cs.HCSep 19, 2026
SaveAdaptive Scaffolding Needs Contingency: An AI Tutor That Escalates and Fades on What the Learner Does
Organizations: National Institute of Education, Nanyang Technological University Singapore, Singapore · School of Computer Science and Engineering, Nanyang Technological University Singapore, Singapore · School of Art, Design and Media, Nanyang Technological University Singapore, Singapore · College of Computing and Data Science, Nanyang Technological University Singapore, Singapore ·
Abstract
Coding assistants raise task performance, but learners plan and monitor less. Giving less away, the usual fix, conflates two things: how much work a system carries (cognitive load) and what the learner must decide before help arrives (metacognitive demand). Our principle, preserved metacognitive demand, holds the second constant and lets the first vary. CoMeT implements it: support rises when a learner fails at a decision point and fades on take-up. Within subjects, 131 adult learners used CoMeT, an unrestricted assistant and a question-only tutor on three Python tasks. CoMeT matched the question-only tutor's demand, delivered artifacts twice as often as the assistant, and frustrated learners less than the question-only tutor, with delegation and load unchanged. Learners often did not answer. Fading held when their turn addressed the decision under support, and CoMeT surrendered the full answer in one session in sixteen, against one in six for the question-only tutor.
Figures & tables
| System | Withholds | Structures | Gates | Escalates | Fades |
|---|---|---|---|---|---|
| Prior LLM systems | |||||
| Unrestricted | – | ✓ | – | – | – |
| Hint tutor ( Bassner et al., 2024 ) | ✓ | – | – | – | – |
| CodeHelp ( Liffiton et al., 2023 ) | ✓ | ✓ | – | – | – |
| CodeAid ( Kazemitabaar et al., 2024 ) | ✓ | ✓ | – | – | – |
| Question-only ( Westbye et al., 2026 ) | ✓ | – | – | – | – |
Table 1. The design space as five properties of a design. Withholds : keeps the answer back. Structures : supplies structuring alongside. Gates : makes release contingent on what the learner has shown. Escalates : increases support on demonstrated failure. Fades : withdraws it on success.
Figure 2. The contingency dimension. Grey points are prior systems; an arrow is a release made once, labelled with its trigger; Tutor C escalates one rung at a time and returns to the first on take-up. Shaded bands mark where each failure mode becomes more likely. Positions are illustrative. A two-dimensional plot with cognitive load carried by the system on the horizontal axis and metacognitive demand on the learner on the vertical axis. Grey points mark prior systems near the low-demand region. Tutor A sits at high load and low demand, Tutor B at low load and high demand. Tutor C starts beside Tutor B and steps rightward along three labelled rungs, with a dashed arrow returning it to the first rung on take-up. Shaded bands near each corner are labelled unproductive success and unproductive failure.
Figure 3. The CoMeT interface under Tutor C across the three phases of one task: (a) Planning, editor locked and sketchpad active; (b) Monitoring, editor unlocked; (c) Evaluating, self-review before the gate opens. Numbered regions are named in the key. Four quadrants. Three are annotated screenshots of the same web interface, one per phase. In (a) Planning, a task statement sits at top left, a phase stepper runs across the top, a sketchpad panel is in the centre, the code editor is greyed out, and a chat panel is on the right. In (b) Monitoring, the code editor is active with a terminal below it. In (c) Evaluating, the stepper has advanced to review and submit, the chat panel carries a self-review prompt that the learner answers before the tutor replies, and a submit button has appeared in the top bar. The fourth quadrant is a key naming the eight numbered regions: task statement, phase stepper, the gate, sketchpad, code editor, tutor chat, terminal, and the self-review turn.
Figure 4. The two baselines. (a) Tutor A answers or builds on the first ask. (b) Tutor B returns one metacognitive question and releases only through the welfare floor, on surrender. Placement, the gates and the event record are shared with Tutor C. Two flow diagrams side by side. Left, a learner turn enters Tutor A and exits in one step as either a direct answer or a built artifact placed in the workspace. Right, a learner turn enters Tutor B and exits as one metacognitive question; a separate branch labelled welfare floor releases the artifact only when the learner states they are giving up. Both diagrams end in the same placement and event-record box.
