Large language models support personalized tutoring, content generation, and adaptive learning paths. However, capability development requires these resources and interactions to form a coherent learning process that evolves with the learner. A central challenge is to align goals, activities, and assessment while using learning evidence to decide what should follow. We introduce Human-Participatory Multi-Agent Organization and instantiate it in Agentic University (AU). In AU, learners participate as organization members and work with specialized agents to maintain a revisable Capability Program. The program connects personal goals, learning activities across modalities, and layered assessment, while learner feedback and performance guide subsequent support. Our evaluation shows advantages in program organization, assessment design, and revision quality. Users achieved larger mean assessment gains (12.19 versus 5.61 points on a 50-point scale) and reported stronger cognitive support. This work contributes an organizational approach to adaptive learning that makes learner participation a continuing basis for multi-agent coordination. The Agentic University project is available at \href{https://github.com/AgenticOrgLab/AgenticUniversity.git}.
Figures & tables
Figure 1: Two ways of organizing multi-agent work around a learner. Left: a learner provides a request, and agents collaborate to generate learning artifacts. Right: a learner participates as a member of a multi-agent organization. Its activities support learning, while learner feedback and performance inform subsequent work.
Figure 2: AU connects the Human Node and agent roles through a Capability Program. The program includes a Learning Path, Interactive Classroom, Assessment. Classroom contributions and assessed performance provide Learning-State Evidence and Capability Evidence. The agents use this evidence to update learner state and revise remaining work. The four icons denote functional roles; the replanning role includes three diagnostic agents.
Figure 3: AU uses intake dialogue to build the Human Node. The learner describes goals, background, preferences, and constraints. AU stores this information in structured records for later planning and interaction. The displayed name and avatar are fictional.
Figure 4: AU breaks the learning goal into stages and linked knowledge-point nodes. The classroom supports each node through learning materials, Pause-QA, and mentor dialogue. Chapter assessments, stage micro-projects, and an enterprise simulation examine performance at increasing scopes. Names and avatars shown in the enterprise simulation are fictional and do not represent actual participants.
Figure 5: AU uses learner evidence to propose a local path revision. The Replanner identifies a supported difficulty and defines additional learning within the existing sequence. AU applies the revision after learner confirmation and records the updated program state.
Study
Research question
Evaluation focus
Study 1
RQ1: Representation. What aspects of learning-program quality does the Capability Program representation support?
Quality of initial learning programs.
Study 2
RQ2: Evidence conditioning. How effectively can learner evidence guide program revisions that improve learning support and maintain a coherent path?
Quality of program revisions based on learner evidence.
Study 3
RQ3: Human participation. What learning outcomes and experiences arise when learners participate in a Capability Program?
Learning outcomes, system experience, and experienced revision.
Table 1: Research questions and evaluation focus of the three studies.
Figure 6: Comparison of AU and three baselines across four outcome dimensions. The table reports mean ± standard deviation (SD). Horizontal bars show mean paired differences in score points for AU minus each baseline. Whiskers show BCa 95% confidence intervals. Purple, green, and gold bars denote comparisons with General, PLPP, and GenMentor, respectively. The four bar groups align with the table rows. The dashed vertical line marks zero difference; positive values favor AU.
Figure 7: The figure shows Capability Coverage and ten rubric submetrics for the four methods. Bars show means. Whiskers show 95% confidence intervals across 96 matched cases. Shading groups the submetrics by their main measure. The figure shows Capability Coverage separately.
Figure 8: The table reports mean ± standard deviation (SD) on a normalized 0–100 scale. The right panel compares mean scores across the four revision measures. Paired AU − GenMentor mean differences (95% paired bootstrap confidence intervals), in the plotted order, are −0.94[−3.28,1.56] , +5.78[3.28,8.28] , +8.01[5.27,10.74] , and +2.19[0.31,4.06] .
Figure 9: Overall adaptation quality for AU and GenMentor under early and late replanning. Points show individual cases. Boxplots summarize the distributions, and shaded curves show score densities. Dashed lines and labels indicate means.
Figure 10: Overview of the Study 3 procedure, including pilot testing, participant screening and allocation, pre- and post-study measures, system use, and explanatory interviews.
Figure 11: Learning outcomes before and after system use. Raincloud plots show Objective Capability, Self-Efficacy, and Metacognition for 16 participants in each group. Lines link the two observations for each participant. Labels report mean ± SD. Objective Capability uses a scale from 0 to 50. The self-ratings use scales from 0 to 100.
Figure 12: Experience ratings collected during and after system use. Bars show the means of four measures. Error bars show ±1 SD. Each group includes 16 participants.
Figure 13: Participant-level association between Learner Agency & Engagement and pre–post Metacognition gain.
Construct
AU ( n=13 )
GenMentor ( n=7 )
U
praw
pHolm
rrb
Adaptive Diagnosis
74.04±12.97
80.36±9.83
30.0
.189
.378
−.34
Instructional Responsiveness
78.85±12.89
64.29±13.36
72.0
.030
.090
.58
Post-Revision Recovery
75.00±14.43
71.43±17.25
51.0
.640
.640
.12
Table 2: Revision experience among participants who reported revision. Values are mean ± SD on a scale from 0 to 100. Each participant provides one score for each measure across their revision experience. Two-sided Mann-Whitney U tests compare the groups. Holm correction covers the three comparisons. Positive rank-biserial effect sizes ( rrb ) favor AU.
Education is one of the most promising real-world applications for Large Language Models (LLMs). However, current LLMs rely on static pre-training knowledge and lack adaptation to individual learners, while existing RAG systems fall short in delivering personalized, guided feedback. To bridge this gap, we present DeepTutor, a fully open-source agentic framework that unifies citation-grounded problem tutoring with difficulty-calibrated question generation. A hybrid personalization engine couples static knowledge grounding with dynamic learner memory, continuously adapting each interaction to the student's evolving needs. The same personalization substrate further extends to adaptive learning workflows, interactive books, and proactive multi-channel tutoring agents. To evaluate personalized tutoring, we introduce TutorBench, an interactive benchmark incorporating customized learner profiles grounded in university-level curricula across five domains. We further propose an LLM-based first-person interactive evaluation protocol that conducts assessments via a profile-driven student simulator. Complementary evaluations on established benchmarks, supported by human-alignment and ablation studies, confirm the framework's robustness and general utility. Results show that DeepTutor improves personalized metrics by 10.8% on average and strengthens general agentic reasoning across five backbone models by 29.4%.
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.
Jaward Sesay, Yue Yu, Siwei Dong +1
Beijing Institute of Technology · Peking University · Beijing Academy of Artificial Intelligence
Good tutoring adapts to the individual: it tracks what a learner knows, notices why they go wrong, and asks the next question that will help most. Most deployed tutoring tools instead serve fixed item banks and treat a wrong answer as a single bit of signal. We present CoLearn, an interactive, agentic tutor that supports an iterative tutoring loop: the learner practises, and the system builds an evidence-grounded memory of the learner's mastery and misconceptions. This memory is updated as evidence accumulates and is used to generate the next personalised question. CoLearn has three components: (i) a persistent learner-state memory that updates per-topic mastery with a soft-evidence variant of Bayesian Knowledge Tracing, where a large language model acts as a continuous observation function; (ii) adaptive question generation that targets the learner's weakest topic and recurring misconceptions; and (iii) an evidence view that makes personalisation visible and testable through live progress visualisation and blind A/B comparison. In blind A/B evaluation, questions conditioned on this memory are preferred over non-personalised ones 68-69% of the time, and in persona simulations with hidden ground-truth mastery the agent's belief converges toward the learner's true mastery.