PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic University: Collaborative Agents for Personalized Learning

Jingru Fan, Tianle Zhou, Yunjie Zhou, Zihao He, Yixuan Zhou, Kezhao Wu, Xinyu Zheng, Xiyuan Wang, Yuheng Wang, Zhi Li, Songhua Liu, Weinan E, Chen Qian

arXiv:2610.40492v1cs.AIcs.CL

TL;DR

Personalized tutoring and learning materials need to form a coherent process that aligns goals, activities, and assessment as learner evidence accumulates. Agentic University organizes learners and specialized agents around a revisable Capability Program. Across its evaluations, AU showed advantages in program organization and assessment design, and AU users had larger mean assessment gains than the comparison group.

  • Problem

    Personalized materials and tutoring need to form a coherent learning process aligning goals, activities, and assessment while using accumulating evidence to guide what follows.

    Source: §1 · p. 1

  • Method

    AU makes learners members of a multi-agent organization that maintains a revisable Capability Program connecting personal goals, multimodal learning activities, and layered assessment.

    Source: §1 · p. 2

  • Results

    AU users gained 12.19 mean assessment points versus 5.61 for the comparison group on a 50-point scale, and reported stronger cognitive support.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper frames continuing learner participation as a basis for multi-agent coordination in a learning program that remains revisable as the learner develops.

    Source: §8 · p. 25

    Evaluation used selected synthetic cases and a short human study of 32 learners aged 18–30, recruited with a focus on university students.

    Source: §7 · p. 24

Abstract

from arXiv · show

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}.

1 Introduction

The introduction argues that personalized materials and tutoring need to form a coherent process aligning goals, activities, and assessment as learning evidence accumulates. It presents a learner-participatory multi-agent organization and Agentic University, with evaluations reporting program and assessment advantages and larger mean assessment gains.

1 Introduction

  • Personalized materials and tutoring alone do not provide a coherent learning process aligning goals, activities, assessment, and evidence-guided next steps.The introduction notes that AI-assisted practice performance may not translate to unaided performance, so assistance conditions and assessed performance matter when revising learning paths.Source: §1 · p. 1
  • AU users gained 12.19 mean assessment points versus 5.61 for the comparison group on a 50-point scale, and reported stronger cognitive support.Across three studies, model-based evaluation identified program organization and assessment design as AU’s clearest advantages; AU also received higher ratings for revision appropriateness, program improvement, and path coherence.Source: §1 · p. 2
  • Human-Participatory Multi-Agent Organization makes learners organization members who work with agents to organize and adapt ongoing learning activities.AU instantiates this framework through a revisable Capability Program connecting personal goals, multimodal activities, and assessment requirements.Source: §1 · p. 2, Figure 1 · p. 2
  • The paper reports three studies covering program formation, evidence-guided revision, and human use, with evidence on program quality, revision quality, and learner outcomes and experiences.The introduction identifies organization and assessment design as the clearest program-formation advantages.Source: §1 · p. 2

2 Related Work

Related work traces personalization from learner models and instructional interactions toward coherent learning programs, coordinated multi-agent work, and continuing learner participation. AU combines these strands with aligned capability assessment and program revision.

2.1 From Personalized Tutoring to Coherent Learning Programs

  • Personalized tutoring, learner models, and adaptive paths expanded learning support, while AU organizes goals, multimodal activities, assessment, and learner-approved path changes in a revisable program.Prior systems include MALPP [71], Pxplore, GenMentor [65], and DeepTutor [74]; AU situates the learner as an organization member alongside specialized agents.Source: §2.1 · pp. 2–3

2.2 From Role Specialization to Shared-State Coordination in Multi-Agent Systems

  • Multi-agent systems organized specialized roles through shared artifacts, coordination rules, and persistent state; AU’s Capability Program links educational dependencies, evidence, diagnoses, and learner-confirmed revisions.Examples include role-based collaboration [42], staged workflows [56], task ledgers [28], and governed memory [23].Source: §2.2 · p. 3

2.3 Human Participation and Learner Agency

  • Prior work gave people control over learning decisions or agent plans, while AU connects learner goals, corrections, and confirmation to subsequent planning, mentoring, assessment, and replanning.AU’s Human Node retains learner contributions, and the Capability Program carries them across roles.Source: §2.3 · pp. 3–4

2.4 Connecting Capability Assessment to Program Revision

  • Constructive alignment, evidence-centered design, and authentic assessment inform AU’s layered assessments and its distinction between learning-state and capability evidence for targeted revisions.The system interprets evidence in light of task and assistance conditions, then links revisions to affected objectives, activities, and assessment requirements.Source: §2.4 · p. 4

3 Formative Study

The formative study examined how learners and educators connect backgrounds and goals with activities, progress judgments, and next steps. Its interviews surfaced needs around context continuity, feasible routes, task-sensitive capability judgments, and interpreting difficulties to guide support.

