Publications
Papers and posters, written up as the work matures. Negative results included.
Subspace dictionaries across depth: comparing SASA and sparse autoencoders in Gemma 3 1B
Does a model encode why it is uncertain, not just how much? Subspace dictionaries vs. SAEs across all 26 layers of Gemma 3 1B. Both carry uncertainty signal; the split by cause is a dataset confound.
Read the write-up →From commands to invitations: the pragmatic softening of child-directed speech, 1960s–2000s
2.47 million parent tokens from CHILDES, read as an accidental historical corpus. Commands fall 41%, softeners rise 76%, and the structural features that should not move do not.
See the poster and write-up →Impact of network topology on Byzantine resilience in decentralized federated learning
Byzantine-robust aggregators, tested on sparse peer-to-peer graphs instead of the fully connected ones they assume. They break.
Read the write-up →Stanza Extravaganza: improving Latin American Spanish morphological classifiers
Stanford's Stanza mis-tags Spanish verbs that carry clitic pronouns. A clitic-aware preprocessing pass fixes the 18.6% it missed, on child-acquisition corpora from Paraguay and Argentina.
See the poster and write-up →