Figure 5. Tutor C. One ladder over the task’s authored decision points: take-up returns it to the first rung and moves to the next open point; no take-up raises it one rung, through three asks, after which that point alone is conceded. A flow diagram of Tutor C. A learner turn enters a support ladder over the current open decision point: the tutor asks, the learner replies, and the reply is judged. A reply judged to take the support up returns the ladder to its first rung and the tutor advances to the next open point; a reply that does not raises support one rung, and once the asks for that point are spent the tutor concedes that point alone and moves on. Side boxes show the authored coverage points, how a reply is judged, the ladder reset, the stage gates, the welfare floor, where artifacts are placed, and the events recorded each round.
| % | ||
| Age | ||
| 18–20 | 17 | 13.0 |
| 21–23 | 79 | 60.3 |
| 24–26 | 31 | 23.7 |
| 27 or older | 4 | 3.1 |
| Gender |
Table 2. The analysed sample ( ). The lower block gives counts across the five response options, from 1 (never, or not at all confident) to 5.
| Construct | Measure | Unit | Source | RQ |
|---|---|---|---|---|
| Self-report, after each task | ||||
| Cognitive effort (demand as experienced) | effort on the task, effort on the approach (5) | per task | learner, report | 1 |
| Perceived task load | this task itself was complex (5, agreement) | per task | learner, report | 1 |
| Germane load | I really had to make sense of the task (5, agreement) | per task | learner, report | 1 |
| Help fit | direct help received, help wanted (5) | per task | learner, report | 2 |
| Ownership | tutor authorship, who did the work (4); could do it alone (5); understands the solution (5) | per task | learner, report | 2 |
Table 3. Measures. Scale points in parentheses. Self-report items are answered once per task unless noted, so each participant supplies three readings, one per tutor. Source says who produced the record: the tutor, the learner, or the learner’s self-report. The RQ column names the question each measure serves; – marks the process measure that carries the manipulation check rather than a question.
Figure 6. Design and procedure. (a) The six tutor orders, one per sequence, with the realised allocation n . (b) One session in the order met, with questionnaire item counts, and the three gated phases of a task. The strip is a sequence, not a time axis; no part of the session was timed. Two panels. (a) A six-row table. Each row is one of the six orders in which the three tutors can be met, shown as tinted cells lettered A, B or C under the headings Task 1, Task 2 and Task 3, with the number of participants assigned to that order at the right, summing to 131. (b) A strip of eight blocks in order: consent, intake, the three tasks, comparison, reflection and debrief, each block carrying its questionnaire item count. A wedge widens from the Task 1 block into an expansion below, showing the three phases of one task — Planning, Monitoring and Evaluating — each box carrying at its foot the gate condition that ends that phase: plan submitted for review, one clean run, self-evaluation then checks. The last arrow leads to the after-task questionnaire.
| Code | What the learner supplied | Example |
|---|---|---|
| Take-up evidenced | ||
| Demonstration | The substance of a decision, correct or not | Base fare first, 0.60 for children and seniors, 1.40 for adults; then the peak surcharge, adults only. |
| Artifact | A pasted plan, program or run result, with no articulation | age = input("Enter age: ") hour = input("Enter hour (0--23): ") |
| Take-up not evidenced | ||
| Assertion | A claim to be done or to understand, with no task content | All three test cases have run and the output matches. |
| Request or resistance | Asks the tutor to decide, or pushes back; reasoned argument against the tutor codes as demonstration | Can I start writing the code now? What else needs changing in my plan? |
Table 4. Warrant codes for fading episodes, ordered from highest to lowest warrant, each with one learner turn from the corpus. Each episode receives exactly one code, the highest that applies.
| Tutor C | Tutor A | Tutor B | C – A ( ) | C – B ( ) | |
| RQ1 · Metacognitive demand and cognitive load | |||||
| Effort | 3.44 | 2.98 | 3.34 | *** (0.09) | (0.09) |
| Effort on approach | 3.28 | 2.95 | 3.27 | *** (0.09) | (0.09) |
| Planned first | 3.49 | 3.25 | 3.48 | (0.09) | (0.09) |
| Perceived task load | 3.36 | 3.30 | 3.29 | (0.08) | (0.08) |
| Germane load | 3.85 | 3.74 | 3.78 | (0.08) | (0.08) |
Table 5. After-task self-report ( participants, 393 sessions). Means by tutor and mixed-effects contrasts against Tutor C (positive = Tutor C higher), with Holm-corrected within each family. All items 1–5 except tutor authorship (1–4, higher = more of the work done by the tutor). Bold contrasts survive correction; the asterisks carry the same information.
