Researchers · Interviewed at COLM 2026
UC San Diego

COLM 2026 · October 8, 2026 · 17 min
Unlike English, DNA has no spaces, no punctuation and no natural word boundaries - so every DNA language model has to decide for itself what counts as a "word". Most borrow BPE from NLP and let frequency decide. But the most frequent sequences are not necessarily the ones that matter biologically.
This episode of Women in AI Research features Nan Huang, PhD student at UC San Diego, presenting EvoLen - a tokenizer that uses cross-species evolutionary signals to decide which DNA segments should become tokens. The model architecture and training objective stay fixed, only the tokenizer changes, and that alone is enough to move the results: