Do you want proof that humans sometimes generate sentences with low probability?

Here is one:

Four years of PhD have persuaded me that I want to stay in academia.

Because there is no other place where you can do various things simply out of curiosity or a sense of purpose without being motivated by money. (Says a person who just finished reviewing 😅)

Or because you are exposed to the students’ way of thinking, which prevents you from seeing the world too seriously.

Or because nothing can replace the feeling of walking into a university anywhere on this planet and starting to work in a study room on the eduroam wi-fi?

Anyway, I am looking for a post-doc position! 👨‍🎓

And along with that, I would like to start working on some of the questions that I am genuinely curious about:

  • To what extent is the intelligent behavior of LLMs built on solid world models (as opposed to memorization and shallow statistics)?
  • Can we identify the minimal combination of data, training objectives, and model architectures needed to achieve proficiency on specific tasks?
  • Are there problems beyond the scope of current architectures (such as planning ahead with autoregressive decoding)?
  • And how is this all related to the way humans think?

Some of my works already focus on similar questions, investigating the generalization abilities of pretrained LMs and behaviors of open LLMs on the task of data-to-text generation.

In my future research, I would like to move even closer to understanding what the models are doing.

My dream position would be 👇️

  • Within the field of NLP interpretability, low-resource learning, world models, reasoning, knowledge representations, and such,
  • At a European university or research center,
  • Starting from September 2024.

If you know of a position that would be a good fit, please let me know!

And if you wonder whether I am a good candidate:

  • Here is my CV, my Google Scholar page, and the summary of my research projects.
  • Even though I have not (directly) worked on the aforementioned topics, I am deeply familiar with the related NLP and machine learning techniques.
  • I already spent quite some time getting familiar with the literature on these topics.
  • I am very enthusiastic about these topics, ready to learn and experiment.
  • And if you are still not persuaded, I am ready to talk to you!

I am attending ICLR 2024, so I am also open to an in-person meeting in Vienna.