PaperlayerView exact revision on arXiv

Source-linked AI summary

A pedagogically guided LLM framework for scaffolding human-AI cocreation and preserving learners' agency and creativity

Ijaz Ul Haq

arXiv:2610.33051v1cs.CY

TL;DR

LLMs can perform creative work and decision-making that learners are meant to practise, while existing cocreative learning lacks a way to derive appropriate AI support from activity requirements. This paper proposes educator-approved specifications and a constrained workflow for applying them; worked examples illustrate support that varies by activity while remaining traceable to educator decisions.

  • Problem

    LLMs can generate and direct ideas that students are expected to develop, but how their behaviour should follow the cognitive and pedagogical requirements of a creative activity remains unresolved.

    Source: §1 · p. 2

  • Method

    Educators specify learner-owned work and permitted support for each activity, then a proposed workflow applies that specification through prompting, retrieved guidance and pre-delivery checks.

    Source: §1 · pp. 2–3, §7 · p. 17

  • Results

    The worked examples illustrate activity-specific support whose constraints remain traceable to educator decisions.

    Source: §1 · p. 3, §7 · p. 17

  • Takeaways & Limitations

    Educators and researchers can reuse and revise the specifications to examine AI support while students retain agency over ideas and learning decisions.

    Source: §7 · pp. 17–18

    The workflow awaits integrated testing, including whether it provides useful support while detecting replies that perform learners’ work or choose their direction.

    Source: §6 · p. 17

Abstract

from arXiv · show

Large language models (LLMs) can generate alternatives, elaborate ideas and supply rationales, overlapping with cognitive work learners are expected to perform. However, human-AI cocreation still lacks a mechanism for determining how AI can scaffold learners' creative decision-making. This study develops a cognitive-pedagogical framework for deriving the behaviour of a pedagogically constrained LLM from an educator-approved specification of each activity in a cocreative process in project-based learning. The creative-process phase supplies the activity's purpose, typical processes and pedagogical conditions; the active section of a creative technique specifies how the group works on its ideas. For each task, an educator approves four fields: a cognitive and metacognitive objective, a pedagogical objective, the task and group artefact as context, and the learner-owned product. The objective identifies the creative operations and decisions about progress, focus, criteria and next steps that remain with the group. Four translation rules derive permitted support, timing, feedback scope and prohibitions. The framework is illustrated with Viacocrea's six phases and two activities: Plan Brainstorming and a constructed Build activity using PPCo (Pluses, Potentials, Concerns, Overcoming concerns). In the proposed workflow, a controller applies the approved specification to student requests and feedback offers. Prompt engineering and approved pedagogical guidance retrieved through retrieval-augmented generation (RAG) inform a draft, which is checked for form, pedagogical function and safety before delivery. The worked examples illustrate how support can differ between activities while its constraints remain traceable to educator decisions. The framework offers educators and researchers a reusable way to derive LLM support from creative activities while preserving learners' agency and creativity.

1. Introduction

Cocreative learning involves iterative group work, but LLMs can perform the creative operations and decisions students are meant to practise. The paper proposes educator-approved activity specifications and a constrained workflow to derive and apply permitted support.

1. Introduction

  • LLM support raises the question of whether students practise creative thinking or the model performs work assigned to them.Prior studies reported improved ratings for individual writers’ ideas but reduced story diversity, or mostly null effects on divergent association; neither examined groups learning through structured creative techniques.Source: §1 · p. 2
  • An educator specifies each activity through cognitive and metacognitive objectives, pedagogical objectives, task and artefact context, and the product learners must author.The resulting permissions and prohibitions must be traceable to the specification and prevent the model from doing creative work or directing decisions reserved for the group.Source: §1 · p. 2
  • The proposed workflow uses prompting, approved retrieved guidance, controlled activation and response checks to apply the specification to student requests and educator-approved feedback offers.Prompting expresses permitted and prohibited behaviour; retrieval supplies current-activity guidance; pre-delivery checks assess drafts against the specification, learner-owned work and safety requirements.Source: §1 · pp. 2–3
  • The framework provides a human-in-the-loop procedure, a constrained workflow and a reusable template for relating cocreative activity design to permitted LLM support.Cocreative work may unfold across phases and be revisited as projects change direction; the paper describes Viacocrea’s six phases and structured techniques in a shared digital workspace.Source: §1 · pp. 1–3

2. Theoretical and technical foundations

The foundations connect creative-process research, group cognition and educational AI to the need to specify what learners do in a particular activity. They motivate activity-sensitive support that leaves assigned creative work and decisions with the group.

