Location: IDEAS NCBR, Warsaw, Poland Website: www.mimuw.edu.pl/~pmilos Available projects: 1
I am interested in methods that can deliver a robust decision-making capability in complex scenarios. Such methods can be applied in control to obtain broadly intelligent agents. At the core of this effort is solving credit assignment over prolonged horizons. I like to put, perhaps unorthodoxly, many problems under this umbrella. The tasks requiring extreme reasoning, like math solving, on the one side and classical continuous control in robotics, on the other side.
LLM models generally lack planning capabilities. There are two extremal views on the reason, one advocates that scaling, and proper datasets, could resolve the problem. The other claims that autoregressive objective cannot lead to advanced reasoning. The goal of the project is to make an initial study towards understanding the boarder-line. In the first step, we plan to evaluate how much the performance of a large model (e.g. Llama) will increase on hard planning problems like Rubic’s cube https://twitter.com/laion_ai/status/1714883417383346493.
python, experience in deep learning
experience in llms
Contact: Piotr Milos (pmilos [ at ] mimuw.edu.pl)
See project's page