Mode' for interacting with existing materials. The system aims to bridge the gap between static content generation and dynamic, embodied instructional delivery tailored to individual learner profiles.

Entities (11)

LectūraAgents · framework · 100%LecturePlanner · agent · 100%ProfessorAgent · agent · 100%ResearchAgent · agent · 100%ScriptAgent · agent · 100%SlideAgent · agent · 100%SpeechAgent · agent · 100%Study Mode · mode · 100%TASA · algorithm · 100%TasaAgent · agent · 100%Teach Mode · mode · 100%

Relation Signals (5)

TASA → aligns → Teaching Action and Speech

confidence 100% · a Teaching Action-Speech Alignment (TASA) algorithm that employs salience-based heuristics... to generate coherent teaching action sequences

LectūraAgents → contains → TASA

confidence 100% · LectūraAgents consists of four interconnected modules: LLM, Agent, TASA, and Memory.

ProfessorAgent → coordinates → ResearchAgent

confidence 100% · the ProfessorAgent leads a collaborative team of specialized subordinate agents... ResearchAgent, SlideAgent, ScriptAgent, SpeechAgent, and TasaAgent

ProfessorAgent → leads → LecturePlanner

confidence 100% · the ProfessorAgent leads a collaborative team of specialized subordinate agents... the ProfessorAgent instructs the LecturePlanner

LecturePlanner → manages → ResearchAgent

confidence 90% · the validator agent (LecturePlanner), who in turn manages executor agents

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract: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.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

100,201 characters extracted from source content.

