Summary (Overview)

• Hierarchical multi-agent architecture for end-to-end personalized learning: Lect¯uraAgents mirrors a professor-student relationship, where a ProfessorAgent coordinates specialized assistant agents (lecture planner, research, slide, script, speech, and TASA agents) through planning, research, review, and creation of lecture content tailored to individual learner profiles.
• Teaching Action–Speech Alignment (TASA) algorithm: A novel technique that uses LLM-based semantic analysis, temporal content segmentation, and salience heuristics to generate coherent sequences of embodied teaching actions (e.g., highlight, underline, handwriting) aligned with corresponding speech segments and learner profiles.
• Embodied lecture delivery mechanism: The ProfessorAgent performs visible, pedagogically motivated teaching actions directly over slide contents while speaking, using rough notations (RN) and handwriting (HW) actions executed by a 3D quill-holding hand in a dynamic teaching environment.
• Consistent improvements across metrics: Evaluated on 280 lectures covering high school, undergraduate, and graduate levels, Lect¯uraAgents achieves an overall AAR of 80.4% (Gemini 3 Pro), outperforming existing frameworks like Instructional Agents, GenMentor, and Learn Your Way by 11–19 percentage points in lecture content quality, personalization, and assessment quality.
• Positive student learning outcomes: A small-scale efficacy study with 45 real students shows that learners using Lect¯uraAgents scored higher on post-learning assessments and reported stronger perceived understanding (100% agreement) and future learning support (87%) compared to Learn Your Way and Adobe Reader.

Introduction and Theoretical Foundation

Adaptive personalized AI-assisted learning aims to tailor instruction to individual learners, improving motivation, engagement, and learning outcomes [1–3]. However, existing solutions primarily focus on what content is recommended, rather than how it is delivered [4]. Research on embodied teaching demonstrates that performing actions such as writing, pointing, or gesturing during a lecture guides attention, fosters conceptual understanding, and enhances learning outcomes [5–7]. This underscores the need for solutions that integrate adaptive content with embodied delivery.

Recent LLM-powered agent frameworks enable planning, tool-use, and multi-step problem solving [8–11], leading to educational agents that simulate virtual classrooms [17–19], generate personalized materials [22,23], or provide tutoring support [24–26]. However, these approaches are often limited to text-only modalities, controlled simulations, or static content generation—lacking a unified model that connects personalized content generation with adaptive embodied delivery. Key pedagogical features such as coordinated lesson planning, iterative review, embodied teaching, and alignment of teaching behavior with learner needs remain insufficiently addressed.

Lect¯uraAgents addresses these limitations by introducing a hierarchical multi-agent framework for end-to-end personalized lecture generation and embodied teaching, managing the entire lifecycle of a lecture from preparation to delivery while adapting to individual learning preferences.

Methodology

Lect¯uraAgents is built on four interconnected modules—LLM, Agent, TASA, and Memory—that span two main stages: Lecture Preparation and Lecture Delivery.

Lecture Preparation Session

The ProfessorAgent leads a collaborative team through planning, research, alignment, review, and creation of personalized lecture artifacts. Multi-agent collaboration is mediated by a Swarm-of-Ranks Group Chat, where a coordinator agent (ProfessorAgent, Rank 1) supervises a validator agent (LecturePlanner, Rank 2), who in turn manages executor agents (Rank 3: ResearchAgent, SlideAgent, ScriptAgent, SpeechAgent, TasaAgent). Agents communicate via nine message types (e.g., [Task], [Approval], [Revisal]).

The process (Algorithm 1) follows these steps:

1. Planning: The LecturePlanner drafts a lecture plan based on the topic and learner profile, which is reviewed and approved by the ProfessorAgent.
2. Generation: Executor agents sequentially generate slides (HTML-based, supporting text/image/video/speech), scripts (conditioned on learner attention, understanding, preferences), and speech (synthesized via Kokoro TTS with word-level timestamps from Whisper ASR).
3. Alignment: The TasaAgent performs temporal semantic segmentation and salience heuristics to map teaching actions to slide regions.
4. Self-reflection: Each agent validates its own work against criteria before submission.
5. Personalization: All artifacts are conditioned on the learner’s profile, learning preferences, and usage history in memory (short-term, long-term, and dynamic memory).

Lecture Delivery Session

The ProfessorAgent assumes the role of an embodied instructor, executing teaching actions (RN and HW) over slide contents while speaking. Two action types are supported:

• Rough Notation (RN): Marking existing content (highlight, underline, circle, box) to draw attention, using a hand-drawn annotation library.
• Handwriting (HW): Writing new information in a natural handwriting style (RNN-based or preset font), used to reinforce learning.

Teaching Action–Speech Alignment (TASA) Algorithm (Algorithm 2)

TASA generates an ordered list of pedagogically informed teaching action-speech sequences for each slide:

$$AS_{seq} = \{ S_1[a_1, a_2, \dots, a_n], \dots, S_n[a_1, a_2, \dots, a_n] \}$$

Each action $a_n$ is defined as:

$$a_n = \{ actiontype_n, start_n, end_n, cfg_n \} \tag{1}$$

where $actiontype_n \in \{RN, HW\}$, $(start_n, end_n)$ defines the duration, and $cfg_n$ contains metadata.

