A Computational Model of Teacher-Student Co-Adaptation
Authors: Francesco Balzan, Pedro P. Santos, Maurizio Gabbrielli, Mahault Albarracin, Manuel Lopes
Organizations: University of Bologna, Computer Science Department, Via Zamboni 33, 40126 Bologna, Italy · University of Pisa, Computer Science Department, Largo B. Pontecorvo, 56127 Pisa, Italy · Instituto Superior Técnico, University of Lisbon, Av. Rovisco Pais 1, 1049-001 Lisbon, Portugal · INESC-ID, Rua Alves Redol 9, 1049-001 Lisbon, Portugal · VERSES.ai, Los Angeles, California, USA
Effective teaching involves mutual adaptation: teachers adjust instruction based on student responses, while students actively seek information to resolve uncertainty. Yet many computational models of teaching treat instruction as a one-way process. This paper presents a Bayesian computational framework for studying teacher-student co-adaptation under partial observability in a stylized concept-learning task. Students are organized in heterogeneous groups that differ in which features of a target concept they can observe, and the teacher infers these constraints from limited feedback. We compare five interaction modes (active teaching, active learning, turn-based interaction, adaptive teaching and full co-adaptation) and three teacher-led controls across 100 independent matched simulation runs. Learner-initiated queries resolve information gaps left by the baseline teachers' stopping rules, while teacher-led continuation also achieves complete identification in the observed episodes. Bayesian belief tracking reduces the mean interaction steps required by full co-adaptation as teacher experience accumulates. At experience level 40, full co-adaptation requires fewer interaction steps on average than active learning, turn-based interaction and all three teacher-led controls, with complete observed coverage shared by these policies. Analyses by group and student type, together with sensitivity analyses, clarify the conditions and distributional limits of these results. The framework is interpretable and modular. Its findings provide a computational proof of concept for investigating interaction structure in heterogeneous groups, rather than direct evidence about classroom practice.
Figures & tables
Educational concept
Computational implementation
What the mapping captures
Where the analogy breaks down
Learner diversity
Binary observability types (each type blind to 5 of 15 features)
Differential access to information
Does not capture cognitive, affective, or cultural differences
Guided instruction / scaffolding
Information-gain-based feature selection
Teacher provides most informative cues
No dynamic difficulty adjustment; no linguistic scaffolding
Learner agency / inquiry
Student selects a feature to ask about
Learner contributes information to the interaction
Limited to predefined feature queries; no open-ended inquiry
Teacher adjusts strategy based on accumulated evidence
Assumes rational updating; no teacher uncertainty about goals
Zone of proximal development
Not directly implemented
Effective instruction targets what the learner can almost achieve
No ZPD mechanism in the model; inspiration only
Equitable outcomes
Universal task completion across all groups
All students reach correct identification
Does not capture equity of access, effort, or participation
Table 1: Mapping between educational concepts and computational implementation. The rightmost column identifies where the analogy between theory and model is limited.
Figure 1: Illustrative schematic of the teacher–student interaction setting. For visual clarity, the schematic uses a simplified configuration (15 students in 3 groups of 5, two student types, three concepts, three features); the actual simulations use a larger setting (90 students in 3 groups of 30, three student types, ~150 learnable concepts, 15 features), as specified in Table 3 . The top-left table shows the binary feature values for each of the three example concepts (Alex, Mary, John). The set of student types is { , } . The bottom-left tables display the concepts from the perspective of each type: students of type cannot distinguish hair color, and students of type cannot distinguish whether characters wear a hat. Each student type is therefore unable to observe a subset of the features, illustrating the partial-observability assumption central to the framework.
Figure 2: Mean percentage of students correctly identifying the target concept across interaction steps for each of the five interaction modes. Shaded regions show pointwise 95% percentile intervals from 30,000 whole-run resamples, preserving the three dependent groups and matched strategies in each of 100 independent runs. For episodes ending before step 15, the final percentage is retained at subsequent plotted steps when calculating the average. Left: no prior experience (0 concepts previously taught). Middle: moderate experience (20 concepts). Right: experienced teacher (40 concepts). Dashed lines indicate adaptive modes (with Bayesian belief tracking); solid lines indicate non-adaptive modes.