Expand or collapse full text

Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Jaward Sesay Beijing Institute of Technology Yue Yu † Beijing Institute of Technology Siwei Dong * Peking University Yemin Shi * Peking University Guangyao Chen * Cornell University Börje F. Karlsson * Beijing Academy of Artificial Intelligence Figure 1: Overview of Lect ̄ uraAgents: a hierarchical multi-agent framework for end-to-end adaptive personalized learning experiences. Given a lecture prompt or learning materials and a learner profile, a ProfessorAgent leads a collaborative team of specialized agents through research, planning, design, evaluation and embodied delivery of lecture and study contents that adapt to the individual learner. The framework provides students with access to real-time adaptive, personalized teaching and study sessions. Abstract Effective personalized AI-assisted learning demands learning sys- tems that can not only generate accurate learner-specific educa- tional materials, but also dynamically adapt their instruction to diverse learners. However, existing educational agent frameworks have primarily focused on lecture content automation and simula- tions, which often fall short of modelling multimodal and embodied instructional methods tailored for the individual learner. To this end, we propose Lect ̄ uraAgents—a multi-agent framework that en- ables personalized learning through end-to-end adaptive embodied teaching. At its core, Lect ̄ uraAgents mirrors a professor-student relationship, in which the ProfessorAgent leads a collaborative team of specialized subordinate agents through research, planning, re- view, and embodied delivery of lecture contents that adapt to a learner’s needs. The framework offers three main contributions: (1) † Corresponding author: yuyueanny@hotmail.com * Equal contributing authors The dataset for this work is available at: https://huggingface.co/datasets/Jaward/ lectura-agents-data. 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 con- tents in a teaching environment while speaking; 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 Lect ̄ uraAgents on diverse courses at high school, under- graduate, 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 Lect ̄ uraAgents as a pedagogically well-grounded frame- work for personalized learning at scale. 1 arXiv:2606.16428v1 [cs.CL] 15 Jun 2026 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching 1 Introduction Adaptive personalized AI-assisted learning has emerged as a promis- ing approach for tailoring instructions to individual learners, with studies reporting gains in motivation, engagement, and learning outcomes, especially in online educational settings [1–3]. However, contemporary personalized learning solutions and frameworks typ- ically focus on adapting what is recommended, rather than how instructional content is delivered to the learner [4]. Research on embodied teaching shows that performing teaching actions (e.g., writing, pointing or gesturing) during a lecture can help guide atten- tion, foster conceptual understanding, and enhance overall learning outcomes [5–7]. These findings point to the need for personalized learning solutions that well integrate adaptive learning contents with embodied instructional delivery. Recent frontier models demonstrate strong reasoning and agen- tic capabilities that have enabled planning, tool-use, and multi-step problem solving, opening new possibilities for applications in per- sonalized learning [8–11]. This breakthrough has led to the explo- ration of LLM-powered agent frameworks for education, where specialized agents automate learning and teaching tasks to support students and educators [12]. Moreover, recent efforts have further demonstrated the potential of leveraging multiple agents to act as personal tutors and learning companions that provide on-demand teaching and learning support based on individual needs [13–16]. However, the predominant focus of most related frameworks has been on simulations, where agents enact roles in virtual classrooms [17–19] simulate teacher–student dialogues to evaluate teaching behaviors and feedback strategies [20,21], or coordinate agent workflows for generating personalized learning materials [22,23]. These are important proof of concepts, but their impact is limited to controlled virtual environments that do not capture the myriad nu- ances of adaptive personalized learning in real life scenarios. Other works have explored single-agent or prompt-engineered LLM tu- toring systems [24–26] that generate explanations, feedback, or instructional contents, but without rigorous review or modeling of how such contents should be contextualized and adapted to diverse learning profiles. Few related works extend beyond these scopes to adopt a broader personalized instructional perspective that is often centered on automating course content generation [27,28], which is primarily delivered in text-only modality, with no account for personalized embodied instructional delivery. Collectively, these systems offer valuable contributions to AI-assisted learning but remain fragmented in scope, lacking a unified model that connects personalized content generation with adaptive embodied delivery. Consequently, key pedagogical features, including coordinated les- son planning, iterative content review, embodied teaching, and alignment between teaching behaviour and learner needs, remain insufficiently addressed. To address these limitations, we propose Lect ̄ uraAgents, a hierar- chical multi-agent framework for end-to-end personalized lecture generation and embodied lecture delivery. Our framework moves beyond simulations and static content generation, to managing the entire life cycle of a lecture (i.e., from preparation to delivery, as shown in Figure 1), while adapting to individual learning prefer- ences. Lect ̄ uraAgents offers three primary contributions: 1.A hierarchical multi-agent architecture for end-to- end personalized learning: we propose the first multi- agent framework with end-to-end personalization for learn- ing. It mirrors a professor–student relationship, where a ProfessorAgent coordinates specialized assistant agents (at different hierarchies) to plan, research, review, and create lecture contents tailored for the individual learner. 2.A Teaching Action-Speech Alignment (TASA) algo- rithm: a novel technique that uses LLM-based semantic analysis, temporal content segmentation, and salient heuris- tics to accurately align relevant teaching actions to regions or contents in a teaching environment (e.g., over a slide). 3.An embodied lecture delivery mechanism: our frame- work enables a ProfessorAgent to perform visible, inter- pretable teaching actions (e.g., highlight, handwrite, under- line, etc.) directly over contents in the teaching environ- ment (in our case, lecture slides) with a clear pedagogical rationale for each action taken. Lect ̄ uraAgents decomposes personalized instruction into agents operating at three hierarchies across two sessions: Lecture Prepa- ration and Lecture Delivery. In the preparation session, the Pro- fessorAgent leads a team of validator and executor agents through planning, research, generation, and evaluation of lecture artifacts. During the delivery or teaching session, the ProfessorAgent utilizes these artifacts to enact an embodied teaching role, executing visible and pedagogically motivated teaching actions on contents in the learning environment. We conducted extensive evaluations of the framework on di- verse courses at high school, undergraduate, and graduate levels, assessing lecture quality, teaching quality and personalization. Our experiments show that Lect ̄ uraAgents can produce high quality lecture artifacts, while effectively adapting personalized teaching strategies to diverse learner profiles through coherent embodied teaching action sequences. 2 Related Work 2.1Adaptive Personalized AI-Assisted Learning The idea of personalized learning predates LLMs and LLM agents. Early theories of memory, such as Atkinson and Shiffrin’s model of how information is encoded and rehearsed [39] and Cowan’s ac- count of short-term and long-term memory capacities [40], helped establish the cognitive foundations for adapting instruction to the ways learners process and retain information. These insights in- spired models of personalized learning that emphasized learner- centered pathways, individualized pacing and tailored support. Be- fore LLMs became widely adopted, deep learning models were used in intelligent tutoring systems (ITS) to monitor learners’ perfor- mances, adjust task difficulty, and deliver personalized feedback [41–43]. Reviews show that AI-assisted personalized learning has a positive impact on students’ engagement and learning outcome across diverse learning settings [44–46]. More recent empirical studies of AI-driven adaptive platforms in university and language- learning contexts, report gains across performance, satisfaction, and self-directed learning [47–49]. Collectively, these findings make clear the significance of adaptive personalized learning, forming the foundations upon which our framework is built. 2 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Table 1: Comparison of Lect ̄ uraAgents with existing multi-agent frameworks in this domain FrameworkTeaching ModalityEmbodied Agent(s) Teaching Action Alignment Personalization Multi-Agent Collaboration EduAgent [29]Text✗ Agent4Edu [30]Text (simulation)✗✓ EducationQ [31]Text (simulation)✗ FACET [32]Text✗✓✗ KELE [33]Text✗ Instructional Agents [34]Text✗✓ EduPlanner [35]Text✗✓ GenMentor [36]Text✗✓ SimClass [37]Text (simulation)✗✓ WikiHowAgent [38]Text✗✓ Lect ̄ uraAgentsMultimodal (Embodied)✓ 2.2 LLM Agent Frameworks for Education Early works on LLM agents demonstrated how language models can plan, use tools, decompose tasks, and coordinate multi-step reasoning across multiple collaborating agents [50–53]. These apa- bilities soon inspired educational multi-agent frameworks [54]. For instance, EduAgent [29] models diverse student personas us- ing cognitive-science priors, Agent4Edu [30] simulates learner responses with memory-based generative agents, and Educa- tionQ [31] stages multi-agent teacher-student-evaluator interac- tions to assess teaching behaviours. Similarly, systems like Sim- Class [37] and WikiHowAgent [38] extend simulation to classroom dynamics and procedural learning. Course-content automation then became a focus, with Instructional Agents [34] generating full course materials through role-based collaboration, and EduPlan- ner [35] iteratively refining lesson plans via evaluator–optimizer agent loops. More recent works have also introduced personaliza- tion: FACET [32] creates learner-adapted worksheets, KELE [33] provides concept-level enrichment and feedback, and GenMen- tor [36] builds personalized learning paths from learner goals. While these contributions demonstrate how multi-agent systems can en- hance learning, they lack relevant capabilities (as summarized in Table 1) that integrates personalized content generation with em- bodied instructional delivery. 2.3 Embodied Teaching Agents Embodied teaching in digital settings refers to instructional meth- ods that combine verbal instruction with spatial teaching actions (e.g., writing, highlighting, underlining, or pointing) over learning contents in a virtual teaching environment. These actions help guide attention, reduce cognitive load, and support concept forma- tion [55,56]. Earlier models like AutoTutor and its variations [57,58] demonstrated the benefit of animated pedagogical agents capable of conversational scaffolding. Recent systems have explored program- matic video-based approaches, for example, Xu et al. [59] explored how AI-generated lecture videos compare with real lectures, while AutoLectures [60] converts slides into narrated videos with high- light actions (using Levenshtein and LLM-based matching), and PASS [61] automated slide and speech generation from word doc- uments. These efforts emphasize the importance of action-based instructional cues, but fall short of delivering a coherent end-to- end personalized, adaptive, and pedagogically informed embodied instruction. 