3 Formative Study

  • Learners and educators highlighted the need to connect backgrounds and goals with activities, progress, and next steps, informing AU’s four design goals.The formative study used purposive recruitment of seven people with learning, technical-work, or teaching experience, including five primary learners, one practitioner, and one instructor.Source: §3 · p. 4, §3.1 · p. 5

3.1 Participants and Procedure; 3.2 Analysis

  • Interviews asked how participants planned learning, judged progress, and responded to difficulty; two researchers used inductive thematic analysis to produce 22 codes and four themes.Researchers compared interpretations, reviewed supporting and contrasting accounts, and retained an audit trail.Source: §3.1 · p. 5, §3.2 · p. 5

3.3.1 F1: Learner context was fragmented and sometimes inaccurate

  • Participants described fragmented context across courses and tools, including inaccurate or outdated assumptions that could misdirect generated learning routes.They called for context continuity and correction of learner representations.Source: §3.3.1 · p. 5

3.3.2 F2: Learning routes left gaps in action and completion criteria

  • Learners said routes need feasible activities, explicit completion criteria, and planning horizons suited to competing responsibilities and immediate demands.Accounts also identified fragmented links between generated routes and learning resources.Source: §3.3.2 · p. 6

3.3.3 F3: Capability judgments depended on tasks and available support

  • Participants distinguished completing an activity from applying knowledge, and noted that performance evidence depends on the task and assistance available.Running a project with AI repair support did not by itself indicate much, while workplace capability could include coordinating resources to complete work; non-completion could also reflect time, priorities, or unsuitable resources.Source: §3.3.3 · p. 6

3.3.4 F4: Interpreting difficulties guided subsequent learning choices

  • Learners and an instructor described using interpreted difficulties or conflicting assessment results to choose further explanations, practice, resources, or priorities.Support choices distinguished conceptual misunderstanding from insufficient practice.Source: §3.3.4 · p. 6

3.4 Design Goals

The four themes informed design goals for learner representation, activity organization, performance interpretation, and program revision. The goals shape AU’s shared representations and use learner evidence and input to revise subsequent work.

3.4 Design Goals

  • AU maintains a persistent, correctable learner representation that retains relevant context, records its sources, and lets learners correct inaccurate or outdated assumptions.This goal informed the Human Node and selective SharedMemory.Source: §3.4 · p. 6
  • AU connects goals to activities, resources, prerequisites, completion criteria, and tracked work within a program scope suited to the learner’s planning horizon.This goal informed the Capability Program and follows constructive alignment.Source: §3.4 · p. 6
  • AU interprets performance against the target capability, task, conditions, timing, and allowed support, distinguishing Learning-State Evidence from Capability Evidence.Classroom contributions and assessed performance inform learner-state updates and revisions to remaining work.Source: §3.4 · p. 6, Figure 2 · p. 7
  • AU uses new evidence and learner input to revise subsequent work, address difficulties, preserve relevant work, and allow learner inspection.Learner review is supported when evidence conflicts or time and other constraints may explain performance; this follows formative assessment’s use of evidence to guide the next instructional step.Source: §3.4 · pp. 6–7

4 Agentic University

Agentic University (AU) organizes personalized capability development around a learner-maintained program linking goals, learning activities, and assessment. Learner evidence and confirmation guide state updates and local revisions across that program.

4 Agentic University

  • AU coordinates learners and specialized agents through a Human Node and a shared Capability Program that links goals, learning activities, assessment, progress, and revisions.Learners contribute goals, feedback, task work, and corrections; agents use those contributions to guide subsequent support.Source: §4 · p. 7, §4.1 · pp. 8–9

4.1 Organizational Model and Human Node

  • AU's Human Node records learner profile, current state, and shared memory, with intake dialogue turning goals and constraints into structured records for planning.The four agent roles plan learning paths, mentor, assess at multiple levels, and propose changes to remaining work.Source: §4.1 · p. 8