2.1 Creative process, phases and techniques in creativity research

  • Creative-process phases indicate purposes and plausible work, but do not establish which cognitive operation a group is performing at a given moment.Viacocrea’s phases move from forming a group and defining a challenge through planning and developing ideas to selecting and communicating conclusions; techniques further specify how learners work on ideas.Source: §2.1 · pp. 3–4

2.2 Group creativity, scaffolding and creative metacognition

  • A group’s written entries trace contributions to an emerging artefact but cannot reveal who framed, evaluated or redirected the work.Creative metacognition includes monitoring progress and controlling focus, strategy, criteria and timing; those decisions may be negotiated among group members.Source: §2.2 · pp. 4–5
  • Group brainstorming research reports both stimulation and narrowing of idea variety, while LLM-assisted studies report differing product, diversity and responsibility outcomes across settings.Scaffolding accounts distinguish structuring activity from prompting deeper examination; assistance that supplies an idea may improve a product while removing learners’ opportunity to construct or negotiate it.Source: §2.2 · pp. 4–5

2.3 LLMs and the need for pedagogical constraint

  • Because LLMs can generate, compare and direct ideas, support must be judged against the creative work assigned to students in the activity.In cocreative tasks, a question can introduce an option or choose a focus, while a factual definition may be appropriate when students are meant to apply the concept.Source: §2.3 · p. 5

2.4 Prompting, retrieval and response checks

  • Prompting, retrieved activity guidance and response checks address different parts of constrained support, but checks alone cannot establish whether a reply takes over learners’ work.Retrieved material does not override educator instructions; an LLM judge may assess pedagogical function, but its reliability must be tested against human-labelled examples because judgements can be biased or mistaken.Source: §2.4 · pp. 5–6

2.5 Contemporary work on creativity and AI

  • The framework treats the group’s current activity as the unit to specify, combining phase purpose, technique-section instructions and task context to determine learner work and permitted assistance.Table 1 is described as showing what each contemporary approach contributes and the question it raises for a group working within a particular creative technique.Source: §2.5 · pp. 6–7, Table 1 · p. 6

3. The cognitive–pedagogical framework

The framework connects educator specification of learner-owned activity work to a constrained workflow that applies the approved support boundaries.

3. The cognitive–pedagogical framework

  • For RQ1, educators combine the creative-process phase, active technique section and task to specify learner work and permitted LLM support.The specification prevents functional substitution of learners’ cognitive work and attentional substitution of the group’s focus, criteria or next step.Source: §3 · p. 7, Figure 1 · p. 7
  • For RQ2, prompting, retrieved guidance and pre-delivery checks apply the approved specification to student requests and educator-approved feedback offers.The framework figure presents educator specification and controller application as connected parts.Source: §3 · p. 7, Figure 1 · p. 7

3.1 From phase and creative technique to an activity specification

The specification starts from an interpreted phase purpose, then combines it with the active technique section and task. Educators document the basis for these interpretations and approve the work assigned to learners.

3.1 From phase and creative technique to an activity specification

  • For each Viacocrea phase, the framework interprets its stated purpose, group activity and object of work in light of creative-cognition accounts.Table 2 presents plausible cognitive processes and pedagogical conditions; these interpretations guide educators alongside the task and active technique section.Source: §3.1 · pp. 7–8, Table 2 · p. 8
  • Framework labels such as FRAME, GENERATE, EVALUATE and SELECT describe possible processes and are proposed interpretations, not Viacocrea prescriptions.The specification also records how the group monitors progress and decides criteria, focus and when to move on.Source: §3.1 · p. 8
  • Researchers document phase descriptions and technique instructions, then combine the phase, active technique step and current task to specify an activity.A technique step receives a separate specification when it changes assigned learner work; uncertain interpretations are resolved by the educator before AI support is approved.Source: §3.1 · pp. 8–9

3.2 From activity specification to AI-support specification

The approved activity specification defines learners’ work, and four translation rules derive the LLM support permitted for that work. The rules set the support objective, role and timing, feedback scope, and behavioural constraints.

3.2 From activity specification to AI-support specification

  • The four translation rules turn approved learner work into a support objective, AI role and timing, feedback scope, and behavioural constraints.These outputs together form the AI-support specification.Source: §3.2 · p. 9

3.2.1 Cognitive and metacognitive objective to support objective (TR1)

  • TR1 derives permissible help from the cognitive and metacognitive objective, avoiding help that generates ideas or decides when the group should stop.In Plan Brainstorming, the LLM may ask what different directions students see among their entries, but must not supply a candidate idea or declare idea generation complete.Source: §3.2.1 · p. 9

3.2.2 Pedagogical objective to AI role and timing (TR2)

  • TR2 uses the pedagogical objective to determine the AI’s role and when help may be requested or offered.Examples include a participation cue in Plan Brainstorming and a question about uncertainty in a Build representation; students may defer or dismiss offers.Source: §3.2.2 · p. 10

