Petros Karypis

I'm a PhD student at UC San Diego advised by Julian McAuley. I work on language model pretraining and architecture with a focus on long context, most recently on positional encodings.

Writing

Oct 29, 2026 LeRoPE: letting RoPE learn its own frequencies
LeRoPE learns RoPE's rotation frequencies during pretraining and improves long-context performance. It's now used in Mistral Large 4.