| Tutor C | Tutor A | Tutor B | C vs A | C vs B | |
| Delegation | |||||
| Asked the tutor to build | 50.4% | 51.1% | 23.7% | , | , |
| Learner turns | 16.8 | 5.0 | 14.0 | , | , |
| Delivery | |||||
| Tutor-written artifact delivered | 48.1% | 23.7% | 0 9.2% | , | , |
| Lines of code the tutor wrote | 25.3 | 15.4 | 0 4.5 | , | , |
Table 6. Behaviour from the event log, per session ( , 393 sessions). Rates are the share of sessions in which the event occurred; turns, lines, programs and runs are means. Contrasts are population-averaged GEE estimates clustered on participant, with task and position as covariates: odds ratios ( ) for the rates, rate ratios ( ) for the counts, each stated as Tutor C against the named tutor.
Figure 7. (a) Share of decisions still open at fading that the tutor later conceded, by what the learner had supplied; 95% bootstrap intervals resampled on participant, rows grouped by aim. (b) Take-up rate by rung; the first three bars are the tutor’s asks, the last the spent ladder, where the point has already been conceded. (a) A dot-and-interval chart of concession rate by warrant code, in two groups. Aimed at the decision under support: the learner explained it 30.3 percent (95 percent interval 25.3 to 35.1, 669 open decisions), pasted a plan or program 25.9 (18.0 to 33.4, 332), said they were fine 20.4 (12.5 to 28.6, 49), asked the tutor to build 33.7 (22.8 to 44.8, 86). Aimed somewhere else: wrote about another decision 65.0 (47.1 to 80.0, 20), wrote about the interface 54.9 (23.1 to 89.2, 51). The first four intervals overlap one another; the interval for another decision lies above them, while the interface interval is too wide to separate. (b) A bar chart of take-up rate by rung, with a dashed line after the third bar: the tutor's three asks at 37.0 percent of 718 rounds, 27.5 percent of 338 and 21.8 percent of 188, then rounds on a spent ladder, where the tutor concedes and builds, at 14.7 percent of 273.
| Asked the tutor to build | Full answer surrendered | Distress detected | ||||
|---|---|---|---|---|---|---|
| Tutor | % | % | % | |||
| Tutor A | 51.1 | 59.5 | 0 0.8 | |||
| Tutor B | 23.7 | 16.0 | 16.0 | |||
| Tutor C | 50.4 | – | 0 6.1 | – | 0 8.4 | – |
Table 7. What it took to get the full answer ( participants, 393 sessions). Each cell is the share of sessions in which the event occurred at any point; full answer surrendered is an end state, the whole answer given away, not the delivery rate of Table 6 , which counts a session as soon as the tutor places any artifact, whole or partial. A build request is the same act under every tutor, but Tutor A builds what is asked and no more, so its cell counts sessions in which it produced the complete program, in chat or in the workspace; Tutor B and Tutor C surrender only through the welfare floor. : McNemar exact, paired within participant, against Tutor C.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| # | Identifier | Prompt | Scale |
|---|---|---|---|
| 1 | task_authorship | Who did the work on the task you just finished? | 4-point: I did it myself, without the AI / I did it, with the AI helping / I told the AI what to do and it did it / The AI did it |
| 2 | task_alone | Could you write this again on your own, without the tutor? | 5-point: Definitely not / Probably not / Maybe / Probably yes / Definitely yes |
| 3 | task_understand | How well do you understand the solution you just submitted? | 5-point: Not at all / A little / Parts of it / Most of it / Completely |
| 4 | task_planned | How much did you work out your approach before writing code? | 5-point: Not at all / A little / Some / Quite a lot / A great deal |
| 5 | task_effort | How hard did you have to think during this task? | 5-point: Hardly at all / A little / A moderate amount / Hard / Very hard |
| 6 | task_effort_approach | How hard did you have to think about HOW to solve it — the approach, not the typing? | 5-point: Hardly at all / A little / A moderate amount / Hard / Very hard |
Table 8. The after-task questionnaire, in the order asked.
| Identifier | Prompt |
|---|---|
| session_think | With which tutor did you do the most thinking yourself? |
| session_plan | With which tutor did you think most about HOW to solve it before you started coding? |
| session_monitor | With which tutor did you check your own work most carefully as you went? |
| session_effort | With which tutor did you have to think hardest? |
| session_own_work | With which tutor did the solution feel most like your own work? |
| session_learning | Across the three tasks, which tutor helped you LEARN the most? |
Table 9. The comparative items, in the order asked.