3 Lect ̄ uraAgents We designed Lect ̄ uraAgents to be both domain-specific and exten- sible given the nature of the problem we are trying to solve. Its framework integrates planning, research, and pedagogical embodi- ment within a cohesive, end-to-end hierarchical architecture that supports both personalization and continual learning in tandem. Lect ̄ uraAgents consists of four interconnected modules: •LLM – The LLM module provides agents with access to frontier models (e.g., GPT-5, Gemini 3 pro, Claude Sonnet 4, Deepseek V3.2, Qwen 3, and Kokoro TTS [62]) through their respective custom APIs. It serves as the brain behind agents, handling text, image, and speech modalities. •Agent – This module holds the core logic for each agent’s role, capabilities, and tools. It also enables coordinated multi-agent collaboration through dynamically invoked ac- tions for assigned tasks. The framework adopts a three-tier hierarchical collaborative mechanism with a lead coordina- tor agent managing a validator agent, who in turn manages executor agents for lecture content generation. To complete tasks, agents execute a series of actions utilizing custom tools. •TASA – The Teaching Action-Speech Alignment (TASA) module handles alignment between embodied teaching ac- tions and their corresponding lecture speeches. It provides logic for salient heuristic analysis and temporal semantic segmentation, which help provide context when curating relevant teaching action sequences. •Memory – This module implements short-term, long-term, and dynamic memories, which together allow agents to preserve context, track learner needs, and adapt their be- haviour over time. 3 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Figure 2: Lect ̄ uraAgents Architecture. The framework adopts a hierarchical multi-agent architecture, modeled after a professor–students’ relationship. One in which a coordinator agent (or ProfessorAgent) guides a collaborative team of validator and executor agents through planning, research, design, and delivery of personalized lecture contents. Multi-agent collaboration is mediated through an orchestration layer with group-chat communication that enables iterative planning, self-evaluation, and continuous refinement of generated materials. This architecture is supported by four interconnected modules: Agents, LLM, TASA, and Adaptive Memory. These four modules span across the framework’s two main stages: Lecture Preparation Session and Lecture Delivery Session. Moreover, as shown in Figure 3, the delivery session supports two modes: Teach Mode, which generates a new personalized lecture based on the learner’s profile and provided learning materials, and Study Mode, which allows learners to upload existing materials, such as notes, books or projects, and interact with the ProfessorAgent through real-time Q&A. 3.1 Lecture Preparation Session In this stage, the ProfessorAgent leads a collaborative team of spe- cialized agents through planning, research, alignment, review, and creation of personalized lecture artifacts (e.g., lecture plan, slides, scripts, speech, teaching actions, notes, etc.). A quick overview of the entire process can be found in Algorithm 1. Lecture Prompts and Configs. Lecture preparation begins by processing the learner’s prompts along with a range of configu- ration choices that define the scope, style, and preferences of the lecture. The prompt captures the lecture topic, its intended cover- age, and the learner profile, so the framework can adapt content depth and learning preferences, while optional syllabus or refer- ence materials help anchor the lecture to a course context or source material. Additional settings specify the instructor persona, target academic level (high school, undergraduate, masters, or PhD), lan- guage of instruction (which currently includes English, Chinese, French, or Spanish), and the approximate number of slides to be generated. Learners can also choose their preferred voice model, handwriting mode (either Handwriting RNN Model or Preset Font Handwriting), LLM model, and research method (using Wikipedia or Google search). Together, these inputs provide the initial condi- tions that guide downstream multi-agent collaboration, planning, research, content generation, and embodied teaching. Our teach- ing and learning environment can be accessed via a browser (as shown in Figure 3) for easier entry of all inputs. Additional details on lecture prompts and configurations can be found in Appendix B. Multi-agent Collaboration. When a lecture is prompted, the ProfessorAgent first initiates the preparation session, creating a col- laborative group chat named Swarm-of-Ranks Group Chat (shown in Figure 4) – where agents at different ranks collaborate to com- plete assigned tasks. In this group chat we have a coordinator (Pro- fessorAgent), a validator (LecturePlanner), and different executor (ResearchAgent, SlideAgent, ScriptAgent, SpeechAgent, and Teaching Action-Speech Alignment agent or TasaAgent) agents. The coor- dinator agent (Rank 1) supervises the validator agent (Rank 2), who in turn manages executor agents at Rank 3. This hierarchical structure allows for efficient review and successful completion of 4 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Figure 3: An example of adaptive personalized learning experience with Lect ̄ uraAgents assigned tasks. Agents communicate by sending messages in the group chat through a communication layer. There are nine message types: [Task], [TaskAcknowledged], [Progress], [TaskCompleted], [Approval], [Revisal], [Handoff ], [RevisalSucceeded], and [Revisal- Failed]. Table 2 shows the message-types respective agents can send in the group chat. Table 2: Message types for respective agents at different ranks Rank(s) Message TypesAgents 1, 2[Task], [Approval], [Revisal] ProfessorAgent and LecturePlanner 1, 2, 3[TaskAcknowledged], [Progress], [TaskCompleted] All agents 3[Handoff ], [RevisalSucceeded], [RevisalFailed] ResearchAgent, SlideAgent, ScriptAgent, SpeechAgent, TasaAgent 5 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Planning. The lecture preparation process starts with planning, wherein the ProfessorAgent instructs the LecturePlanner to draft a lecture plan based on the requested lecture topic and learner pro- file. The LecturePlanner first conducts preliminary research on the topic, then writes a detailed plan, and submits it for review and approval by the ProfessorAgent. The plan contains lecture metadata, learner profile, and detailed descriptions of tasks for each execu- tor agent with respective criteria for completing assigned tasks. Once the plan is approved, the LecturePlanner then instructs and coordinates executor agents to generate lecture contents based on the plan. Subsequent preparation stages will involve sequential execution of tasks by executor agents and iterative validation by the LecturePlanner. Generation. This stage starts with the SlideAgent, which is tasked with designing each slide (in HTML format), using a cus- tom slide builder tool, and generating respective contents based on structural and pedagogical criteria from the lecture plan. Each slide is designed to support contents in text, image, video, and speech modalities, via structured content blocks. Slide images can be either generated or sourced online via web search. Next, the ScriptAgent utilizes the generated slide contents (along with lecture plan and research findings) to create a personalized and pedagogically in- formed script for each created slide. Scripts are conditioned to capture the learner’s attention, level of understanding, and learning preferences. Finally, the scripts are then passed on to the SpeechA- gent which performs speech synthesis, converts scripts to speech (in the learner’s desired instructor voice), and creates word-level timestamps for each speech action using Whisper ASR [63]. These artifacts will later be used during alignment and review. Alignment. Given the generated speech timestamps, scripts, slide contents, and learner profile, the TasaAgent first performs a preliminary teaching action analysis using segmentation and salient heuristic tools in the TASA module. This analysis starts with the temporal semantic segmentation of slide contents and scripts to identify segments that should receive teaching actions; it then applies salience-based heuristics to provide rationale for each teaching action application. Currently, Lect ̄ uraAgents supports two kinds of teaching actions: Rough Notation (RN), e.g., highlight, underline, circle, box, etc., and Handwriting actions (HW), i.e., writing down key points in natural human-like handwriting style, while speaking. This analysis results are then added to the agent’s context when mapping pedagogical teaching actions to contents in the slide teaching environment. The ProfessorAgent will later utilize the resulting teaching action sequences during embodied teaching in the lecture delivery session. Self-reflection. In addition to the hierarchical review mecha- nism present in multi-agent collaboration, we ensure each agent self-reflects on any completed tasks to find and fix issues before submitting results for review by the validator agent. They do this by first reviewing completed tasks, then self-validating them against required criteria detailed in the lecture plan. Personalization. We ensure personalization across all gener- ated lecture contents—slides, images, quizzes, lecture notes, scripts and teaching actions—by conditioning generation on the learner’s profile, learning preferences, and usage history in memory. For example, slide contents, as shown in Figure 5, are adapted to the learner’s interests by framing concepts around a favourite sport Algorithm 1 Lecture Preparation Session Input:Lecture Prompt 퐿 푃 , Learner Profile푈 Parameters:Coordinator/validator agents퐴 푃 ,퐴 퐿푃 ; executor agents 퐸 퐴 =퐴 푅 ,퐴 푆 ,퐴 푆푐 ,퐴 푆푝 ,퐴 푇 ; preparation plan 푃=푃 1 ,푃 2 , . . .,푃 푛 ; adaptive memory 푀 퐴 =푀 푠 ,푀 퐿 ,푀 푑 ; Swarm-of-Ranks group chat 퐺 chat Output: Lecture artifacts퐿 퐴 =Plan, Slides, Script, Speech, TeachingActions, Notes, Assessments 1: Initialize memory푀 푠 ,푀 퐿 ,푀 푑 and agents퐴 푃 ,퐴 퐿푃 2: 퐴 푃 starts prep session and instantiates 퐺 chat 3: 퐴 푃 debriefs 퐴 퐿푃 and requests lecture plan 푃 4: repeat 5: 퐴 퐿푃 drafts 푃 from(퐿 푃 ,푈) → 푀 푑 6: 퐴 푃 reviews 푃 and gives feedback→ 푀 푠 7: 퐴 퐿푃 updates 푃 based on feedback→ 푀 푑 8: until 퐴 푃 approves 푃 or max iterations reached 9: Initialize executor agents 퐸 퐴 10: for each 푃 푖 in 푃 do 11: 퐴 퐿푃 debriefs 퐸 퐴 on assigned tasks 12: for each executor 퐸 푖 in 퐸 퐴 do 13:repeat 14:퐸 푖 plans and executes task 15:퐸 푖 self-reflects and submits 16:퐴 퐿푃 reviews task and gives feedback 17:until 퐴 퐿푃 approves task or max iterations reached 18: end for 19: end for 20: 퐴 퐿푃 submits artifacts 퐿 퐴 for final review by 퐴 푃 21: 퐴 푃 reviews and validates 퐿 퐴 22: return 퐿 퐴 Figure 4: Swarm-of-Ranks group chat or hobby, or can be tailored into an easier-to-follow learning path (e.g., more scaffolding or simpler analogies) when the student pro- file indicates lower prior knowledge. Slide images are generated 6 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Figure 5: Screenshot of a personalized slide for an undergraduate student whose favorite sport is tennis, with key concepts explained using tennis-themed visuals and embodied teaching actions over slide contents. to match the same themes and difficulty level, while quizzes are personalized in both content and phrasing to assess understanding using familiar scenarios. The resulting notes, scripts, and teaching actions mirror these choices to ensure a coherent, learner-specific narrative throughout the lecture. Review. Finally in this session, generated lecture artifacts are as- sembled by the LecturePlanner and submitted to the ProfessorAgent for final review. During review the ProfessorAgent again validates lecture artifacts based on lecture content quality, teaching quality, action alignment, and personalization. Once review is successful, the ProfessorAgent agent then takes on the role of teacher in the subsequent lecture delivery session. 3.2 Lecture Delivery Session During this stage, the ProfessorAgent assumes the role of an em- bodied instructor that executes pedagogical teaching actions in the slide environment using lecture artifacts from the lecture prepa- ration session. In this work, we define a Teaching Action as a semantically bounded, visually interpretable and pedagogically mo- tivated operation performed by the ProfessorAgent over contents in the teaching environment, while speaking. Each action comes with a rationale for why it was taken at a particular time. We experiment with two types of teaching actions: 1. Rough Notations (RN): These are actions that involve marking or emphasizing existing contents on the slide. Ex- amples include highlighting key terms, underlining impor- tant phrases, circling diagrams, or boxing critical points. RN actions are used to draw the learner’s attention to spe- cific areas of the slide that are relevant to the current topic being discussed. For improved user experience, we make use of a hand-drawn annotation library [64] that simulates human-like rough notations for these actions. 2.Handwriting (HW): These actions involve writing new in- formation directly onto the slide canvas in a natural, human- like handwriting style. This can include jotting down defi- nitions, drawing diagrams, or annotating existing content. HW actions serve to reinforce learning by actively engag- ing the learner with newly introduced concepts during the lecture. We utilize both a handwriting recurrent neural net- work model based on Graves [65] and a preset font-based handwriting synthesis for this teaching action. These actions undergo preliminary review, analysis, and alignment using our proposed Teaching Action-Speech Alignment (TASA) algorithm, summarized in Algorithm 2. Teaching Action–Speech Alignment (TASA) Algorithm. TASA uses a combination of LLM-based salience heuristics analysis and temporal semantic segmentation to help guide the TasaAgent with prospective relevant teaching action se- quences. The agent’s objective is to emit an ordered list of ped- agogically informed teaching action-speech sequences퐴푆 seq = 푆 1 [푎 1 ,푎 2 , . . .,푎 푛 ], . . .,푆 푛 [푎 1 ,푎 2 , . . .,푎 푛 ], for each slide푆 푛 , where each action 푎 푛 is given by: 푎 푛 =actiontype 푛 , start 푛 , end 푛 , cfg 푛 (1) 푎푐푡푖표푛푡푦푝푒 푛 can be either RN or HW, (푠푡푎푟푡 푛 ,푒푛푑 푛 ) gives the dura- tion for the action, and푐푓푔 푛 holds additional metadata or configu- ration specific to the action type, as illustrated in Figure 6. Figure 6: Data structure for Rough Notation and Handwriting teach- ing actions in json. Temporal Semantic Segmentation. Before performing salience heuristics analysis, we first segment slide contents and speech semantically (see Figure 7), in order to augment our agent’s context for better teaching action sequences. Segment labels include Pedagogical, Personalized, Salient, Adaptive, and Assessment, each of which helps provide insight into the kind of teaching actions to apply. For each slide region푟푒푔푖표푛 푛 ∈ 푅 푠 and corresponding speech segment with label푙푎푏푒푙 푛 , the TasaAgent creates a segment 푠푒푔푚푒푛푡 푛 given by: segment 푛 =label 푛 , region 푛 , speech_segment 푛 (2) specifically, for each candidate segment푠푒푔푚푒푛푡 푛 in a slide푆 푛 , TASA analyses the segment data and assigns a suitable teaching action along with a rationale푟 푛 in natural language, explaining why this action is appropriate for that specific region. The final heuristics analysis data for a given slide is recorded as: H(푆 푛 )=segment 푛 ,푎 푛 ,푟 푛 (3) which provides the TasaAgent with a structured context when generating the resulting teaching action-speech sequences AS seq . 