In June we presented two papers at the International Conference on Web Engineering (ICWE 2026) in Lyon. On June 11, our full paper received the Best Student Paper Award. It describes the architecture behind the grading logic of ScormIQ.
The award is external proof of something you can't claim for yourself: a jury of web technology researchers picked our method out of 39 submitted papers.
ICWE is one of the established conferences for web architectures and web technologies. The proceedings are published by Springer (LNCS), and selected papers are invited to the Journal of Web Engineering. This year the conference took place at INSA Lyon.
Out of 39 submissions, the program committee accepted 8 full papers into the research track, an acceptance rate of 21%. The Best Student Paper Award goes to exactly one of those papers, and that one is ours: awarded by General Chair Andrea Mauri and Program Chairs Lola Burgueño and Riccardo Tommasini. The research behind it is part of Jonas Gwozdz' PhD, which we support at Netresearch and whose results feed directly into ScormIQ.
The title: Toward Reliable LLM-Integrated Web Architectures for Teacher-Aligned Automatic Student Grading. The question behind it: can a language model take part in grading exams without the teacher handing over control of the judgement?
The answer is a modular web architecture with a feedback loop at its core. Teachers confirm or correct individual gradings, and those confirmed cases become the reference for all following grading cycles. Cases where the model is unsure go back to the teacher automatically. The system's autonomy grows with the evidence: the more confirmed cases exist, the more it grades on its own.
We tested this on real exam data from the computer science bachelor programme at HTWK Leipzig. The best setup reached around 91% agreement with the teachers' gradings, against a majority-class baseline of 56%. The most important result concerns the amount of feedback needed: a handful of confirmed examples is enough to match the grading reliably to what the individual teacher expects. Model size matters less than expected.
The real lever is taking the model's uncertainty seriously. The system only grades on its own when its confidence sits above a defined threshold, for example 80%. Everything below that goes to the teacher for review.
How strict that threshold needs to be depends on the setting. A safety briefing at a driving school can barely afford wrongly approved answers. The budget for mistakes is small, so the threshold is high. A refresher on lecture content can automate more generously because a mistake costs less. Same architecture, two settings. The teacher decides where the line sits, and the system sticks to it.
The full paper wasn't the only contribution. At the ICWE PhD symposium we presented a second paper, Toward Trustworthy, Teacher-Aligned Adaptive Learning on the Web, which sketches the bigger picture of the PhD: adaptive, LLM-supported learning on the web is only useful if teachers can trust the system. Both papers have since been officially published by Springer and are citable via their DOIs (full paper: 10.1007/978-3-032-29372-5_8, PhD symposium: 10.1007/978-3-032-29372-5_20).
On demo day we showed the system live. The questions from the audience after both talks were the productive kind: the role of temperature in model queries, comparing several graders on the same answers, transferability beyond individual courses. Exactly the points where the next iteration starts.
The architecture from the paper isn't an academic exercise. It describes how the grading logic in ScormIQ is built. When a trainer reviews their learners' answers in ScormIQ, they train the system on their specific expectations. Grading then follows their judgement, not a generic model default. The paper delivers the scientific evidence that this approach holds, and it describes how the technology behind it has to be built so the result stays reproducible and verifiable.
For us at Netresearch, the method is the point. We test it scientifically and have it evaluated externally before it goes into a product. What holds up in controlled experiments finds its way into our software.
The next step is already set. Since submitting the paper we've extended the experiment to more modules and are checking whether the method holds across several subjects. Add the open questions from the audience in Lyon, like how different graders relate to each other on the same answers. The award confirms the approach. The real test is generalisation across courses and subjects.
If you want to see what teacher-aligned grading looks like in practice, we'll show you in a ScormIQ demo.