4.2 Forming the Capability Program

  • The Capability Program combines an ordered Learning Path, Interactive Classroom, Assessment structure, and Program State to align goals with activities and criteria.The planner uses learner goals, background, current state, preferences, and constraints alongside career references and capability graphs to set stages and prerequisite-linked knowledge nodes.Source: §4.2 · pp. 9–10

4.3 Executing the Program: Instruction, Assessment, and Evidence

  • AU supports each knowledge node with instructional materials and mentor interaction, then assesses understanding, stage-level application, and cross-stage integration.Classroom questions and interactions inform learning-state evidence, while assessed performance updates capability estimates only within tested dimensions and assistance conditions.Source: §4.3 · pp. 10–11

4.4 Closing the Loop: State Updates and Program Revision

  • AU combines compatible evidence to update learner state, preserves conflicts for diagnosis, and uses specialized evaluators to identify gaps against node targets.When revision is warranted, diagnosis determines added learning’s content and placement; the Replanner inserts it locally, preserves completed work and the surrounding sequence, and applies it after learner confirmation.Source: §4.4 · pp. 11–12

5.1 Evaluation Overview

Three studies evaluate AU’s program formation, evidence-guided revision, and use by learners.

5.1 Evaluation Overview

  • Three studies assess initial program quality, revision quality, and learning outcomes and experiences.Table 1 links each study to its research question and evaluation focus.Source: §5.1 · p. 12

5.2 Study 1: Capability Program Formation

Study 1 compared AU with three baselines on learning-program organization and assessment design. AU led on all four measures, with its largest advantage in Assessment Validity.

5.2.1 Study Design and Measures

  • The study crossed three learning goals, 16 profile settings, and two language versions to create 96 simulated learner–goal cases.Profile variations included goal clarity, time frame, AI expertise, and target-field expertise.Source: §5.2.1 · p. 13
  • The baselines were a general-purpose LLM, personalized-path system PLPP, and GenMentor, the closest system-level comparison.The study focused on learning-activity organization and evidence for assessing target capabilities.Source: §5.2 · p. 12, §5.2.1 · p. 13
  • The four methods generated one program per case; a shared model rubric evaluated outputs, and human and model composites showed pooled correspondence of Spearman’s ρ=.721 and Lin’s concordance correlation coefficient of .708.Two experts rated AU and GenMentor outputs for 24 matched cases.Source: §5.2.1 · p. 13
  • Measures covered capability coverage, program quality, assessment validity, and personalization; paired comparisons used Wilcoxon tests with Holm correction and BCa 95% confidence intervals.Coverage was judged from program titles, while the other outcomes were rubric-based.Source: §5.2.1 · p. 13

5.2.2 Higher Ratings for Program Organization and Assessment

  • AU exceeded GenMentor by 19.32 points on Assessment Validity, 10.52 on Program Quality, and 5.89 on Capability Coverage; all three differences remained significant after Holm correction.The paired-difference bars and confidence intervals are shown in Figure 6.Source: §5.2.2 · p. 14, Figure 6 · p. 14
  • AU had higher Program Quality ratings in 85 of 96 cases and higher Assessment Validity ratings in 94 of 96 cases against GenMentor.GenMentor had the highest baseline mean on each measure.Source: §5.2.2 · p. 14

5.2.3 Applied Tasks and Clear Success Criteria

  • AU exceeded GenMentor by 23.54 points on Applied Practice Validity, 22.50 on Integrated Transfer Validity, and 27.29 on Success Criteria Clarity.These ratings concern application and integration tasks and the clarity of standards for judging task outcomes.Source: §5.2.3 · p. 14
  • General and PLPP mainly produced plan text, GenMentor added session documents and quizzes, and AU linked materials, assessments, applied tasks, and projects across program levels.This structure placed practice and assessment within the learning plan.Source: §5.2.3 · p. 14

5.3 Study 2: Evidence-Conditioned Program Evolution

Study 2 compared evidence-guided revisions by AU and GenMentor across learner profiles, trajectory positions, and assessment scores. AU received higher ratings on three revision measures, while both systems scored highly on evidence-revision alignment.