7 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Figure 7: The left slide shows already taken RN and HW teaching actions on a slide, while the right slide shows temporal semantic segmentation of slide contents with segment labels, action types, rationales, regions and their respective speech and script segments. Embodied Teaching. Given the generated teaching action- speech sequences, the ProfessorAgent dynamically schedules and in- vokes respective teaching action functions over regions in the slide environment (in sequence), while speaking. Each action function is tied to a specific speech segment (with word-level timestamps) and applies a targeted visual operation such as handwriting, highlight- ing, circling, or underlining, directly on the corresponding slide region, as illustrated in Figure 8. To ensure accurate and realistic embodiment, the agent is pro- vided with a discrete world view of the slide environment and its contents, while using a 3D quill-holding hand to execute the embod- ied teaching actions with precise spatial targeting of regions and their corresponding action types. As a result, embodied teaching actions like handwriting, highlighting, circling, etc., are executed in a natural, interpretable, and pedagogically grounded manner that closely mirrors human instructional behaviour. Figure 8: Illustration of Embodied Teaching in Lect ̄ uraAgents Algorithm 2 Teaching Action-Speech Alignment (TASA) Algo- rithm Input:Slides 푆=푆 1 ,푆 2 , . . .,푆 푛 , scripts 푆 푐 =푆 푐1 ,푆 푐2 , . . .,푆 푐푛 , word-level speech timestamps푇 푑 , learner profile푈 Parameters: TasaAgent, regions 푅, labels Pedagogical, Personalized, Salient, Adaptive, Assessment; RN and HW action types, salience dataH , dynamic memory 푀 푑 Output:퐴푆 seq 1: Initialize 푆 , 푆 푐 ,푇 푑 , and푈 2: for each slide 푆 푛 in 푆 do 3:Parse slide contents and identify regions 푅 in 푆 푛 4:Analyze script for current slide 푆 푛 5: for each region 푅 푛 in 푅 do 6:퐿 푛 ← assign segment label 퐿 to region 푅 푛 7:푆 푐푛 ← add appropriate speech segment 8:푠푒푔푚푒푛푡 푛 ← write segment data to 푀 푑 9: end for 10: for each 푠푒푔푚푒푛푡 푛 in slide 푆 푛 do 11:Analyze 푠푒푔푚푒푛푡 푛 12:푎 푛 ← assign suitable action (RN or HW) 13:푟 푛 ← give rationale for action 14: H ← write salience heuristic data to 푀 푑 15: end for 16: T ← save segmentation and analysis data to 푀 퐿 17: end for 18: TasaAgent utilizesT to generate 퐴푆 seq 19: return 퐴푆 seq 4 Experiments We conducted extensive quantitative and qualitative evaluations of Lect ̄ uraAgents through diverse experiments, assessing the frame- work’s performance on the following pedagogical metrics: lecture content quality, teaching quality, assessment, and personalization. 8 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Our main goal is to provide answers to two fundamental research questions: 1.RQ1: How does leveraging an adaptive hierarchical multi- agent architecture create high-quality personalized lecture contents that align with diverse learning profiles? 2.RQ2: How can an embodied tutor agent utilize generated materials to execute coherent, visual, and pedagogically informed teaching actions in a teaching environment (e.g., lecture slides presentation)? 4.1 Experiment Setup The experiments were designed to assess the framework from end- to-end, evaluating both personalized lecture generation and em- bodied teaching capabilities. We start by performing pedagogical evaluation on 280 personalized lectures generated using the frame- work under the seven frontier models reported in Table 4. For each model, we generate 40 lectures, with 10 lectures for each academic level, using the same prompts, learner profiles, and text-to-speech model (Kokoro TTS [62]) to ensure a fair comparison. Details on these lectures can be found in Appendix A.2.2. We worked with five expert educators, including subject teachers and university instructors with experience in curriculum design and instructional assessment, to define pedagogical rubrics grounded in recognized instructional quality standards [66], as summarized in Table 3. Ad- ditional details on the recruitment of these experts can be found in Appendix A.2.4. We then adopted the evaluation method in Tutor- Bench [67], with scoring primarily done by the expert educators in order to avoid induced bias from an LLM judge. Thus, for a j-th lecture, the framework’s overall performance score for each session under a given model or baseline framework, is computed as the weighted average of all passed rubric criteria 퐴푅 푗 푤 , given by: 퐴푅 푗 푤 = Í 푁 푗 푖=1 푤 푗 푖 · 1 푟 푗 푖 Í 푁 푗 푖=1 푤 푗 푖 · 1 푤 푗 푖 >0 (4) where푁 푗 is the number of rubric criteria for the j-th lecture,푤 푗 푖 ∈ -5, -3, -1, 0, +1, +3, +5, is the weight assigned to the i-th criterion, and푟 푗 푖 ∈ 0,1 indicates whether criterion i is satisfied. When a criterion is satisfied푟 푗 푖 =1, it contributes a positive reward of +5, +3, or +1, corresponding to it being a highly desirable, desirable and important, or nice-to-have behavior, respectively. When a criterion is not satisfied푟 푗 푖 =0, it is explicitly treated as a failure state and contributes a non-positive score, spanning a 0, -1, -3, and -5 range: 0 denotes the lowest-severity failure (no credit), -1 is used for a minor failure, -3 for a moderate failure, and -5 represents a critical failure (highly undesirable behavior). 4.1.1 Lecture Generation Evaluation Here, we evaluate Lect ̄ uraAgents as a personalized lecture con- tent generator. For each model, we generated 40 personalized lec- tures covering math, science, engineering, art, and history, with 10 lectures each for high school, undergraduate, master’s, and PhD learning profiles. Topics were randomly selected with emphasis on science subjects. Each lecture targeted one individual learner profile, covering learners aged 13–35, with profiles varying by academic level, prior knowledge, learning goals, learning style, and expected Table 3: Evaluation metrics and their respective rubrics Lecture Generation Evaluation MetricRubrics Lecture Content Quality (LCQ) Accuracy, Clarity, Coherence, Cognitive Load, Syllabus Coverage, Instruction-following Personalization Quality (PQ) Adaptive Emphasis, Preference Alignment, Engagement, Motivation, Tone/Style Assessment Quality (AQ) Concept Coverage, Cognitive Appropriateness, Answer Validity; Rationale Lecture Delivery Teaching Action Quality (TAQ) Temporal Alignment, Accurate Handwriting Action, Accurate Rough Notation Action, Spatial Accuracy, Active Learning, Embodied Teaching difficulty. The resulting output after generation contains the fol- lowing lecture artifacts: a detailed lecture plan, a research report, syllabus, learner profile, 15 slides with images, per-slide scripts, lec- ture speeches, personalized lecture notes and study guide, teaching actions, teaching action–speech alignment, and assessments. Evaluation Metrics. Using expert-defined rubrics detailed in Ta- ble 3, we assess the framework’s personalized lecture content gen- eration capability across three main evaluation metrics: Lecture Content Quality (LCQ), Personalization Quality (PQ), and Assess- ment Quality (AQ). LCQ measures accuracy, clarity, coherence, cognitive load, and instruction-following rubric dimensions. PQ evaluates adaptation to a learner profile (adaptive emphasis) and learning preferences (preference alignment), engagement, motiva- tion, and instructor’s tone or style. AQ measures concept coverage, cognitive appropriateness, answer accuracies, and rationale quality. Each lecture’s metric score is computed using the weighted aver- age of all passed rubrics and then averaged across all 40 lectures generated under each model. 4.1.2 Lecture Delivery Evaluation Next, we evaluate the embodied and multimodal teaching capability of the framework. For each generated lecture, the ProfessorAgent is tasked with teaching all 15 slides using lecture artifacts created in the lecture generation session. This stage evaluates the agent’s teaching action quality, independent of content generation, allow- ing us to assess multimodal alignment and embodied instructional delivery capabilities specifically. Evaluation Metrics. Lecture delivery is evaluated using the Teach- ing Action Quality (TAQ) metric, which has six rubric dimensions (detailed in Table 3). These include temporal and spatial alignment of teaching actions, accurate handwriting and rough notation ac- tions, active learning, and overall embodied teaching experience. TAQ assesses how well each model exploits the frameworks archi- tecture to deliver accurate, coherent, and pedagogically informed teaching action sequences. For each slide, script, and teaching action sequence, an expert educator judges whether each rubric criterion is satisfied, and the overall average TAQ score is computed using Equation 4. 9 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Table 4: (RQ 1) Evaluation of Lect ̄ uraAgents across pedagogical metrics under frontier models Rank ModelLCQ (%) PQ (%) AQ (%) TAQ (%) AAR (%) 1Gemini 3 Pro80.283.381.676.580.4 2GPT-5.176.180.582.376.278.8 3Claude 4.5 Sonnet72.478.676.280.476.9 4Gemini 2.5 Pro70.575.280.172.374.5 5DeepSeek V3.268.973.175.277.873.5 6GPT-4o67.571.472.873.271.2 7Qwen 3 Omni65.470.356.564.364.1 Figure 9: (RQ1 and RQ2) Results across rubric dimensions for each evaluation metric under each frontier model. Results. TAQ results indicate that Lect ̄ uraAgents enables gener- ally accurate and coherent teaching action sequences across mod- els. As shown in Figure 9, models perform strongly on spatially grounded criteria, particularly spatial accuracy, handwriting ac- tions, rough notation actions, and embodied teaching. This sug- gests that the framework can reliably convert generated lecture materials into visible instructional actions. Figure 10 further shows that teaching-action-related scores are distributed across multiple lecture artifacts, indicating that embodied delivery is maintained across the broader lecture package rather than appearing only in isolated outputs. A key factor behind this stability is the TASA mod- ule, which provides the ProfessorAgent with a structured view of slide regions and aligns teaching actions with corresponding speech segments. While temporal alignment remains comparatively more 10 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Figure 10: (RQ2) Average distribution of Personalization Quality and Teaching Action Quality across diverse learning profiles at various academic levels. Figure 11: (RQ2) Overall average distribution of Lecture Content Quality scores across generated Lecture Materials from all models. variable due to the difficulty of fine-grained action–speech synchro- nization, Figure 11 shows that TAQ and personalization-related performance remain broadly stable across all learner profiles. This suggests that the embodied teaching mechanism generalizes across academic levels, while timing-sensitive action selection remains an area for addiitonal improvement. 4.1.3 Comparative Evaluation with Related Frameworks We further assess Lect ̄ uraAgents against existing frameworks in this domain. Due to varying capabilities between baselines, we only compare performances on shared capabilities to ensure fairness. We identify two closely related open-source frameworks and one learning system with publicly available lecture data: Instructional Agents [34], GenMentor [36], and Google’s Learn Your Way sys- tem [3]. Our comparative evaluation assesses each framework or system based on lecture content quality (LCQ), assessment quality (AQ), and personalization (PQ) evaluation metrics, using the same evaluation method described in Section 4.1. For InstructionalAgents and GenMentor, we generated 20 lectures using their publicly re- leased implementations. For Learn Your Way, we used the publicly available lectures provided on its website. Additional details about the lecture set and selection process are provided in Appendix B. We then generated the same lectures with Lect ̄ uraAgents using identical topics, prompts, and learner profiles, and evaluated all outputs using the methodology described in Section 4.1. Results. As shown in Table 5, Lect ̄ uraAgents obtains higher scores than the baseline systems across LCQ, PQ, and AQ. The most notable difference is observed in personalization quality, indicating that the framework is better able to adapt generated materials to learner profiles. Its performance in lecture content and assessment quality further suggests that the framework supports not only learner- specific adaptation, but also coherent instructional organization and alignment between lecture materials and assessment tasks. 4.1.4 Efficacy Study with Students The preceding evaluations assessed the pedagogical capabilities of the framework across multiple topics, models, and personalization Table 5: Performance comparison of Lect ̄ uraAgents with existing related frameworks N (number of lectures) = 20 Framework / System LCQ (%) PQ (%) AQ (%) Overall (%) Instructional Agents [34] 52.153.251.452.2 GenMentor [36]50.864.646.654.0 Learn Your Way [3] 58.960.162.560.5 Lect ̄ uraAgents 70.373.571.271.6 settings. However, the impact of the framework is better exam- ined when these capabilities are tested on real learners. Therefore, we conducted a small-scale efficacy study with real students to measure the holistic pedagogical value of Lect ̄ uraAgents in terms of learning support and learner experience. To provide a broader comparison, we included both Learn Your Way, representing a modern AI-assisted learning system, and Adobe Acrobat Reader v23.008.20555, representing a widely used traditional digital study Figure 12: Average scores from immediate assessment on topics learned using Lect ̄ uraAgents, Learn Your Way, and Adobe reader. 