Check out our ACL paper! We use shapley interactions to see which words (and phones) interact non-linearly -- what we lose when we assume linear relationships between features. Chat to Diganta in Vienna!
Interested in how in-context learning (ICL) works in LLMs? Curious about improving ICL and why ICL with LM heads may not be optimal? Check out our new #ICLR2025 and #NAACL2025 papers! Huge applause to our PhD student, Hakaze, for leading these exciting projects!👏