First, slide contents and speech are segmented semantically with labels {Pedagogical, Personalized, Salient, Adaptive, Assessment}:

$$segment_n = \{ label_n, region_n, speech\_segment_n \} \tag{2}$$

For each segment, TASA assigns a suitable action and rationale $r_n$, recording:

$$H(S_n) = \{ segment_n, a_n, r_n \} \tag{3}$$

The TasaAgent uses this structured context to generate the final teaching action sequences $AS_{seq}$ for each slide.

Evaluation Metrics

Lecture generation is evaluated on Lecture Content Quality (LCQ), Personalization Quality (PQ), and Assessment Quality (AQ); lecture delivery on Teaching Action Quality (TAQ). Each rubric criterion carries a weight $w_i \in \{-5,-3,-1,0,+1,+3,+5\}$, and the overall score for a lecture $j$ is:

$$AAR_j^w = \frac{\sum_{i=1}^{N_j} w_i^j \cdot \mathbb{1}_{r_i^j}}{\sum_{i=1}^{N_j} w_i^j \cdot \mathbb{1}_{w_i^j > 0}} \tag{4}$$

where $r_i^j \in \{0,1\}$ indicates whether criterion $i$ is satisfied.

Empirical Validation / Results

Experimental Setup

• Models tested: 7 frontier LLMs (GPT-5.1, Gemini 3 Pro, Claude 4.5 Sonnet, Gemini 2.5 Pro, DeepSeek V3.2, GPT-4o, Qwen 3 Omni) + Kokoro TTS.
• Lectures generated: 280 personalized lectures (40 per model), 10 each for high school, undergraduate, master’s, PhD levels, covering math, science, engineering, art, history.
• Evaluation: Five expert educators used rubric-based scoring; each lecture produced 15 slides with full artifacts (plan, scripts, speech, teaching actions, assessments, notes).

Main Results (RQ1 and RQ2)

Table 4: Evaluation of Lect¯uraAgents across pedagogical metrics under frontier models. Gemini 3 Pro leads in overall AAR (80.4%), while Claude 4.5 Sonnet achieves the highest TAQ (80.4%), indicating strong embodied delivery capability. The framework enables accurate and coherent teaching action sequences across models, with TASA providing structured region-speech alignment. Temporal alignment remains an area for improvement.

Comparative Evaluation with Related Frameworks

Table 5: Performance comparison with existing frameworks. Lect¯uraAgents outperforms all baselines, with the largest gain in personalization quality (+8.9% over GenMentor), demonstrating effective adaptation to learner profiles across content, assessments, and teaching actions.

Efficacy Study with Students

45 students (15 per system, 5 per academic level) used Lect¯uraAgents, Learn Your Way, or Adobe Reader for a learning session, followed by an immediate assessment and survey.

Figure 12 (summary): Lect¯uraAgents achieved the highest average assessment scores across all learner groups (high school, undergraduate, master’s), with consistent improvements over Learn Your Way and Adobe Reader.

Table 6: Student survey responses. Lect¯uraAgents consistently receives the highest agreement across all dimensions, with 100% of students feeling it helped them gain a good understanding.

Key Observations

• TAQ scores are broadly stable across all learner profiles (Figure 10), indicating the embodied teaching mechanism generalizes across academic levels.
• Personalization-related performance is consistent across profiles, though timing-sensitive action selection remains a challenge.
• The TASA module provides a structured view of slide regions, enabling spatially accurate and pedagogically grounded teaching actions.

Theoretical and Practical Implications

Lect¯uraAgents makes several contributions to AI-assisted learning:

• Theoretical: It bridges the gap between adaptive content generation and embodied instructional delivery, operationalizing cognitive theories of attention, motivation, and information processing (e.g., Atkinson-Shiffrin, Cowan’s capacity model) into a computational framework. The TASA algorithm formalizes the alignment of speech and action through salience heuristics and semantic segmentation, providing a principled approach to multimodal pedagogy.
• Practical: The framework enables end-to-end personalized learning at scale—from lecture planning to embodied teaching—without requiring human teachers to manually adapt materials. It supports multimodal content (slides, scripts, speech, assessments, teaching actions) conditioned on learner profiles, and runs in a browser-based teaching environment. The efficacy study provides preliminary evidence that this approach improves learning outcomes and student experience compared to both traditional digital tools and other AI-assisted systems.
• Limitations: The current TASA relies on offline heuristics with a limited set of teaching actions; multi-agent orchestration introduces latency; and the framework can inherit LLM failure modes (factual errors, inconsistent reasoning). Future work should expand the action taxonomy, transition to learned policies (e.g., reinforcement learning), improve grounding to reduce hallucinations, and optimize orchestration for efficiency.

Conclusion

Lect¯uraAgents is a hierarchical multi-agent framework for end-to-end adaptive, personalized AI-assisted learning. It addresses two fundamental questions: (1) how to adaptively personalize instructional content to diverse learners, and (2) how to deliver such content in embodied, pedagogically meaningful ways. By mirroring a professor-student relationship and integrating a Teaching Action–Speech Alignment (TASA) algorithm, the framework generates high-quality lecture materials and executes coherent teaching action sequences (rough notations and handwriting) over slide contents. Extensive evaluations across 7 frontier models and 280 lectures show consistent improvements over existing frameworks in lecture quality, personalization, assessment, and embodied teaching. A small-scale efficacy study with real students further validates that the framework improves learning outcomes and learner experience. Lect¯uraAgents is positioned as a pedagogically grounded framework for personalized AI-assisted learning at scale.