11 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Table 6: Student responses to a survey given after assessment To what extent do you agree or disagree with the following statements? % somewhat agree or strongly agree Lect ̄ uraAgents N = 15 Learn Your Way N = 15 Adobe Reader N=15 I felt adequately prepared to complete the assessment after using today’s educational tool. 95%80%72% I felt like today’s educational tool helped me gain a good understanding of the topic. 100%92%65% I would like to use today’s educational tool to support my learning needs in the future. 87%73%63% The educational tool I used today would make me more effective at learning compared to other educational tools I currently use at home or in school. 84%67%44% reading software without generative AI capabilities. The study in- volved 45 students divided equally across the three learning systems, with 15 participants per system. Each group comprised five stu- dents from each educational level—high school, undergraduate, and master’s—with ages ranging between 15 to 25 years. Students were recruited through a short pre-study topic-familiarity screening and provided informed consent prior to participation. Result. Figure 12 compares students’ post-learning assess- ment performance across learning systems. The results show that Lect ̄ uraAgents achieved the strongest performance across all learner groups, followed by Learn Your Way and Adobe Reader. Al- though the improvement is modest, its consistency suggests that the framework’s personalized and embodied teaching capabilities sup- ported better short-term comprehension and content recall, rather than merely improving students’ subjective learning experience. Consistent with this pattern, Table 6 shows that students using Lect ̄ uraAgents reported stronger perceived content understanding, assessment readiness, future learning support, and overall learning experience than those using Learn Your Way or Adobe Reader. 5 Limitations and Future Work We acknowledge several limitations that may inform future work. First, while Lect ̄ uraAgents performs well on lecture content gener- ation and embodied delivery, the current teaching action–speech alignment module relies heavily on offline heuristics with a limited set of supported teaching actions. This may constrain the richness of embodied instruction and robustness across diverse slide layouts. Second, the multi-agent orchestration can introduce latency and compute overhead. Finally, the framework can sometimes inherit common LLM failure modes such as factual errors, inconsistent reasoning, and tool or prompt-sensitivity. Future work will (1) ex- pand the teaching action taxonomy and improve action fidelity; (2) transition from heuristic action–speech alignment to learned poli- cies (e.g., training policies in a presentation slide environment with preference optimization or reinforcement learning); (3) strengthen grounding to reduce hallucinations; and (4) optimize orchestration for efficiency while preserving pedagogical coherence and control- ling compute costs. 6 Conclusion In this paper we introduced Lect ̄ uraAgents, a hierarchical multi- agent framework for end-to-end adaptive, personalized AI-assisted learning experiences. The framework addresses two major issues in personalized AI-assisted learning: (1) How can AI adaptively personalize instructional contents to best meet the needs of diverse learners? (2) How can such instructional contents be delivered in embodied and pedagogically meaningful ways to ensure better learning outcomes? In order to effectively address these issues, Lect ̄ uraAgents is first modelled on a professor-student relationship framing, wherein a ProfessorAgent leads a collaborative class of specialized subordinate agents through research, planning, evalu- ation, and embodied delivery of instructional contents that adapt to diverse students. The framework’s personalized and embodied capabilities (e.g., TASA algorithm) offer students enhanced learning and study experiences. We evaluated Lect ̄ uraAgents through two main experiments: a pedagogical evaluation under frontier models across high school, undergraduate, and graduate-level topics, and an efficacy study with real students. Experimental results show substantial improvements over baseline frameworks in lecture con- tent quality, personalization, assessment quality, and embodied teaching performance. In addition, these findings are validated by results from our efficacy study with students, which provide prelim- inary evidence that the framework can improve learning outcomes while enhancing learner experience. In conclusion, we position Lect ̄ uraAgents offers as a pedagogically grounded framework for personalized AI-assisted learning at scale. References [1]Ming Yang and Feng Wen. 2023. AI-powered personalized learning journeys: Revolutionizing information management for college students in online platforms. J. Inf. Syst. Eng. Manag. 8, 1 (2023). Article 23196. doi:10.55267/iadt.07.14079 [2]Hadi Farhood, Matthew Nyden, Amir Beheshti, and Stephan Müller. 2025. Artificial intelligence-based personalized learn- ing in education: A systematic literature review. Discover Artif. Intell. 5 (2025). Article 331. doi:10.1007/s44163-025-00598-x [3]LearnLM Team, Alicia Martín, Amir Globerson, Amy Wang, Anirudh Shekhawat, Anna Iurchenko, Anisha Choudhury, Avinatan Hassidim, Ayça Çakmakli, Ayelet Shasha Evron, 12 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Charlie Yang, Courtney Heldreth, Diana Akrong, Gal Eli- dan, Hairong Mu, Ian Li, Ido Cohen, Katherine Chou, Komal Singh, Lev Borovoi, Lidan Hackmon, Lior Belinsky, Michael Fink, Niv Efron, Preeti Singh, Rena Levitt, Shashank Agarwal, Shay Sharon, Tracey Lee-Joe, Xiaohong Hao, Yael Gold-Zamir, Yael Haramaty, Yishay Mor, Yoav Bar Sinai, and Yossi Ma- tias. 2025. Towards an AI-augmented textbook. Preprint at https://arxiv.org/abs/2509.13348. arXiv:2509.13348 doi:10. 48550/arXiv.2509.13348 [4]Jian Peng and Yu Li. 2025. Frontiers of artificial intelligence for personalized learning in higher education: A systematic review of leading articles. Appl. Sci. 15, 18 (2025). Article 10096. doi:10.3390/app151810096 [5] Meghan Novack and Susan Goldin-Meadow. 2015. Learning from gesture: How our hands change our minds. Educ. Psychol. Rev. 27, 3 (2015), 405–412. doi:10.1007/s10648-015-9325-3 [6]Hanne Marie Hegna and Trine Ørbæk. 2024. Traces of em- bodied teaching and learning: A review of empirical studies in higher education. Teach. High. Educ. 29, 2 (2024), 420–444. doi:10.1080/13562517.2021.1989582 [7]Mina C. Johnson-Glenberg and Cathy Megowan-Romanowicz. 2017. Embodied science and mixed reality: How gesture and motion capture affect physics education. Cogn. Res.: Princ. Implic. 2 (2017). Article 24. doi:10.1186/s41235-017-0060-9 [8]Shuo Wang, Tianyi Xu, Haoliang Li, Chen Zhang, Jian Liang, Jian Tang, Philip S. Yu, and Qingsong Wen. 2024. Large lan- guage models for education: A survey and outlook. Preprint at https://arxiv.org/abs/2403.18105. arXiv:2403.18105 doi:10. 48550/arXiv.2403.18105 [9]Lijia Chen, Peng Chen, and Zhijun Lin. 2020. Artificial intel- ligence in education: A review. IEEE Access 8 (2020), 75264– 75278. doi:10.1109/ACCESS.2020.2988510 [10]Shubham Sharma, Piyush Mittal, Mohit Kumar, and Vineet Bhardwaj. 2025. The role of large language models in person- alized learning: A systematic review of educational impact. Discover Sustain. 6 (2025). Article 243. doi:10.1007/s43621-025- 01094-z [11]Emma Chen, Jae-Eun Lee, Jialin Lin, and Kenneth Koedinger. 2024. GPTutor: Great personalized tutor with large language models for personalized learning content generation. In Pro- ceedings of the 11th ACM Conference on Learning @ Scale (L@S ’24). ACM, Atlanta, GA, USA. doi:10.1145/3657604.3664718 [12]Songlin Xu, Xinyu Zhang, and Lianhui Qin. 2024. EduA- gent: Generative student agents in learning. Preprint at https://arxiv.org/abs/2404.07963.arXiv:2404.07963 doi:10. 48550/arXiv.2404.07963 [13]Tianfu Wang, Yi Zhan, Jianxun Lian, Zhengyu Hu, Nicholas Jing Yuan, Qi Zhang, Xing Xie, and Hui Xiong. 2025. LLM-powered multi-agent framework for goal-oriented learn- ing in intelligent tutoring systems. In Companion Proceedings of the ACM Web Conference 2025. ACM, New York, NY, USA, 510–519. doi:10.1145/3701716.3715244 [14]En Sun and Li Tai. 2025. MultiTutor: Collaborative LLM agents for multimodal student support. In Proceedings of the Innovation and Responsibility in AI-Supported Education Work- shop (Proc. Mach. Learn. Res., Vol. 273). 174–190. Available at https://proceedings.mlr.press/v273/sun25a.html.https: //proceedings.mlr.press/v273/sun25a.html [15]Michael Vaccaro, Matthew Friday, and Andrea Zaghi. 2025. Multi-agentic LLMs for personalizing STEM texts. Appl. Sci. 15 (2025). Article 7579. doi:10.3390/app15137579 [16] Qiang Yang, Yifan Yang, Shiqi An, Tianhao Hao, and Guang Xu. 2025. LLM-based collaborative agents with pedagogy- guided interaction modeling for timely instructive feedback generation. In Proceedings of the 34th International Joint Conference on Artificial Intelligence (IJCAI ’25). 9972–9980. doi:10.24963/ijcai.2025/1108 [17] Yifan Ma, Shuo Hu, Zhu Bo, Yifan Wang, Yuxuan Kang, Shiyu Liu, and Kang Hao Cheong. 2025. EduVerse: A user-defined multi-agent simulation space for education scenarios. Preprint at https://arxiv.org/abs/2510.05650. arXiv:2510.05650 doi:10. 48550/arXiv.2510.05650 [18] Meng Yue, Weinan Lyu, Walid Mifdal, Jaemin Suh, Yiming Zhang, and Ziyu Yao. 2024. MathVC: An LLM-simulated multi-character virtual classroom for mathematics education. Preprint at https://arxiv.org/abs/2404.06711. arXiv:2404.06711 doi:10.48550/arXiv.2404.06711 [19]Hao Li, Jifan Yu, Xinyu Cong, Yifan Dang, Daniel Zhang-Li, Lin Mi, Yi Zhan, Huiqin Liu, and Zhiyuan Liu. 2025. Which type of students can LLMs act? Investigating authentic sim- ulation with graph-based human–AI collaborative systems. Preprint at https://arxiv.org/abs/2502.11678. arXiv:2502.11678 doi:10.48550/arXiv.2502.11678 [20]Yao Shi, Rongkeng Liang, and Yong Xu. 2025. EducationQ: Evaluating LLMs’ teaching capabilities through multi-agent dialogue frameworks. In Proceedings of the 63rd Annual Meet- ing of the Association for Computational Linguistics (ACL ’25). 32799–32828. doi:10.18653/v1/2025.acl-long.1576 [21]Haein Jin, Minji Yoo, Jihye Park, Yejin Lee, Xin Wang, and Joonhwan Kim. 2025. TeachTune: Reviewing pedagogical agents against diverse student profiles with simulated stu- dents. In Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI ’25). ACM, New York, NY, USA. Article 1073. doi:10.1145/3706598.3714054 [22]Rui Jia, Mingyuan Zhang, Feng Liu, Bin Jiang, Kun Kuang, and Zhiwei Dai. 2025. EduAgentQG: A multi-agent workflow framework for personalized question generation. Preprint at https://arxiv.org/abs/2511.11635. arXiv:2511.11635 doi:10. 48550/arXiv.2511.11635 [23]Jiahuan Pei, Fanghua Ye, Xin Sun, Wentao Deng, Koen Hindriks, and Junxiao Wang. 2025. Conversational educa- tion at scale: A multi-LLM agent workflow for procedural learning and pedagogic quality assessment.Preprint at https://arxiv.org/abs/2507.05528.arXiv:2507.05528 doi:10. 48550/arXiv.2507.05528 [24] Jing Liu, Zhi Huang, Tian Xiao, Jun Sha, Jiarui Wu, Qi Liu, Shuo Wang, and Enhong Chen. 2024. SocraticLM: Exploring Socratic personalized teaching with large language models. In Advances in Neural Information Processing Systems, Vol. 37. [25]Sayan P. Chowdhury, Vojtěch Zouhar, and Mrinmaya Sachan. 2024. AutoTutor meets large language models: A language model tutor with rich pedagogy and guardrails. In Proceedings of the 11th ACM Conference on Learning @ Scale (L@S ’24). ACM, New York, NY, USA, 5–15. doi:10.1145/3657604.3662041 13 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching [26]Xuejiang Wang, Chee-Poh Lee, and Bilge Mutlu. 2025. Learn- Mate: Enhancing online education with LLM-powered per- sonalized learning plans and support. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA ’25). ACM, New York, NY, USA. doi:10.1145/3706599. 3719857 [27] Jana Gonnermann-Müller, Jennifer Haase, Konstantin Fackeldey, and Sebastian Pokutta. 2025.FACET: Teacher-centredLLM-basedmulti-agentsystems— Towards personalized educational worksheets.Preprint at https://arxiv.org/abs/2508.11401.arXiv:2508.11401 doi:10.48550/arXiv.2508.11401 [28]Kia Karbasi, Kevin Hong, Mohammad Amin Samadi, and Gre- gory Pottie. 2025. Multi-agent collaborative framework for math problem generation. In Proceedings of the 18th Inter- national Conference on Educational Data Mining. 613–618. doi:10.5281/zenodo.15870246 [29] Songlin Xu, Xinyu Zhang, and Lianhui Qin. 2024. EduA- gent: Generative student agents in learning. Preprint at https://arxiv.org/abs/2404.07963.arXiv:2404.07963 doi:10. 48550/arXiv.2404.07963 [30]Weibo Gao, Qi Liu, Linan Yue, Fangzhou Yao, Rui Lv, Zheng Zhang, Hao Wang, and Zhenya Huang. 2025. Agent4Edu: Generating learner response data by generative agents for intelligent education systems. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 23923–23932. doi:10.1609/aaai.v39i22.34565 [31]Yao Shi, Rongkeng Liang, and Yong Xu. 2025. EducationQ: Evaluating LLMs’ teaching capabilities through multi-agent dialogue framework. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. 32799–32828. doi:10.18653/v1/2025.acl-long.1576 [32]Jana Gonnermann-Müller, Jennifer Haase, Konstantin Fackeldey, and Sebastian Pokutta. 