My interview process changed at several point as the team "figured out what they were looking for." After passing the ML rounds, the recruiter told me they decided to do LC style DS/algorithms coding interviews as well, and to expect a LC style easy/medium problem. Ended up getting a LC hard and promptly rejected.

Chat with recruiter followed by coding phone screen. The phone screen involved 30min of coding where I was asked to implement a directed graph class with topological sort + cycle detection. The interviewer was very opinionated about how I wrote the directed graph class even though I had an equivalent time/memory implementation. I tried to explain it to him but he only seemed to understand one way of implementing it, so I did it his way in the end. The second part was to write topological sort using the data structure. No problem solving, he literally just wanted to know if I've memorized that algorithm or not. I'm sure they'll be able to find someone who's memorized it, but I doubt they'll be a good ML Research Scientist. This was an unreasonable questions given the time constraint in my opinion.