I recently attended the European Conference on Computer Vision (ECCV) 2026 in Malmo, Sweden and wanted to share some highlights from the research, workshops and discussions I participated in that are relevant for Eedi, and, I believe, for the broader edtech community.

Main Takeaways for Eedi

• Some areas have become generalised and no longer specifically on computer vision. This was reflected in the more general machine learning workshops at the conference which I found very useful.
• The use of Large Language Models, Vision Language Model, and Multimodal Language Models, and the study of what they can and can’t do, was quite a big theme. This was beneficial for Eedi as using an LLM is often an obvious starting point for a project to understand feasibility.
• There was quite a lot of discussion on uncertainty in machine learning models. This is relevant for Eedi as we are interested in understanding when a model is unsure in its prediction, as well as understanding the ability for a model to generalise to new domains.
• There were lots of interesting approaches to tackle safety issues with LLMs, which points to this being a growing area of interest. This is particularly relevant for the EdTech community with AI tutor deployment interacting with school children.

Workshop on Uncertainty Quantification

The workshop on uncertainty quantification in machine learning models is very relevant for Eedi and our ongoing research. It also covered handling out of distribution data, which is something we deal with at Eedi when scaling to new markets. There was also work presented on the generated outputs of LLMs focusing on the language they use. They advocate for judging LLMs by human standards, arguing that LLMs should have anthropomimetic uncertainty in LLM generated content, but that LLMs should not anthropomorphise. This is very relevant for Eedi given our use of LLMs in production.

Curated Data for Efficient Learning Workshop

Data curation and improving our datasets is of importance for Eedi as we strive to build better models. This workshop presented a collection of interesting and different approaches in that area. There was a lot of work utilising large pre-trained models to act as a prior when operating in an area with limited training data. There was also a lot of discussion about evaluating how well LLMs can solve a specific task, which is a relevant approach for Eedi for building faster solutions for new tasks. Finally, there were also talks about efficient LLMs for resource constrained devices, which could be a relevant direction for Eedi in the future if cost becomes a significant factor in the use of LLMs for different markets.

General Research of Interest

I saw multiple research projects working on unlearning, the concept of training a model to not be able to understand certain data. This feels like relevant work for Eedi to explore in the area of student simulators given the concern that the underlying LLM models know all of the answers. There were also many interesting pieces of work around LLM safety showing novel approaches like using input images to improve the safety of LLM generated content. There seems also to be an increase in the amount of research related to more agentic approaches to using LLMs which I found interesting.