Exp.
Interaction mode
Final % correct (SD)
Completed/total
Interaction steps (SD)
Mean coverage %
0
Active learning
100.00±0.00
300/300
9.84±1.54
54.39
0
Active teaching
85.12±27.93
84/300
8.50±2.30
49.92
0
Turn-based
100.00±0.00
300/300
9.35±1.96
57.47
0
Adaptive teaching
81.47±32.98
102/300
10.18±2.31
39.23
0
Full co-adaptation
100.00±0.00
300/300
10.03±1.84
51.55
20
Active learning
100.00±0.00
300/300
9.98±1.60
54.44
Table 2: Quantitative results summary across experience levels. Final identification and completion include all 300 group-episodes. Interaction steps to completion are conditional on success; successful/total counts are explicit. Mean coverage averages identification across all 15 plotted steps, carrying terminal states forward. SDs describe variation across group-episodes.
Figure 3: Per-group performance trajectories at experience level 40 (40 concepts previously taught). Each panel shows one student group. Lines show mean identification across 100 independent simulation runs; shaded regions show ±1 SD across runs and describe between-run variability rather than confidence intervals for the mean. For episodes ending before step 15, the final identification proportion is retained at subsequent plotted steps. Under full co-adaptation, each group completes all 100 observed episodes; these descriptive results do not establish equivalence between group trajectories.
Figure 4: Robustness to observability severity at experience level 40. Each panel shows results for a different number of unobservable features per student type (3, 5, and 7 out of 15). Each condition uses 30 matched runs; shading gives pointwise whole-run 95% bootstrap intervals. The baseline teacher-only policies remain below universal completion, while active learning and full co-adaptation complete all observed episodes. Concept-space filtering varies across conditions, so difficulty is not held fixed.
Figure 5: Robustness to group homogeneity at experience level 40. Each panel shows results for a different within-group type distribution: high homogeneity (90%/5%/5%, matching the primary analysis), moderate (60%/20%/20%), and uniform (34%/33%/33%). Each condition uses 30 matched runs; shading gives pointwise 95% whole-run bootstrap intervals. The five core strategies are displayed.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Concept space
Candidate concepts before filtering
500
Number of features ∣Φ∣
15
Feature activation probabilities
[0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.1, 0.4,
0.4, 0.9, 0.8, 0.7, 0.5, 0.6, 0.2]
Actual concepts after filtering (across runs)
M=145.45 , SD=13.30 , range: 115–188
Appendix
Table 3: Experimental configuration. All parameter values are reported for reproducibility.
Agent
Known
Unknown
Teacher
Target concept and its feature values; group identity; features previously suggested; whether the selected respondent observed each suggested feature
Type of each student; which features each student can observe; how many students have correctly identified the target
Student
Feature values they can observe (given their type); all suggestions and answers exchanged within the group
Features they cannot observe; the target concept (initially); types and observations of other students
Appendix
Table 4: Information structure: what each agent knows and does not know.
Mode
Teacher
Students
Adapts
Belief
Analogous
initiates
query
to groups
tracking
setting
Active learning
No
Yes
No
No
Student-led inquiry
Active teaching
Yes
No
No
No
Uniform lecture
Turn-based
Yes → No
Yes
No
No
Lecture + Q&A
Adaptive teaching
Yes
No
Yes
Yes
Differentiated instruction
Full co-adaptation
Yes → No
Yes
Yes
Yes
Diff. instruction + inquiry
Appendix
Table 5: Interaction Modes.
