ERC funding for two projects by Basel researchers

Through its Synergy Grants, the ERC supports ambitious research projects that a single research group could not carry out on its own. Two to four researchers and their teams pool their expertise and methods. Funding typically amounts to up to EUR 10 million over six years.

Therapy for blindness

Professor Botond Roska of the IOB and the University of Basel has been awarded one of these interdisciplinary grants together with Valentina Emiliani of the Institut de la Vision in Paris and José-Alain Sahel of the University of Pittsburgh, USA. Their project aims to help people who have lost their sight due to retinal diseases see in greater detail. It builds on a therapy that genetically modifies the remaining nerve cells in the retina to make them sensitive to light. Special glasses convert camera images of the surroundings into light signals that are projected into the eye.

The researchers have already used this therapy to restore rudimentary vision to some people who have lost their sight. However, their vision is still a long way from that of people with healthy eyesight. So far, those treated have not been able to recognize faces or read.

As part of the ERC Synergy project, a team combining expertise in molecular biology, physics, neuroscience and medicine will develop new methods to direct light stimuli more precisely at the modified nerve cells. In parallel, the team is developing a new generation of the therapy with even more sensitive light sensors in the cells. The aim is to make finer details visible and, in the long term, enable people to recognize faces and read.

AI beyond one-size-fits-all

The ERC has also awarded a Starting Grant worth around EUR 1.5 million to Professor Ilija Bogunovic. This brings the total number of Starting Grants awarded to University of Basel researchers in the 2026 funding round to seven. In his project, Bogunovic aims to develop new machine learning methods to help AI language models better meet different needs.

The project focuses on post-training: the final stage of training, in which language models learn from feedback what makes a good answer. But this depends on the language, the task and the person asking. Being as helpful as possible can also clash with safety requirements. Current post-training methods merge these different requirements into a single goal. The result favors the majority, while smaller languages and less common viewpoints receive poorer answers.

Training a separate model for each language or user group would be too costly. The team therefore aims to develop algorithms that keep these requirements separate during post-training. This should allow a single model to balance them and adapt quickly to new languages and individual needs. The team also aims to make this final training stage significantly more efficient.