3.2.3 Task and group artefact to feedback scope (TR3)

  • TR3 sets feedback scope from the task and group artefact, so an unranked idea set is not prematurely narrowed to one entry.A question about a single entry may become permissible after the group selects it in a separately approved activity.Source: §3.2.3 · p. 10

3.2.4 Learner-owned product to behavioural constraints (TR4)

  • TR4 prohibits the LLM from supplying learners’ assigned product or making decisions reserved for the group, unless an exception is approved.An educator may approve a model-generated idea for critique or sourced factual clarification when supplying that fact is outside the assigned work.Source: §3.2.4 · p. 10

3.3 Protecting learners’ creative work and decisions

The framework defines reserved learner work by the current phase, technique section or step, and task. It uses five questions to assess both proposed AI moves and draft responses against those boundaries.

3.3 Protecting learners’ creative work and decisions

  • Reserved group work includes the product learners author, the work needed to produce it, and their monitoring and decisions about focus, criteria, and when to move on.Even a question may cross the boundary by supplying content or directing a choice.Source: §3.3 · p. 10
  • Five questions test whether a response supplies assigned content, bypasses creative work, redirects group decisions, or carries out the approved supportive move.The questions also ask whether learners could adopt the response as their product and whether it declares enough work has been done.Source: §3.3 · p. 10
  • Educators resolve uncertain move types before interaction, and uncertain drafts are withheld before delivery.A single response may both supply content and redirect attention.Source: §3.3 · p. 10

3.4 Applying the framework to two creative techniques

The framework is illustrated through Plan Brainstorming and a constructed Build activity using PPCo. The examples assign different support to the work required in each technique and step.

3.4 Applying the framework to two creative techniques

  • Plan Brainstorming has the group generate possible actions without ranking them, while each PPCo step calls for different work and therefore receives its own specified support.In the example, the group revises its idea in PPCo’s final step.Source: §3.4 · pp. 10–11
  • The examples pair Plan Brainstorming with a Build PPCo sequence and show the AI support derived for each.PPCo starts from an idea selected earlier and moves through Pluses, Potentials, Concerns, and Overcoming concerns.Source: §3.4 · p. 10, Table 5 · p. 11
  • The Build placement, process labels, and assignment of revision to PPCo’s final step are proposed for this example, not reported classroom use.The source guide places PPCo in its Develop stage; the article uses it here to illustrate Viacocrea’s Build phase.Source: §3.4 · pp. 10–11

4. Operationalisation in a constrained LLM workflow

The proposed controller applies an approved activity contract to student requests and feedback offers, using prompts and retrieved guidance to inform drafts. It checks drafts before delivery, while the paper identifies validation needs and presents constructed—not observed—response examples.

4. Operationalisation in a constrained LLM workflow

  • The controller selects the current activity contract, screens input, checks eligibility, retrieves guidance, and checks the draft and activity state before delivery.The LLM cannot change the contract or advance the group’s phase.Source: §4 · p. 11

4.1 Prompt architecture

  • The prompt states the current activity’s permitted help and learner-reserved work, while approved guidance is retrieved by identifier and version rather than similarity.Retrieved guidance informs a response but does not authorize an AI move or provide project answers; the educator approves the contract and guidance together.Source: §4.1 · p. 12, §4.2 · p. 12

4.3 Responding to students and offering feedback

  • Student requests and feedback offers use the same contract but differ in who initiates help; offers require an approved checkpoint, sufficient group work, frequency limits, and an interval.Students may defer or dismiss an offer, and the controller remains silent when offer conditions are unmet.Source: §4.3 · pp. 12–13

4.4 Checking responses before delivery

  • Automated checks assess draft form, pedagogical function, and safety, withholding failed or uncertain drafts before delivery.Passing the checks does not establish that a response is useful to learners.Source: §4.4 · p. 13

4.5 Constructed response examples

  • Before classroom use, human-labelled tests must assess judge errors across intended languages, age groups, activities, and both help routes.Table 8’s response pairs are constructed examples, not observed LLM outputs; the judge may share the generator’s errors.Source: §4.4 · pp. 13–14, §4.5 · p. 14

5. Discussion

The discussion argues that LLM support must be specified for the work of each activity, not inferred from a phase label or a general AI role. It also stresses that compliance checks cannot establish educational value: support must preserve learners’ decisions, and its effects in group work remain to be examined.