5.3 Study 2: Evidence-Conditioned Program Evolution

  • Study 2 tested whether revisions based on new assessment evidence and questions improved program fit while maintaining path coherence.It followed Study 1’s evaluation of programs built from learners’ starting profiles.Source: §5.3 · p. 14

5.3.1 Study Design and Measures

  • The study matched 32 source cases across early and late trajectory positions and target scores of 20/100 and 50/100, yielding 128 AU–GenMentor case pairs.Four measures assessed improvement against each system’s initial program, path coherence, decision appropriateness, and evidence-revision alignment.Source: §5.3.1 · pp. 15–16
  • AU retained learner questions about difficult content as conversational evidence alongside assessment results, while each system revised through its own workflow.A shared external judge assessed revisions using the original program, revised program, and case context.Source: §5.3.1 · p. 15

5.3.2 Higher Revision Quality

  • AU’s mean exceeded GenMentor’s by 8.01 points on Program Improvement, 5.78 on Revision Decision Appropriateness, and 2.19 on Post-Revision Path Coherence.All three paired bootstrap intervals were above zero; Figure 8 presents the comparisons.Source: §5.3.2 · p. 16, Figure 8 · p. 16
  • Evidence-Revision Alignment ratings were high for both systems: AU scored 94.06 and GenMentor 95.00, with a mean difference of −0.94 points (95% CI [−3.28, 1.56]).The ratings indicate that revisions were well supported by evidence available in each system’s workflow.Source: §5.3.2 · p. 16, Figure 8 · p. 16
  • Overall Adaptation Quality rose from early to late replanning for AU (92.56 to 95.96) and GenMentor (87.73 to 93.26); the AU–GenMentor gap was descriptively larger early (4.83 vs. 2.70).The gap comparison does not establish a statistically significant difference.Source: §5.3.2 · p. 16, Figure 9 · p. 17

5.3.3 Addressing Learning Difficulties with Structured Tasks

  • In CASE-0025, GenMentor added in-session explanations and practice, while AU added sequenced prerequisite learning, applicability checks, and a small task before returning to the original node.AU’s assessment identified confusion about prompting, retrieval-augmented generation, and agent-based approaches.Source: §5.3.3 · pp. 16–17
  • In CASE-0008, GenMentor added explanations and code examples, while AU prescribed independent deployment practice, error recovery, and explicit acceptance criteria.The task specified how learners should use added content and check their results.Source: §5.3.3 · p. 17

5.4 Study 3: Learning Outcomes and Learner Experience

Study 3 compared the complete AU and GenMentor workflows for a shared learning goal, measuring assessment outcomes, learner experience, and revision. AU users showed a larger Objective Capability gain and stronger Cognitive Support, while interviews described how learners used feedback and learning resources.

5.4.1 Study Design and Procedure

  • Thirty-two learners pursued the same product-manager goal, with 16 allocated to each complete workflow and pre- and post-tests measuring Objective Capability on a 0–50 scale.The groups used different learning formats and revision processes during an approximately 54-hour completion window.Source: §5.4 · p. 17, §5.4.1 · p. 18

5.4.2 Data Analysis

  • Quantitative comparisons used within-group Wilcoxon tests and between-group Mann-Whitney U tests, while interviews were analyzed thematically by two researchers.Thematic codes came from participants’ accounts, and each participant counted at most once per theme.Source: §5.4.2 · p. 19

5.4.3 Findings

  • AU users gained 12.19 Objective Capability points versus 5.61 for GenMentor; the 6.58-point advantage was significant (p_raw = .035).Both groups improved, and AU's higher mean gains in Self-Efficacy and Metacognition did not show statistically significant group differences.Source: §5.4.3 · p. 19, Figure 11 · p. 20
  • Interviewees described using feedback to revisit concepts and revise work, while targeted added lessons helped some AU learners address difficulties and continue to later tasks.Three of six AU interviewees described reviewing materials or revising work after feedback; five of six described added learning addressing a difficulty.Source: §5.4.3 · pp. 20, 22
  • Participants described differing standards for judging progress, including fatigue and reassessment of understanding, which may accompany self-ratings that did not significantly differ between groups.AU’s self-rating gains were higher on average, but neither Self-Efficacy nor Metacognition showed a statistically significant group difference.Source: §5.4.3 · pp. 19–20
  • Cognitive Support averaged 78.13 under AU versus 65.63 under GenMentor, a significant 12.50-point difference after Holm correction (p_Holm = .048).AU also had higher means on Instructional Alignment, Learner Agency & Engagement, and Learning Efficiency, but those comparisons were not significant after correction.Source: §5.4.3 · p. 20, Figure 12 · p. 21
  • Within AU, Learner Agency & Engagement was positively associated with Metacognition gain (ρ = .557, p = .025), unlike in GenMentor.The authors describe this participant-level association as exploratory.Source: §5.4.3 · p. 21, Figure 13 · p. 22
  • Among participants who reported revision, AU scored higher on Instructional Responsiveness, but none of the three revision comparisons reached the corrected significance threshold.Instructional Responsiveness was 78.85 versus 64.29 (p_Holm = .090); revision-experience ratings came from 13 AU and seven GenMentor participants.Source: §5.4.3 · p. 21, Table 2 · p. 21

6 Discussion

The discussion frames AU as an organization that keeps learner contributions consequential across agent roles and links goals, activities, assessment, and revision as learners change. The studies offer bounded support for this approach, while broader accountability and organizational learning remain directions for further work.

