Jean Marie Ntsama
MSCA Project
Cognitive Trapeze: Modeling Language–Thought Co‑Development in AI‑Augmented Learning
A Computational Approach to Literacy Acquisition in Typical and Neurodiverse Learners
Executive Summary
The Cognitive Trapeze project investigates how language and thought co-evolve within AI-augmented learning environments.
The project integrates conceptual analysis, cognitive experimentation, multilingual data, and computational modeling to understand how learners navigate between linguistic structures, cognitive representations, and technological mediation.
Core pedagogical objective: The project specifically targets guided reading in priority education zones (ZEP) in Cameroon. It aims to demonstrate how AI-assisted multimodal tools help learners bridge the gap between linguistic decoding and deep semantic comprehension.
Field setting: 4 pilot schools representing the 4 types of Cameroonian educational institutions (CES, CETIC, General Lycée, Technical Lycée), totaling approximately 160 learners. All field activities are officially supervised by the Regional Delegation of MINESEC (Adamaoua) and its Regional Inspectors (ICR) for IT and Languages.
Technical Proof of Concept – ASR Results from 60+ ZEP Learners
To validate our ASR engine choice, we conducted a comparative evaluation campaign under real classroom conditions in Cameroon’s Adamaoua ZEPs. Over 60 learners were tested on 18 corpus texts, comparing Whisper models (Tiny, Medium, Large).
The results are clear:
| Text | Grade | Whisper Tiny | Whisper Medium | Whisper Large |
|---|---|---|---|---|
| Text 1 | 6th grade | 50.6% | 23.4% | 1.3% |
| Text 2 | CE1 (Grade 1) | 75.5% | 29.6% | 24.5% |
| Text 3 | CE2/CM1 | 97.4% | 32.9% | 23.1% |
| Text 4 | CE2 | 83.8% | 31.5% | 25.7% |
| Text 5 | CM1/CM2 | 78.0% | 24.4% | 14.9% |
| Text 7 | CM2/6th grade | 75.5% | 12.9% | 14.8% |
(WERs shown are averages for Whisper models. Large is the most reliable, delivering remarkable accuracy even in the most challenging conditions).
🔴 The Critical Challenge: Semantic Hallucinations
Despite the excellent performance of Whisper Large, qualitative analysis reveals a fundamental problem for pedagogical use: semantic hallucinations. The model transcribes the sounds correctly, but inverts the meaning of the sentence or shifts the verb tense.
Real examples observed with our learners:
- “La maman prépare le repas” → “La maman ne fait pas le repas” (Intrusive negation)
- “Je regarde par la fenêtre” → “Je ne regarde pas la fenêtre” (Meaning inversion)
- “Papa a acheté” → “Papa achetait” (Tense drift)
✅ The Architectural Solution: A Dual-Layer ASR + LLM
These semantic errors are invisible to a simple phonetic corrector. They can only be detected by a Language Model (LLM) capable of understanding context.
This is why the TDR v2 of the Cognitive Trapeze application integrates a mandatory dual-layer architecture:
- ASR Engine (quantized Whisper Large): for phonetic decoding.
- LLM Engine (Qwen 0.5B or Phi-3-mini): for semantic verification, detecting meaning inversions, and correcting tense drifts before delivering feedback to the student.
This architecture is not an option: it is the guarantee of reliable pedagogical feedback.
Project Status (August 2026)
✅ Host institution and primary supervisor confirmed: The project is hosted by the University of Copenhagen (UCPH), Denmark, under the supervision of Prof. Daniel Spikol. The official Letter of Support and Intended Supervision has been received and signed by Prof. Spikol on 21 August 2026, confirming his full commitment to the project and the planned field missions in Cameroon.
✅ Co‑supervision and secondment confirmed: Professor Marcelo Worsley (Northwestern University, USA) has officially agreed to serve as co‑supervisor for the computational and AI prototyping components, with a 6‑month secondment (M10–M16) at the tiilt lab.
✅ Field supervision: The Regional Delegation of MINESEC (Adamaoua) and its Regional Inspectors (ICR) will officially supervise all field activities. Prof. Pius Ondoua (University of Yaoundé I) acts as academic advisor for the philosophical coherence of the model.
✅ Proposal fully validated: The full proposal (Part A, Part B-1, Part B-2) has been reviewed and validated by the UCPH Research Support Office and the supervision team. The final submission to the EU Funding & Tenders Portal is scheduled before the 9 September 2026 deadline.
PART B1 — EXCELLENCE
1. Project Title
Cognitive Trapeze: Modeling Language–Thought Co‑Development in AI‑Augmented Learning
Subtitle: A Computational Approach to Literacy Acquisition in Typical and Neurodiverse Learners
2. Summary
This project explores how language and thought co-evolve within AI-augmented learning environments.
Building on the Cognitive Trapeze Model and the field experience of the IA4ZEP programme (supervised by the MINESEC ICRs), it proposes a dynamic computational framework to model how learners move between linguistic structures, cognitive representations, and technological mediation.
3. Objectives
Scientific objectives
- Model the dynamic interplay between language, thought, and technology via Bayesian networks and multi‑agent simulation.
- Analyze how AI tools influence conceptual development in reading.
- Investigate cross‑linguistic and neurodiverse learning patterns (dyslexia, language disorders).
Methodological objectives
- Field mission 1 (M8–M10): Baseline assessment (N=160 learners) and preliminary data collection.
- Secondment (M10–M16): MVP APK development at the tiilt lab (Northwestern) with Prof. Worsley.
- Field mission 2 (M16–M20): Prototype deployment and pilot test (N=30 learners, 15 dyslexic, 15 typical).