2025.FACET: Teacher-centredLLM-basedmulti-agentsystems— Towards personalized educational worksheets.Preprint at https://arxiv.org/abs/2508.11401.arXiv:2508.11401 doi:10.48550/arXiv.2508.11401 [33]Xian Peng, Pan Yuan, Dong Li, Junlong Cheng, Qin Fang, and Zhi Liu. 2025. KELE: A multi-agent framework for structured Socratic teaching with large language models. In Findings of the Association for Computational Linguistics: EMNLP 2025. Association for Computational Linguistics, Suzhou, China, 16342–16362. doi:10.18653/v1/2025.findings-emnlp.888 [34]Hao Yao, Wenjie Xu, Jacob Turnau, Nathan Kellam, and Hui Wei. 2025. Instructional agents: LLM agents on automated course material generation for teaching faculties. Preprint at https://arxiv.org/abs/2508.19611. arXiv:2508.19611 doi:10. 48550/arXiv.2508.19611 [35]Xueqiao Zhang, Chao Zhang, Jianwen Sun, Jun Xiao, Yi Yang, and Yawei Luo. 2025. EduPlanner: LLM-based multi-agent systems for customized and intelligent instructional design. IEEE Trans. Learn. Technol. 18 (2025), 416–427. doi:10.1109/ TLT.2025.3561332 [36]Tianfu Wang, Yi Zhan, Jianxun Lian, Zhengyu Hu, Nicholas Jing Yuan, Qi Zhang, Xing Xie, and Hui Xiong. 2025. LLM-powered multi-agent framework for goal-oriented learn- ing in intelligent tutoring systems. In Companion Proceedings of the ACM Web Conference 2025. ACM, New York, NY, USA, 510–519. doi:10.1145/3701716.3715244 [37] Zheyuan Zhang, Daniel Zhang-Li, Jifan Yu, Linlu Gong, Jin- chang Zhou, Zhanxin Hao, Jianxiao Jiang, Jie Cao, Huiqin Liu, Zhiyuan Liu, Lei Hou, and Juanzi Li. 2025. Simulating classroom education with LLM-empowered agents. In Pro- ceedings of NAACL-HLT 2025. 10364–10379. doi:10.18653/v1/ 2025.naacl-long.520 [38] Jiahuan Pei, Fanghua Ye, Xin Sun, Wentao Deng, Koen Hindriks, and Junxiao Wang. 2025. Conversational educa- tion at scale: A multi-LLM agent workflow for procedural learning and pedagogic quality assessment.Preprint at https://arxiv.org/abs/2507.05528.arXiv:2507.05528 doi:10. 48550/arXiv.2507.05528 [39]Richard C. Atkinson and Richard M. Shiffrin. 1968. Human memory: A proposed system and its control processes. In Psychology of Learning and Motivation. Vol. 2. Academic Press, New York, 89–195. doi:10.1016/S0079-7421(08)60422-3 [40]Nelson Cowan. 2008. What are the differences between long- term, short-term, and working memory? In Prog. Brain Res. Vol. 169. 323–338. doi:10.1016/S0079-6123(07)00020-9 [41]Yu Su, Qi Liu, Qi Liu, Zhenya Huang, Yuting Yin, Enhong Chen, Chao Ding, Shuwei Wei, and Guoping Hu. 2018. Exercise- enhanced sequential modeling for student performance pre- diction. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 32. doi:10.1609/aaai.v32i1.11864 [42]Qi Liu, Shiwei Tong, Chuanren Liu, Hongke Zhao, Enhong Chen, Haiping Ma, and Shijin Wang. 2019. Exploiting cog- nitive structure for adaptive learning. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Dis- covery & Data Mining. ACM, New York, NY, USA, 627–635. doi:10.1145/3292500.3330922 [43]Zhenya Huang, Qi Liu, Yuying Chen, Le Wu, Keli Xiao, En- hong Chen, Haiping Ma, and Guoping Hu. 2020. Learning or forgetting? A dynamic approach for tracking the knowledge proficiency of students. ACM Trans. Inf. Syst. 38, 2 (2020). Article 19. doi:10.1145/3379507 [44]Ilie Gligorea, Marius Cioca, Romana Oancea, Andra-Teodora Gorski, Hortensia Gorski, and Paul Tudorache. 2023. Adaptive learning using artificial intelligence in e-learning: A literature review. Educ. Sci. 13, 12 (2023). Article 1216. doi:10.3390/ educsci13121216 [45]Le Ying Tan, Shiyu Hu, Darren J. Yeo, and Kang Hao Cheong. 2025. Artificial intelligence-enabled adaptive learning plat- forms: A review. Comput. Educ.: Artif. Intell. 9 (2025). Article 100429. doi:10.1016/j.caeai.2025.100429 [46] Graham Hardaker and Louise E. Glenn. 2025.Artificial intelligence for personalized learning: A systematic litera- ture review. Int. J. Inf. Learn. Technol. 42, 1 (2025), 1–14. doi:10.1108/IJILT-07-2024-0160 [47]Raffaella Folgieri, Marisa Gil, Miriam Bait, and Claudio Luc- chiari. 2024. AI-powered personalised learning platforms for EFL learning: Preliminary results. In Proceedings of the 16th International Conference on Computer Supported Education, Vol. 2. 255–261. doi:10.5220/0012672000003693 14 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching [48]María Fernanda Contrino, María Reyes-Millán, Paola Vázquez- Villegas, and José Membrillo-Hernández. 2024. Using an adap- tive learning tool to improve student performance and satis- faction in online and face-to-face education for a more per- sonalized approach. Smart Learn. Environ. 11 (2024). Article 6. doi:10.1186/s40561-024-00292-y [49] Yajun Chen. 2025. Evaluation of the impact of AI-driven personalized learning platform on medical students’ learning performance. Front. Med. 12 (2025). Article 1610012. doi:10. 3389/fmed.2025.1610012 [50] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023.Tool- former: Language models can teach themselves to use tools. In Advances in Neural Information Processing Sys- tems.https://proceedings.neurips.c/paper_files/paper/ 2023/hash/d842425e4bf79ba039352da0f658a906-Abstract- Conference.html [51] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing reasoning and acting in language models. In Proceedings of the International Conference on Learning Representations. https: //openreview.net/forum?id=WE_vluYUL-X [52]Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Re- flexion: Language agents with verbal reinforcement learning. Preprint at https://arxiv.org/abs/2303.11366. arXiv:2303.11366 doi:10.48550/arXiv.2303.11366 [53]Guangyao Chen, Siwei Dong, Yu Shu, Ge Zhang, Jaward Sesay, Börje F. Karlsson, Jie Fu, and Yemin Shi. 2024. AutoAgents: A Framework for Automatic Agent Generation. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24. 22–30. doi:10.24963/ijcai.2024/3 [54]Zhendong Chu, Shen Wang, Jian Xie, Tinghui Zhu, Yibo Yan, Jingheng Ye, Aoxiao Zhong, Xuming Hu, Jing Liang, Philip S. Yu, and Qingsong Wen. 2025. LLM Agents for Edu- cation: Advances and Applications. In Findings of the Associa- tion for Computational Linguistics: EMNLP 2025. Association for Computational Linguistics, Suzhou, China, 13782–13810. doi:10.18653/v1/2025.findings-emnlp.743 [55]Feifei Wang, Wenqing Li, Richard E. Mayer, and Hui Liu. 2018.Animated pedagogical agents as aids in multime- dia learning: Effects on eye-fixations during learning and learning outcomes. J. Educ. Psychol. 110, 2 (2018), 250–268. doi:10.1037/edu0000221 [56]Wenqing Li, Feifei Wang, Richard E. Mayer, and Hui Liu. 2019. Getting the point: Which kinds of gestures by pedagogical agents improve multimedia learning? J. Educ. Psychol. 111, 8 (2019), 1382–1395. doi:10.1037/edu0000352 [57]Arthur C. Graesser, Katja Wiemer-Hastings, Peter M. Wiemer- Hastings, and Roger J. Kreuz. 1999. AutoTutor: A simulation of a human tutor. Cogn. Syst. Res. 1 (1999), 35–51. doi:10.1016/ S1389-0417(99)00005-4 [58]Benjamin D. Nye, Arthur C. Graesser, and Xiangen Hu. 2014. AutoTutor and family: A review of 17 years of natural language tutoring. Int. J. Artif. Intell. Educ. 24 (2014), 427–469. doi:10. 1007/s40593-014-0029-5 [59]Tao Xu, Yuan Liu, Yaru Jin, Yueyao Qu, Jie Bai, Wenlan Zhang, and Yun Zhou. 2025. From recorded to AI-generated in- structional videos: A comparison of learning performance and experience. Br. J. Educ. Technol. 56, 4 (2025), 1463–1487. doi:10.1111/bjet.13530 [60]Alexander Holmberg. 2025. Generating narrated lecture videos from slides with synchronized highlights. Preprint at https://arxiv.org/abs/2505.02966. arXiv:2505.02966 doi:10. 48550/arXiv.2505.02966 [61]Tushar Aggarwal and Aarohi Bhand. 2025. PASS: Presenta- tion automation for slide generation and speech. Preprint at https://arxiv.org/abs/2501.06497. arXiv:2501.06497 doi:10. 48550/arXiv.2501.06497 [62] hexgrad. 2025. Kokoro-82M. Hugging Face model repository. Open-weight text-to-speech model. https://huggingface.co/ hexgrad/Kokoro-82M [63]Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Chris- tine McLeavey, and Ilya Sutskever. 2023. Robust speech recog- nition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning (Proc. Mach. Learn. Res., Vol. 202). 28492–28518. arXiv:2212.04356 doi:10.48550/arXiv.2212.04356 [64]Rough Notation. 2020. Rough notation library. GitHub repository at https://github.com/rough-stuff/rough-notation. https://github.com/rough-stuff/rough-notation [65] Alex Graves. 2013. Generating sequences with recurrent neural networks. Preprint at https://arxiv.org/abs/1308.0850. arXiv:1308.0850 doi:10.48550/arXiv.1308.0850 [66]Mariska H. Knol, Conor V. Dolan, Gideon J. Mellenbergh, and Han L. J. van der Maas. 2016. Measuring the quality of university lectures: Development and validation of the Instructional Skills Questionnaire (ISQ). PLOS ONE 11, 2 (2016). e0149163. doi:10.1371/journal.pone.0149163 [67]Rakshith S. Srinivasa, Zora Che, Chen Bo Calvin Zhang, Diego Mares, Ernesto Hernandez, Jayeon Park, Dean Lee, Guillermo Mangialardi, Charmaine Ng, Ed-Yeremai Hernan- dez Cardona, Anisha Gunjal, Yunzhong He, Bing Liu, and Chen Xing. 2025. TutorBench: A benchmark to assess tu- toring capabilities of large language models. Preprint at https://arxiv.org/abs/2510.02663.arXiv:2510.02663 doi:10. 48550/arXiv.2510.02663 15 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Appendix A A.1 Lect ̄ uraAgents: Detailed Architecture A.1.1 Core Modules and Components The framework is organized into four core modules, each serving a distinct purpose in the lecture generation and lecture delivery stages. These modules provide the infrastructure for agent coordination, LLM integration, teaching action alignment, memory management, and content rendering. The modular design enables easy extension and maintenance of individual components. Table A1: Lect ̄ uraAgents’ Core Modules and Components ModuleLocationFunctionKey Classes AgentsLectura/LecturaAgents/module/agentsThe Agents module implements the core agent architecture with base interfaces, role definitions (Coordinator, Executor, Validator), and state management. This module provides the hierarchical three-tier agent system with collaboration mechanisms (sequential and parallel) and orchestration through SwarmOfRanks. It handles agent lifecycle, task execution, validation, and inter-agent communication. Agent (base class), ProfessorAgent, LecturePlanner, ResearchAgent, SlideAgent, ScriptAgent, SpeechAgent, TasaAgent LLMsLectura/LecturaAgents/module/llmsProvides unified abstraction layer for multiple LLM providers (OpenAI, Google, Anthropic, DeepSeek, Qwen, Local) enabling seamless model switching. Handles authentication, API communication, response formatting, function calling, and streaming. It abstracts provider-specific differences to provide consistent interface for all agents. LLMProvider (base class), OpenAIProvider, GoogleAIProvider, AnthropicProvider, DeepSeekProvider, QwenProvider, LocalLLMProvider TASALectura/LecturaAgents/tasaImplements Teaching Action Salience Analysis (TASA) for generating and aligning synchronized teaching actions (rough notation, handwriting) with speech. This module processes speech scripts to embed action markers, extracts word-level timestamps from audio using Whisper ASR [63], and creates temporal alignment between visual actions and spoken content. TASA Adaptive MemoryLectura/LecturaAgents/memoryImplements a three-layer adaptive memory system: short-term memory for session context, long-term memory for persistent learner data, and dynamic memory for adaptive learning patterns. This module provides a unified AdaptiveMemory interface that enables agents to access learner context, preferences, and learning history for personalization. ShortTermMemory, LongTermMemory, DynamicMemory, AdaptiveMemory A.1.2 Agent Hierarchy and Roles Agents are organized by rank and responsibility. Rank 1 agents (ProfessorAgent) serve as coordinators and validators at the highest level, Rank 2 agents (LecturePlanner, ResearchAgent) coordinate execution and validate outputs, while Rank 3 agents (SlideAgent, ScriptAgent, SpeechAgent, TasaAgent) execute specific tasks. Each agent has clearly defined responsibilities and access to appropriate tools and actions for their role. 16 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Table A2: Agent Hierarchy and Roles RankAgentRoleResponsibilitiesTools / Actions 1ProfessorAgentCoordinator and Tutor• Initiates lecture sessions • Reviews and approves plans • Validates final artifacts • Reviews final lecture artifacts • Delivers embodied lectures research() create_syllabus() review_plan() instantiate_groupchat() create_lecture_notes() create_study_guide() create_assessments() create_personalization() review_artifacts() embodied_teaching() 2Lecture PlannerValidator• Creates lecture plans. • Manages and validates tasks done by subordinate executor agents. • Assembles generated lecture artifacts and submits to ProfessorAgent for final review. research() create_plan() validate_task() assemble_artifacts() 3Executor AgentsExecutors ResearchAgent: Conducts multi-turn web searches on lecture topic, writes a detailed research report and submits for review by LecturePlanner. web_search() SlideAgent: Generates personalized slide contents; Designs and build slides with structured content blocks based on learner’s preferences. slide_builder() research() file_parser() ScriptAgent: Creates engaging, personalized narration scripts that aligns with both slide contents and learner’s preferences. analyze_slide() write_script() SpeechAgent: Synthesizes and generates speech audio from scripts based on learner’s preferred instructor voice. Uses TTS/ASR tool to create word-level timestamps . Whisper [63], Kokoro TTS [62] create_timestamps() TasaAgent: Uses tools in TASA module to segment and annotate slide contents with heuristic based context for prospective action-speech sequences. It then processes speech timestamps and slide contents into synchronized embodied teaching action sequences with embeded action markers (highlight, underline, handwriting, etc.). TASA Module temporal_segmentation() heuristic_analysis() A.1.3 Agent States and Lifecycle Agents transition through a well-defined state machine during task execution. The lifecycle begins with the IDLE