Edustories: A Collection of Real-world Case Studies from Classroom Practices

Synthesis: Štefánik et al. (2026) argue that AI-in-education research has overwhelmingly targeted individualized student assistance while most teaching worldwide still happens in collective classrooms. Edustories is their corrective dataset: 1,492 teacher-written case studies from elementary and high-school classrooms describing challenging student behavior, the pedagogical interventions teachers attempted, and what followed. Its first application is diagnostic — benchmarking whether language models can predict whether a teacher's intervention succeeded — and the answer is a qualified no: the strongest models reached 58% accuracy against 64% for human experts.

Key Findings

1. The dataset covers collective classroom practice, not one-to-one tutoring: authentic situations with disruptive behavior, teacher intervention and outcome, written by practicing teachers.
2. Current models fall short of human expertise but are not far behind — best-model accuracy of 58% versus 64% for expert assessors across four language-model families.
3. Outcome prediction is framed as a prerequisite capability, not an end in itself: without reliable prediction, teacher-facing feedback cannot be trusted.
4. The authors position teacher-facing assistance as the under-served direction of AI in education, contrasting it with the volume of work on student-facing tutoring.
5. Known limitations are stated plainly — outcome prediction understates the complexity of real deployment, and the contributing teachers were mostly in the final years of their university studies, so the corpus under-represents experienced practitioners.

Why a classroom-level dataset matters

Individualized tutoring is the tractable case for AI: one learner, one dialogue, measurable progress. Collective teaching is harder — the unit of analysis is a group with its own dynamics, the teacher's action space includes management as well as instruction, and success is often judged by what did not happen. Building a corpus around teacher-written narratives gives researchers a substrate for teacher-facing systems, and the accuracy gap the authors measure is precisely the kind of baseline that keeps claims about AI as a teaching assistant honest.

From prediction to assistance

The paper is careful about the distance between predicting an outcome and helping a teacher. Useful assistance would likely require free-form rationales or conversational interfaces rather than a single predicted label, and the authors flag that as future work at the intersection of AI and human-computer interaction. Read alongside the wiki's other work on Teaching and Educational Development, it strengthens the case that the next productive wave of AI-in-education research runs through the teacher's decision rather than the student's answer.

What this means for practice

• Teachers. Write up your classroom interventions in the dataset's five-part structure — situation narrative, student background, the problem, the intervention you applied, and the perceived outcome — so your own cases become comparable practice rather than anecdote.
• Teacher educators. Require short- versus long-term outcome to be labeled in collective reflection sessions, because that is the distinction human experts agree on 82.56% of the time (κ = 0.7356) and the one models fail worst, identifying long-term success only 49.88% of the time against 71.97% for human experts.
• Administrators. Do not deploy a language model to decide whether an intervention worked: the best evaluated model (Qwen-3-30B, 0.580) only reaches the lower bound of expert accuracy (0.573–0.587), and 5 of 6 models fall below it.
• Researchers. Build and test teacher-facing feedback interfaces — free-form rationales or conversational systems — instead of more single-label outcome prediction, which the authors treat as a prerequisite capability rather than the assistive product.

Limitations

• All 1,492 case studies were written by 241 trainee teachers in the final years of their university study at Czech elementary and high schools (2023–2026), so the corpus under-represents experienced practitioners and their strategies.
• Only six locally run models from the Llama 3, Qwen 3, Mistral v0.3 and Olmo 3 families were evaluated; models above 100 billion parameters and proprietary API services were excluded on computational grounds, so the 0.580 ceiling is not a claim about frontier systems.
• Outcome labels are the teachers' own perceived results, and expert judges agreed on only 82.56% of cases (κ = 0.7356), which bounds how accurately any system can be scored against this ground truth.
• The English free-text is machine-translated Czech, and the authors estimate that 5–10% of stories may contain a typo or "translationese" text that no human proofreader verified.

Connected Concepts

• Teaching
• Teacher AI Competency
• Large Language Models (LLMs)
• Educational NLP
• Benchmark
• AI Feedback Quality
• Student Engagement
• Collaborative Learning
• Educational Development
• Quantitative Research

Connected Articles

• I code or AI code: A comparative evaluation of AI-rated scores in classroom observations — AI-rated classroom observation scores
• AI Literacy of Teachers: Prompt Engineering and Model Selection as Predictors of AI-Feedback Quality — Teacher AI literacy and feedback quality
• How AI Is Changing Teaching Workflows — How AI changes teaching workflows
• An Activity-Theoretical Approach to Teacher Professional Development in Pedagogical AI Agent Design — Teacher professional development for AI agents

Citation

Štefánik, M., Nehyba, J., Karasova, J., Fico, M., Škarková, L., Košatková, M., & Kosatka, D. (2026). Edustories: A Collection of Real-world Case Studies from Classroom Practices. arXiv:2609.20484.