5.1 From creative process to the work of one activity

  • A phase, technique section and task together locate the group’s work; a cue that broadens Brainstorming may interrupt appraisal of a selected PPCo idea.The discussion uses this contrast to show why a phase name alone does not specify appropriate support.Source: §5.1 · p. 15
  • Learner ownership includes deciding focus, criteria and next steps, since supplying a required concern or choosing which concern matters can substitute for group work.The discussion distinguishes functional substitution from attentional substitution and contrasts monitoring a range of entries with declaring generation complete.Source: §5.1 · p. 15

5.2 From educator judgement to constrained LLM behaviour

  • The four translation rules make permissions and prohibitions traceable to the activity specification, but retrieved guidance grants no extra permission and source faithfulness does not ensure pedagogical fit.The discussion notes that student-authored text may try to redirect the model, and that a reply can still supply work reserved for the group.Source: §5.2 · pp. 15–16
  • A functional response judge may accept subtle substitution or reject useful help, so human-labelled validation must examine both errors across intended activities, languages and learners.Form checks alone cannot determine whether a question leaves idea generation or priority setting to learners; usefulness must be assessed alongside compliance.Source: §5.2 · p. 16
  • Requests and checkpoint offers follow the same learner-owned boundary; students may defer or dismiss offers, while later independence requires evidence from unaided work.A decline in offers measures controller behaviour, not learner independence.Source: §5.2 · p. 16

5.3 What constrained support means for group creativity

  • The role of an AI-generated example depends on the assigned task; prior studies report differing outcomes, and neither examined groups using a structured creative technique.The discussion frames whether students performed the intended creative work and decisions as a question for such learning.Source: §5.3 · p. 16
  • A project workspace could select approved specifications as groups progress, making support specific to each activity compared with an unconfigured general-purpose LLM.The proposed workspace would identify the phase, technique section and group entries, using reusable technique records.Source: §5.3 · p. 16
  • Delivery checks cannot determine how groups use a reply; educational value depends on how members develop and direct ideas together and what they later do without AI.A permissible question may still interrupt exchange or divert attention from a quieter member’s proposal.Source: §5.3 · pp. 16–17

6. Limitations and future research

The paper’s examples apply the framework to selected Viacocrea activities, but broader applicability and system performance remain untested. Future work should evaluate educator specifications, the constrained workflow, and learning outcomes in classroom settings.

6. Limitations and future research

  • The examples cover Viacocrea phases, Plan Brainstorming and a constructed PPCo activity in Build; application to other creative processes and techniques remains unexamined.Researchers are encouraged to document phase and technique sources and examine whether educator specifications are defensible and how much preparation they require.Source: §6 · p. 17
  • The proposed workflow awaits integrated testing to determine whether it delivers useful support while detecting replies that perform learners’ work or choose their direction.Testing should span intended activities, languages, age groups, both help routes, and validated safety procedures and student-text processing arrangements for minors.Source: §6 · p. 17
  • Future classroom studies can combine prior measures of group-idea novelty and flexibility with discussion, revision, participation and subsequent unaided creative work.The paper presents this as a way to connect group-idea measurement, activity-specific AI support and study of what learners do with it.Source: §6 · p. 17

7. Conclusion

The conclusion argues that learners’ creative work and decisions should be specified for each activity before AI support is introduced. It presents a framework and proposed workflow for deriving support from that assignment, with reusable specifications for educators and researchers.

7. Conclusion

  • A polished group product alone cannot reveal who generated alternatives, appraised ideas or chose the direction, motivating specification of learner-owned work for each activity.The conclusion frames this as the central question for human–AI cocreation in project-based learning.Source: §7 · p. 17
  • The framework combines phase purpose, active technique-section instructions and task context with four educator-approved fields to derive permitted support and prohibitions.The rules address functional substitution and attentional substitution; a proposed workflow applies the decisions to requests and feedback offers through prompting, retrieved guidance and pre-delivery checks.Source: §7 · p. 17
  • Educators can reuse and revise specifications as projects develop.The conclusion also identifies researchers’ potential use of specifications to describe interventions and examine effects on idea generation, appraisal and shared control.Source: §7 · p. 17
  • For system researchers, the specifications define requirements that can be implemented and tested.The conclusion presents this alongside study of whether AI support advances creative learning while learners retain agency over ideas and decisions.Source: §7 · p. 18

Declarations

The declarations identify available supplementary materials, note that the conceptual paper collected no new participant data, and report the author’s acknowledgement and competing-interests statement.

Declarations

  • Supplements S1–S3 accompany the preprint, and additional framework materials are available from Ijaz Ul Haq on reasonable request.The examples in Table 8 are constructed; no executable or validated system is provided.Source: Declarations · p. 18
  • The conceptual paper collected no new participant data and did not reanalyse data from earlier studies.Future classroom research will require ethical approval, appropriate consent and safeguards for student data, including data from minors.Source: Declarations · p. 18
  • Dr. Aamir Anwar contributed to manuscript revision and improvement; the preprint version is currently under review.Source: Declarations · p. 18
  • The author declares no competing interests.Source: Declarations · p. 18
Loading 2610.33051v1…