6.1 Human Participation as a Continuing Basis for Coordination

  • AU’s organizational perspective makes whether learner contributions remain consequential across roles a criterion for meaningful participation.Constraints, practice difficulties, and corrections to inferred backgrounds can prompt reconsideration of future work.Source: §6.1 · p. 23

6.2 Maintaining Learning Continuity as Learners Change

  • The Capability Program links goals, activities, assessment requirements, and revisions to track a changing learner state while preserving completed work.Study 1 favored AU on program organization and assessment design; Study 2 favored it on revision appropriateness, improvement, and coherence, though sustained trajectories remain unvalidated.Source: §6.2 · p. 23

6.3 Making Adaptive Organizations Accountable to Learners

  • Study 3 associated AU use with stronger Cognitive Support and larger objective assessment gains, while self-efficacy and Metacognition gains did not differ significantly.The discussion argues that learners should be able to see which evidence supports decisions and how it informs subsequent work.Source: §6.3 · p. 23
  • Accountability should extend across the organization: learner corrections should update shared state, while capability judgments stay bounded by task and assistance conditions.The system-level results motivate this direction but do not isolate individual agent effects or establish a general theory of human-participatory organization.Source: §6.3 · p. 23

6.4 From Personal Programs to Organizational Learning

  • Although the studies evaluate only single-learner programs, recorded activities, revisions, and supporting evidence could inform planning and mentoring across learners.This organizational learning depends on legible evidence and learner corrections.Source: §6.4 · pp. 23–24

7 Limitations and Future Work

The authors identify limits in the empirical scope and evidence processes, and call for testing AU with broader populations, changing goals, other capability domains, and sustained use.

Empirical scope and validation

  • The studies used selected synthetic cases and a short study of 32 learners aged 18–30 recruited with a focus on university students.Broader human review, evaluator comparisons, larger and more diverse samples, and component studies accounting for learning time could clarify validity and generalizability.Source: §7 · p. 24

Reliable evidence and meaningful participation

  • Errors in generated materials or assessment feedback may affect later teaching and revision, and the effectiveness of evidence tracking and learner correction across roles remains untested.The authors propose studying correction episodes and how the organization handles conflicting evidence.Source: §7 · p. 24

Adaptation to different learners and changing goals

  • The human study used a researcher-assigned goal, and evaluated revisions primarily added support within an existing path rather than changing direction or workload.Future deployments could examine changing goals, interruptions, and when further intervention is unnecessary.Source: §7 · p. 24

Capability development beyond technical domains

  • Capability development beyond technical domains remains to be studied, including writing, the humanities, and workplace soft skills.Such work would need domain-specific assessment, such as rubrics and expert or peer feedback.Source: §7 · p. 24

Sustaining a personal learning organization

  • Continuity across educational and career transitions, learner effort, and generation cost and waiting time across the complete workflow remain unmeasured.Longitudinal deployments could assess retention and transfer alongside the resources needed to sustain support.Source: §7 · p. 25

8 Conclusion

The conclusion presents Agentic University as an instantiation of Human-Participatory Multi-Agent Organization. It frames learner participation and a revisable learning program as continuing parts of multi-agent coordination.

8 Conclusion

  • Connecting a correctable Human Node with a persistent Capability Program lets learner contributions and assessed evidence inform coordinated revisions to subsequent work.The conclusion presents this integration as the basis for keeping human participation consequential across agent roles.Source: §8 · p. 25
  • Three studies provide initial support through evaluations of program quality, revision quality, and learner outcomes and experiences.Source: §8 · p. 25
  • The broader design perspective keeps the learning program revisable as the learner develops and human participation consequential across agent roles.Source: §8 · p. 25
  • Future work should examine continuity, learner control, and reliable memory across open-ended, long-term learning trajectories.Source: §8 · p. 25
Loading 2610.40492v1…