state, progresses through acknowledgment and execution phases, and concludes with completion, failure, or revision states. This state management ensures proper task tracking, error handling, and enables agents to revise their work based on feedback from higher-ranking agents. Table A3: Lect ̄ uraAgents’ States and Lifecycles StateDescriptionTransition IDLEAgent is waiting for a task.ACKNOWLEDGED 17 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching StateDescriptionTransition ACKNOWLEDGEDAgent has received and acknowledged given. EXECUTING EXECUTINGAgent is actively working on assigned task. COMPLETED, FAILED or REVISAL COMPLETEDAgent has completed task successfully. IDLE (for next task) FAILEDTask execution was unsuccessful.REVISAL REVISALAgent is revising work based on feedback from self-reflection or review. EXECUTING A.1.4 Multi-agent Collaboration Agents within the same rank can collaborate using two primary mechanisms: sequential collaboration for dependent tasks and parallel collaboration for independent tasks. The SwarmOfRanks mechanism enables hierarchical coordination across multiple ranks, allowing complex workflows where agents at different levels coordinate their activities. These collaboration patterns are essential for orchestrating the multi-stage lecture generation process. Table A4: Collaboration Mechanisms TypeDescriptionUse Case Sequential ColabAgents complete tasks one after another, sharing responses. When tasks depend on previous outputs. Parallel ColabAgents complete tasks simultaneously, while sharing responses. When tasks are independent. Swarm of RanksHierarchical coordination across ranks Multi-rank workflows A.1.5 Tools and Capabilities The framework provides a comprehensive set of tools that agents use to accomplish their tasks. These tools range from web search and file parsing to text-to-speech synthesis and code execution. Each tool is designed to be modular and reusable, with clear interfaces that agents can invoke during their execution. The tools abstract away complex operations like API interactions, file processing, and multimedia generation. Table A5: Tools and Capabilities ToolPurposeUsageDependencies Web SearchMulti-turn web research using SerpAPI. Used by ResearchAgent, ProfessorAgent, LecturePlanner and SlideAgent SerpAPI Slide WorldDynamic slide environment with canvas for teaching sessions. Used by ProfessorAgent for embodied lecture delivery. HTML/CSS/JS/Python Slide BuilderCustom slide design tool.Used by SlideAgent for building and rendering slides. HTML/CSS/JS/Python File ParserParses PDF, TXT, MD files.Used by ProfessorAgent and SlideAgent to extract content from additional materials. PyPDF2, python-docx Command lineFor command execution to create lecture artifacts. Used by all agents to read/write/edit/save/delete files. Bash/Zsh TASA Segmentor / Aligner Segments, annotates and aligns slide contents with speech timestamps for synchronized Used by ProfessorAgent and TasaAgentTASA Module ResearchA unified research tool that makes use of web search plus an LLM to perform deep research on topics. Used by ResearchAgent, ProfessorAgent, LecturePlanner and SlideAgent SerpAPI + Underlying LLM Whisper [63]Extracts word-level timestamps from audio. Used by TasaAgent for action alignment.Whisper ASR model 18 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching ToolPurposeUsageDependencies Kokoro TTS [62]Generate speeches from scripts with desired instructor voice. Used by SpeechAgent for speech synthesis. Kokoro TTS A.1.6 Adaptive Memory Lect ̄ uraAgents utilizes a three-layer memory architecture to support adaptive and personalized learning experiences. Short-term memory captures recent interactions within a session, long-term memory stores persistent learner-specific data across sessions, and dynamic memory adapts to learning patterns and preferences. The adaptive memory module provides a unified interface that combines all three memory types, enabling agents to access relevant context efficiently. Table A6: Memory Types and Functionalities Memory TypeFunctionStorageUpdate Frequency Short-term MemoryHandles recent interactions and context. In-memory (session-based)Per interaction Long-term MemoryManages persistent learner-specific data. File-based (JSON)Per session Dynamic MemoryAdaptive learning patterns and preferences. In-memory + file-basedContinuously updated A.1.7 LLMs We ensure the framework supports multiple frontier models from leading LLM providers through a unified API, allowing seamless switching between different models based on task requirements, cost considerations, and performance needs. Each provider implementation handles authentication, API communication, and response formatting, while the unified interface ensures that agents can work with any supported model without code changes. This design enables flexibility in choosing the most appropriate model for each task. Table A7: Supported LLM Providers and Models LLM ProviderSupported Models OpenAIGPT-5.1, GPT-4o, o3-pro Google AIGemini 3 Pro, Gemini 2.5 Pro, Gemini Flash 2.5 Lite AnthropicClaude 4.5 Sonnet, Claude 4.1 Sonnet DeepSeekDeepSeek V3.2, DeepSeek-R1 A.1.8 Slide Content Block Types To ensure accurate alignment and rubust slide contents, we ensure each slide can support multiple content block types that enable rich, structured presentation of information. Each block type is designed for specific pedagogical purposes, from definitions and equations for core concepts to examples, steps, and questions for engagement. The framework automatically renders these blocks with appropriate styling and formatting, ensuring consistent visual presentation across all slides. Table A8: Various Types of Slide Content Blocks Block TypeDescriptionRenderingUsage BulletsBrief, concise key points about concepts and topics. HTML list elements (<ul></ul>, <ol></ol>, etc.) Holds main contents for topic DefinitionKey term definitions.HTML styled definition divCore concepts ExampleConcrete examplesHTML highlighted example divExamples EquationMathematical equationsLaTeX rendering in a divFormulas, proofs QuestionInteractive questionsHTML Question box divEngagement LinkExternal referenceshyperlink / link elementResources TableStructured dataHTML table elementComparisons, data VideoYouTube video embedsHTML iframe elementEducational short videos ImageIllustrative and educative images with captions HTML image elementIllustration StepsStep-by-step proceduresHTML numbered listAlgorithms, processes, etc. 19 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching A.2 More on Evaluation Methodology A.2.1 Overview Our evaluation adopts a rubric-based methodology for both pedagogical and comparative assessment, with generated learning and teaching artifacts scored and validated by expert educators. The evaluation examines two core capabilities of the framework: its ability to generate high-quality personalized lecture content for diverse learner profiles, and its ability to utilize these generated materials during embodied teaching. Specifically, we evaluate Lect ̄ uraAgents using four main metrics: Lecture Content Quality (LCQ), Personalization Quality (PQ), Assessment Quality (AQ), and Teaching Action Quality (TAQ). These metrics are applied across three evaluation settings: (1) Pedagogical Evaluation under Frontier Models, which assesses personalized lecture generation and embodied lecture delivery across different frontier models; (2) Comparative Evaluation with Related Frameworks, which compares Lect ̄ uraAgents with existing educational agent or personalized learning frameworks, including InstructionalAgents, LearnYourWay, and GenMentor; and (3) Efficacy Study with Students, which examines the framework’s practical learning support and learner experience using real student participants. A.2.2 Lect ̄ uraAgents’ Pedagogical Evaluation Under Frontier Models During this evaluation, we generated 40 lectures per model across seven models, resulting in 280 lectures in total. For each model, the lecture set included 10 lectures per academic level, with 20 learner profiles in total (five profiles per level). The topics covered science, engineering, history, art, and business. Details on these lectures can be found in the released dataset, available at HuggingFace 1 . The generated lecture artifacts were assessed across four evaluation metrics: Lecture Content Quality (LCQ), Personalization Quality (PQ), Assessment Quality (AQ), and Teaching Action Quality (TAQ). The evaluation followed a two-stage procedure. In Stage 1, an LLM analyst provided structured rubric-based analysis for each lecture, identifying evidence relevant to the instructional criteria under each metric, as detailed in Table A9 and Table A10. In Stage 2, expert educators reviewed the LLM-generated analysis, validated the evidence, assigned the final rubric scores, and made corrections where necessary. The verified scores were then aggregated to compute metric-level scores, overall averages, visualizations, and comparative insights into model performance across academic levels and evaluation dimensions. Table A9: Stages in Pedagogical Evaluations StageTaskCommandOutput Stage 1An LLM (GPT 5.2) gives detail analysis of generated lecture contents per academic level based on rubrics or criteria in the evaluation metrics. python3 evaluate.py \ --model model_name \ --lecture lecture_name \ --level level_name \ --llm analysis_model (JSON) Detailed analysis for each generated lecture at each academic level under a model. Stage 2An expert educator validates, scores and aggregate results for respective rubrics. python3 evaluate.py \ --aggregate \ --lecture lecture_name \ --level level_name (JSON, Charts) Comprehensive scores and results. Table A10: Details on Evaluation Metrics, Rubrics, Descriptions and Their Input Files Lecture Generation Evaluation Evaluation MetricRubrics / CriteriaDescriptionInput Files Lecture Content Quality (LCQ) AccuracyVerifies factual correctness across all generated materials. All generated files ClarityAssesses clarity of explanation across teaching materials. lecture_plan.json, learner_profile.txt, syllabus.json, scripts.json, slides_content.json, slides/*.html, lecture_notes_/*.md, quiz.json, and exam.json CoherenceEvaluates logical flow across all materials.All generated files Cognitive LoadAssesses lecture contents alignment with learner’s background or level. learner_profile.txt, syllabus.json, scripts.json, slides_content.json, slides/*.html, lecture_notes_/*.md, quiz.json, and exam.json 1 HuggingFace dataset: https://huggingface.co/datasets/Jaward/lectura-agents-data 20 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Evaluation MetricRubrics / CriteriaDescriptionInput Files Syllabus CoverageVerifies topic coverage.syllabus.json, scripts.json, slides_content.json, slides/*.html, lecture_notes_/*.md, quiz.json, exam.json, and study_guide.md Instruction-followingChecks framework’s adherence to instructions, tasks or prompts. All generated files Personalization Quality (PQ) Adaptive EmphasisAssesses the framework’s ability to adapt instructions to the learner’s learning preferences or profile through. learner_profile.txt, scripts.json, slides_content.json, slides/*.html, lecture_notes_/*.md quiz.json, exam.json, and study_guide.md Preference AlignmentChecks content alignment with learning preferences. teaching_actions.json, scripts.json, slides_content.json, slides/*.html, lecture_notes_/*.md, quiz.json, exam.json, and study_guide.md EngagementEvaluates framework’s capability to consistently engage the learner. teaching_actions.json, scripts.json, slides_content.json, slides/*.html, lecture_notes_/*.md, quiz.json, exam.json, and study_guide.md MotivationEvaluate motivational elements across learning materials. teaching_actions.json, scripts.json, slides_content.json, slides/*.html, lecture_notes_/*.md, quiz.json exam.json, and study_guide.md Tone/StyleEvaluate language appropriatenessscripts.json, slides_content.json, lecture_notes_/*.md, study_guide.md, and learner_profile.txt Assessment Quality (AQ) Concept CoverageVerifies whether assessments covered all topics in the syllabus. quiz.json, exam.json, syllabus.json, slides_content.json Cognitive AppropriatenessEvaluates assessment difficulty and its alignment with the learner’s profile. learner_profile.txt, quiz.json, exam.json, syllabus.json, slides_content.json Answer ValidityChecks accuracy of solutions to assessments. quiz.json, exam.json, syllabus.json, slides_content.json RationaleEvaluates the quality of explanation in solutions. quiz_solutions.json, exam_solutions.json Lecture Delivery Evaluation Evaluation MetricRubrics / CriteriaDescriptionInput Files Teaching Action Quality (TAQ) Temporal AlignmentValidates action-speech alignments.action_speech_alignment.json, scripts.json, speech_timestamps.json Accurate Handwriting ActionChecks accuracy of handwriting actions, i.e., whether words or phrases are written clearly and correctly at the right time frame. slides/*.html (after applied actions), action_speech_alignment.json Accurate Rough NotationChecks accuracy of rough notation actions, i.e., whether notations like highlight, underline, and circle actions are applied correctly in the right region and at the right time frame. slides/*.html (after applied actions, action_speech_alignment.json Spatial AccuracyVerifies annotation precision.slides/*.html (after applied actions), action_speech_alignment.json Active LearningAssesses the effect of teaching actions on the learner’s engagement or focus during teaching. slides/*.html, quiz.json, exam.json, action_speech_alignment.json 21 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Evaluation MetricRubrics / CriteriaDescriptionInput Files Embodied TeachingEvaluates overall embodied teaching experience. tasa_analysis.json, teaching_actions.json, slides/*.html (after applied