Control
Completed
Interaction steps (SD)
Coverage %
Δ interaction steps [95% CI]
Active + continuation
300/300
11.78 (2.76)
52.16
-2.65 [-2.94, -2.37]
Adaptive + continuation
300/300
11.88 (3.03)
65.54
-2.75 [-3.02, -2.49]
Exhaustive teaching
300/300
13.69 (1.32)
23.26
-4.56 [-4.85, -4.27]
Appendix
Table 6: Teacher-led controls at experience 40: 100 matched runs, 300 group-episodes each. All reach 100% identification. Coverage averages identification across 15 plotted steps. The final column is full co-adaptation minus control, with paired-run 95% intervals over the full common episode set. Other checkpoints and contrasts are in Supplementary Code S1.
Group
Mode
Final % correct (SD)
Completed/total
Interaction steps (SD)
Group 1
Active learning
100.00±0.00
100/100
10.18±1.66
Group 1
Active teaching
86.33±25.41
22/100
9.23±2.50
Group 1
Turn-based
100.00±0.00
100/100
9.80±2.10
Group 1
Adaptive teaching
92.20±5.25
11/100
5.91±1.24
Group 1
Full co-adaptation
100.00±0.00
100/100
9.21±2.16
Group 2
Active learning
100.00±0.00
100/100
10.11±1.78
Appendix
Table 7: Group-level results at experience 40, on 100 matched runs. Dominant types are fixed by group position, with stochastic 90%/5%/5% student assignments. Interaction steps to completion exclude unsuccessful episodes.
Group
Type
N
nAT
AT % [95% CI]
nAdT
AdT % [95% CI]
1
1
2732
2516
92.09 [86.28, 97.03]
2732
100.00 [100.00, 100.00]
1
2
148
36
24.32 [14.79, 35.54]
18
12.16 [4.32, 21.68]
1
3
120
38
31.67 [20.16, 43.97]
16
13.33 [4.88, 23.21]
2
1
167
45
26.95 [16.95, 38.04]
16
9.58 [3.61, 16.76]
2
2
2690
2173
80.78 [72.79, 88.17]
2690
100.00 [100.00, 100.00]
2
3
143
59
41.26 [29.53, 53.55]
30
20.98 [11.83, 31.08]
Appendix
Table 8: Within-group type outcomes at experience 40; Type g is dominant in Group g . N counts students and n identifications; intervals resample runs. Active learning, turn-based, full co-adaptation and all three controls have n=N in every row. AT denotes active teaching and AdT adaptive teaching. Boundary empirical intervals are not population guarantees. Full eight-strategy results are in Supplementary Code S1.
Figure 6: Full co-adaptation minus comparator at experience 40, with pointwise 95% intervals. Identification and trajectory coverage use paired whole-run resampling; completion uses conservative group-position binomial bounds. These are not simultaneous or equivalence intervals.
Disabled
Mode
Final % correct (SD)
Completed/total
Interaction steps (SD)
3/15
Active learning
100.00±0.00
90/90
10.24±1.36
3/15
Active teaching
89.78±21.81
31/90
8.23±2.11
3/15
Turn-based
100.00±0.00
90/90
9.37±2.08
3/15
Adaptive teaching
91.81±14.80
22/90
7.18±1.85
3/15
Full co-adaptation
100.00±0.00
90/90
9.06±2.14
5/15
Active learning
100.00±0.00
90/90
9.73±1.50
Appendix
Table 9: Observability-severity sensitivity at experience 40: 30 runs, hence 90 group-episodes, per condition. The five core strategies are shown; all eight policies are included in the executable outputs. The five-disabled reference reuses the first 30 primary runs. Times are conditional on completion.
Condition
Worlds
Concepts: mean (SD)
Range
Mean log2Nc
Primary
100
145.45 (13.30)
115–188
7.18
5 disabled / high homogeneity
30
143.60 (13.45)
115–170
7.16
3 disabled
30
267.70 (16.21)
226–301
8.06
7 disabled
30
56.63 (5.44)
48–70
5.82
Medium homogeneity
30
146.43 (15.54)
118–178
7.19
Uniform mixture
30
147.07 (14.52)
124–182
7.19
Appendix
Table 10: Actual retained concepts and initial uncertainty in the simulations. Five-disabled/high-homogeneity sensitivity uses the first 30 primary worlds. Conditions with different observability need not have equal task difficulty.