- Produce a multilingual dataset of AI-mediated reading interactions.
Career development objectives
- Strengthen expertise in computational cognitive modeling through the Northwestern secondment.
- Build a sustainable international research network (Europe–Africa–USA) through collaboration with Prof. Worsley, UCPH, and MINESEC.
- Publish 4 rank‑A articles (M8, M16, M22).
PART B2 — IMPACT
Impact on the researcher
- Advanced expertise in cognitive modeling and AI pedagogy.
- Strong interdisciplinary profile (philosophy, linguistics, AI).
- Enhanced international visibility via transatlantic and African networks.
Impact on the host institution
- New collaborations in cognitive sciences and digital education.
- Innovative methodologies for AI‑assisted literacy.
- Transcontinental and institutional research framework (MINESEC).
Impact on Europe
- Contribution to digital education and AI ethics frameworks.
- Innovation in cognitive sciences for multilingual learning.
Impact on Africa
- On‑the‑ground impact: Concrete improvement in reading comprehension in Cameroon’s ZEPs (4 pilot schools).
- Strengthening of multilingual digital literacies (French/Fulfulde).
- Inclusive and scalable learning models for neurodiversity.
PART B3 — IMPLEMENTATION
Work Packages
WP1 — Theoretical Foundations (M1–M8)
Formalization of the Cognitive Trapeze model into a predictive Bayesian network and agent‑based simulation (ABM).
Deliverable: Technical report + Article 1 (Theoretical) submitted at M8.
WP2 — Empirical Validation (M6–M22)
- M8–M10: Baseline assessment (N=160).
- M10–M16: Researcher is on secondment at Northwestern. No field activities.
- M16–M20: MVP prototype deployment and data collection in the 4 schools. Supervised by MINESEC ICRs.
- M20–M22: Final statistical analysis (Bayesian regression, multilevel models).
Deliverables: Article 2 (Descriptive) submitted at M16 + Article 3 (Confirmatory) submitted at M22.
WP3 — AI Prototyping and Computational Modeling (M10–M22)
- M10–M16: Intensive secondment at Northwestern (tiilt lab). Android APK development (quantized Whisper Large + Qwen 0.5B LLM).
- M16–M20: Pilot test of the prototype in the 4 schools (N=30).
- M20–M22: Data analysis and technical paper writing.
Deliverables: Functional alpha APK + Article 4 (Technical) — preprint deposited on arXiv in December 2026 for community feedback, then final submission at M22 (co-author Marcelo Worsley).
WP4 — Consortium and Sustainability (M15–M24)
Building a sustainable Africa–Europe partnership (institutional agreements) and ERC Synergy pre‑proposal.
Deliverables: Consortium agreement + ERC pre‑proposal (M23) + dissemination workshop (M24).
Gantt Chart (Textual Overview)
- Months 1–8: WP1 (Theory – Europe)
- Months 8–10: WP2 (Field mission 1 – Cameroon)
- Months 10–16: WP3 (Secondment – Northwestern, USA)
- Months 16–20: WP2/WP3 (Field mission 2 – Cameroon)
- Months 20–22: WP2/WP3 (Final analysis – Europe)
- Months 22–24: WP4 (Consortium & ERC – Europe + Cameroon)
| Publication milestones: Article 1 (M8) | Article 2 (M16) | Articles 3 & 4 (M22) |
Risks and Mitigation
- Low participant availability: Expand recruitment via the IA_4_ZEP network and MINESEC ICRs.
- Modeling complexity: Iterative prototyping with collaboration from Northwestern students.
- Data variability: Mixed‑method triangulation.
Ethics and Open Science
- Institutional compliance: All fieldwork conducted under MINESEC Authorization Ref. N° 658/2026/AR/H/DRES (Regional Delegation of Adamaoua).
- GDPR compliance for all learner data; written parental consent and child assent obtained.
- Anonymized datasets: All learner names replaced with pseudonyms; no biometric or identifiable data shared.
- Open-source code & data: GitHub (MIT license), Zenodo (CC-BY 4.0), with permanent DOIs.
- Four published datasets:
- ASR Corpus — DOI: 10.5281/zenodo.22922869
- Tendani Ito Comic — DOI: 10.5281/zenodo.23065165
- AI Writing Lab (Text-to-Image) — DOI: 10.5281/zenodo.23065607
- AI Writing Lab (Narrative Productions) — DOI: 10.5281/zenodo.23230681
- Preprint on arXiv co-signed with Prof. Marcelo Worsley.
- Ethical treatment of neurodiverse participants in strict adherence to EU and Cameroonian standards.
Indicative Budget (24 months)
- EU contribution requested: €283,502 (100% funding rate)
- Living Allowance (incl. CCC Denmark 115.5%): €176,022
- Mobility Allowance: €17,040
- Family Allowance (if applicable): €15,840
- Research, Training & Networking (RTN): €24,000
- Management & Indirect Costs: €15,600
- Equipment & Software: €6,000
- Travel & Logistics: €16,000
- Open Access Publications: €8,000
- Contingency: €5,000
Host Institution Fit and Training Plan
The European host institution (University of Copenhagen, UCPH) will provide:
- expertise in cognitive sciences and AI;
- access to computational modeling facilities;
- interdisciplinary supervision under Prof. Daniel Spikol;
- integration into European research networks.
The training plan includes:
- advanced courses in cognitive modeling (Bayesian networks, ABM);
- workshops on AI‑mediated learning;
- a 6‑month secondment (M10–M16) at Northwestern under Prof. Marcelo Worsley for technical prototyping.