actions), action_speech_alignment.json, scripts.json, speech_timestamps.json A.2.3 Rating Each rubric or criteria is evaluated as a boolean (satisfied or not), and these boolean scores are weighted and averaged to produce Average Achieved Ratings (AARs) at the metric and overall levels. Thus, for the j-th lecture, the overall performance score for under a given model, is computed as the weighted average of all passed rubric criteria 퐴푅 푗 푤 , given by: 퐴푅 푗 푤 = Í 푁 푗 푖=1 푤 푗 푖 · 1 푟 푗 푖 Í 푁 푗 푖=1 푤 푗 푖 · 1 푤 푗 푖 >0 where푁 푗 is the number of rubric criteria for the j-th lecture,푤 푗 푖 ∈-5, -3, -1, 0, +1, +3, +5, is the weight assigned to the i-th criterion and 푟 푗 푖 ∈ 0,1indicates whether criterion i is satisfied. When a criterion is satisfied푟 푗 푖 = 1, it contributes a positive reward of +5, +3 or +1, corresponding to a highly desirable, a desirable and important, or a nice-to-have behaviour, respectively. When a criterion is not satisfied 푟 푗 푖 =0, it is explicitly treated as a failure state and contributes a non-positive score, spanning 0, -1, -3, -5: 0 denotes the lowest-severity failure (no credit), -1 a minor failure, -3 a moderate failure, and -5 a critical failure (highly undesirable behaviour). A.2.4 Expert Recruitment and Evaluation Procedure Five expert educators were recruited through purposive sampling based on their experience in teaching, curriculum development, and educational assessment. The panel consisted of secondary-school teachers and university instructors from STEM, social science, and humanities disciplines, each with at least five years of teaching experience. Prior to the evaluation, the experts participated in an online workshop, during which the evaluation dimensions, criteria, and weighting scheme were reviewed and refined to ensure pedagogical relevance and consistency across educational levels and subject domains. During the evaluation, experts were assigned respective lecture samples according to their areas of expertise; they reviewed the generated lecture artifacts and assigned final scores based on the agreed-upon rubrics. A.2.5 Comparative Evaluation of Lect ̄ uraAgents with Related Frameworks Comparative analysis was done against two multi-agent frameworks (Instructional Agents and GenMentor) and one system (Google’s Learn Your Way). For the frameworks, we generated 20 lectures (5 for each level spanning 10 profiles) using their released code and then generated the same lectures with Lect ̄ uraAgents and compared performances. Table A11 summarizes generated lecture topics and profiles per framework or system. For Google’s Learn Your Way system, given that no source code was released we instead utilized their already generated sample lectures openly available on their website. We then generated these lectures with Lect ̄ uraAgents and compared performances as well. Our comparative evaluation assesses each framework or system based on lecture content quality (LCQ), assessment quality (AQ) and personalization (PQ) evaluation metrics using the same evaluation method described in Appendix A.2.3 and Appendix A.2.4. Table A11: Generated Lectures for Comparative Analysis Framework / SystemLecture and Learner Profile Details Instructional Agents and GenMentor Lecture Title: Newton’s Laws of Motion Learner Profile: 8th-grade high schooler interested in STEM, enjoys basketball, and prefers visual, hands-on learning through diagrams, examples, and practical activities. Lecture Title: Photosynthesis and Cellular Respiration Learner Profile: 9th-grade high schooler interested in creative writing and music, enjoys sketching, and learns biology best through story-like explanations, visuals, and everyday analogies. Lecture Title: Quadratic Equations and Functions Learner Profile: 10th-grade high schooler preparing for advanced mathematics, enjoys chess, and prefers worked examples, graph-based explanations, and short practice problems. Lecture Title: The Solar System and Planetary Motion Learner Profile: 11th-grade high schooler interested in astronomy and planetary systems, enjoys tennis, and prefers simulations, diagrams, and applied problem solving. Lecture Title: World War I: Causes and Consequences Learner Profile: 12th-grade high schooler interested in modern history and global conflict, enjoys soccer, and prefers timeline-based explanations with cause-and-effect reasoning. 22 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Framework / SystemLecture and Learner Profile Details Lecture Title: Intro to Large Language Models Learner Profile: Undergraduate computer science student interested in artificial intelligence and language technologies, enjoys basketball, and prefers intuitive explanations followed by coding examples. Lecture Title: Machine Learning: Supervised vs Unsupervised Learner Profile: Undergraduate data science student interested in machine learning methods and data patterns, enjoys hiking, and prefers visual comparisons using real datasets. Lecture Title: Molecular Biology: Gene Expression Learner Profile: Undergraduate biology student interested in genetics and molecular regulation, enjoys swimming, and prefers process diagrams with concept checks. Lecture Title: Operating Systems: Process Scheduling Learner Profile: Undergraduate learner interested in environmental science and sustainability, enjoys photography, and learns systems concepts best through visual workflows, resource-allocation analogies, and practical examples. Lecture Title: Thermodynamics: Entropy and Free Energy Learner Profile: Undergraduate chemistry student interested in thermodynamics and energy transformations, enjoys cooking, and prefers equation walkthroughs connected to everyday examples. Lecture Title: Advanced Machine Learning: Deep Neural Networks Learner Profile: Master’s-level engineering student interested in deep learning and neural architectures, enjoys tennis, and prefers model diagrams with optimization intuition. Lecture Title: Advanced Operating Systems Learner Profile: Master’s-level systems student interested in distributed computing and resource management, enjoys cycling, and prefers architecture diagrams with performance trade-offs. Lecture Title: Computational Biology: Sequence Analysis Learner Profile: Master’s-level computational biology student interested in genomics and sequence alignment, enjoys photography, and prefers algorithmic workflows with biological examples. Lecture Title: Cryptography and Network Security Learner Profile: Master’s-level learner interested in ancient history and ethics, enjoys debate, and learns cryptography best through historical examples, trust scenarios, and clear protocol diagrams. Lecture Title: Distributed Systems Architecture Learner Profile: Master’s-level computer science student interested in scalable systems and fault tolerance, enjoys tennis, and prefers system-design scenarios with failure cases.. Lecture Title: Advanced Quantum Field Theory Learner Profile: PhD researcher interested in quantum fields and particle interactions, enjoys baseball, and prefers formal derivations supported by physical intuition. Lecture Title: Non-Equilibrium Statistical Mechanics Learner Profile: PhD researcher interested in statistical physics and complex systems, enjoys tennis, and prefers rigorous mathematical development with simulation examples. Lecture Title: Synthetic Biology: Circuit Design Learner Profile: PhD researcher interested in synthetic biology and programmable cellular circuits, enjoys running, and prefers circuit schematics with lab-oriented examples. Lecture Title: Topological Data Analysis in ML Learner Profile: PhD researcher interested in topology and machine learning geometry, enjoys rock climbing, and prefers visual abstractions grounded in data examples. Learn Your WayLecture Title: Atoms and Molecules Learner Profile: Middle schooler who likes reading. Lecture Title: Carbon Learner Profile: Undergrad who likes painting. Lecture Title: Microeconomics and Macroeconomics Learner Profile: Undergrad who likes food. Lecture Title: Logical Statements Learner Profile: Undergrad who likes writing. Lecture Title: The Ancient Roman Economy Learner Profile: Undergraduate who likes plants.. 23 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Framework / SystemLecture and Learner Profile Details Lecture Title: The'Long-Haired' Comets Learner Profile: Undergraduate who likes movies.. Lecture Title: Early Human Evolution and Migration Learner Profile: Undergrad who like tennis.. Lecture Title: Intro to Data Structures and Algorithms Learner Profile: High schooler who likes basketball. Lecture Title: Critical Reading and Evidence-Based Response Learner Profile: Middle schooler who likes soccer. Lecture Title: Disruptions in the Immune System Learner Profile: Middle schooler who likes food Lecture Title: Earth and Sky Learner Profile: Middle schooler who likes photography Lecture Title: Theories of Slef-development Learner Profile: Undergrad who likes cooking. Lecture Title: What is Learning Learner Profile: Undergrad who likes music. Lecture Title: ”Reading” to Understand and respond Learner Profile: Middle schooler who likes music. Lecture Title: Micronomics and Macronomics Learner Profile: Undergrad who likes cooking. Lecture Title: An Overview of Economic Systems Learner Profile: High schooler who likes movies. Lecture Title: Early Human Evolution and Migration Learner Profile: Undergrad who likes tennis 24 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Appendix B B.1 Code and Data The data supporting this study is currently available on our huggingface repository at: https://huggingface.co/datasets/Jaward/lectura-agents- data. The code can be made available upon reasonable request from the corresponding author. Please follow the installation instructions below or in the readme file to get started. B.1.1 Installation and Usage 1.Add all required api keys inside the .env file in the parent directory. You will need to provide two main api keys (1) for the LLM you want to use (OpenAI, Anthropic, Gemini and Deepseek); (2) A SerpApi key for research, while this is optional, it highly recommended to add one, as it helps reduce hallucination. Get key here: https://serpapi.com/manage-api-key 2. Cd into the parent directory and install all required packages using this command: pip3 install -r requirements.txt 3. If you wish to use the frontend for lecture generation, start the app with this command: python3 main.py Figure 13: Frontend view (with no generated lecture) This will open the teaching environment in your browser at: http://127.0.0.1:8080/. The page should look like Figure 13: There will be a few already generated lectures in the right Lectures pane for you to quickly try or you can also generate new lectures through either the chat pane or in the left prompt pane. Generated lecture materials will appear below the slide as they are generated. 25 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching Figure 14: Swarm-of-Ranks group chat view (during lecture generation) During lecture generation you can follow the whole process unfolds in real-time in the group chat session, as shown in Figure 14. Figure 15: Teaching and learning environment (with teaching actions on lecture or study contents) After lecture generation is complete, the view will automatically update with the slide deck (as shown in Figure 15). Below the deck are controls (Next, Play, Previous, Restart, Temporal Segmentation, and Chat). 4. If you wish to use the terminal for lecture generation, run this command: python3 lecture_prep.py \ --lecture_title "Your Lecture Title Goes Here" \ 26 Lect ̄ uraAgents: A Multi-Agent Framework for Adaptive Personalized AI-Assisted Learning and Embodied Teaching --lecture_desc "Describe the kind of lecture you want here" \ --learner_profile "Add details about yourself, your learning preferences, and your current understanding level here" \ --slides <enter desired number of core slides here> \ --level <enter academic level: highschool, undergrad, masters, or phd> \ --instructor_voice <choose desired instructor voice: professor_lectura, professor_sky, professor_isabella, etc.. > \ --llm <select desired model here: gpt-5.1, gpt-4o, o3-pro, gemini-3-pro, gemini-flash-2.5-lite, claude-4.5, claude-4.1> \ --research <enter research method: llm or google> \ --language <enter output language: english, chinese, french, or spanish> \ --speech_gen <choose speech backend: kokoro-tts, gemini-2.5-tts, or gpt-4o-mini-tts> \ --handwriting_gen <choose handwriting mode: handwriting_rnn_model or preset_font_handwriting> \ --slide_image <choose slide image mode: generate_only, generate_web_search, web_search_only, material_generate_alt, material_web_alt, or material_only> \ --syllabus "Optional syllabus or curriculum text here" \ --additional_materials "Optional reference text or path(s) to .pdf, .txt, or .md files, separated by commas" \ --data_root <optional custom output directory> 5. Example prompt: python3 lecture_prep.py \ --lecture_title "Intro to Data Structures and Algorithms" \ --lecture_desc "A Computer Science lecture for a highschooler who likes basketball. Ensure covering these topics and more: 1. Introduction to Data Types and Abstraction 2. Introduction to Algorithms 3. Algorithm Vs Program. Understanding Data Structures 4. Abstract Data Types: (List, Set, Map, Priority Queue, Graph)" \ --learner_profile "Name: Taylor. Focus: Advanced Computer Science. Interests: Specialized algorithms, system design. Hobby: basketball. Learning style: Deep dive into technical details." \ --slides 24 \ --level highschool \ --instructor_voice professor_sky \ --llm gpt-5.2 \ --research google 6. To view the generated lecture in the teaching environment run this command: python3 lecture_delivery.py --lecture <lecture folder name> The folder could be, for example, intro-to-data-structures-and-algorithms. 27

Instant research discovery