In the talk he described the progress that has recently been made in Generative Information Retrieval, especially concerning document identifiers, aligning token prediction with relevance ranking, and generalizability. He also presented a large number of challenges concerning Generative Information Retrieval that, hopefully, will inspire future work.

Key take-aways

According to chair of the meeting, Anuj Pathania, the key take-aways were:

  • ‘The ERC Starting grant is not straightforward, even for someone as academically accomplished as Prof. Maarten De Rijke. It is especially important to consider the ethical aspects of your proposals carefully, or it may cost you your chance at ERC.’
  • ‘On the technical side, Generative Information Retrieval research is still in its infancy, and many challenges remain, especially on safety.’

About Generative Information Retrieval

De Rijke: ‘Generative Information Retrieval is a relatively new retrieval paradigm that simplifies the traditional index-retrieve-re-rank pipeline by replacing a large external index with an internal index, i.e., model parameters. Knowledge of all documents in a corpus is encoded into the model's parameters in a way that allows for end-to-end optimization, jointly learning document representations and ranking. For retrieval, generative information retrieval models generate relevant document identifiers using beam search.’

Prof. Maarten de Rijke presenting during the All Hands Meeting September 2025The audience during the AHM September 2025