3/ 🏥 A separate team at Northeastern located where certain signals live inside Olmo and made targeted edits that reduced biased clinical predictions. This kind of audit is only possible because Olmo exposes all its components.
→ buff.ly/HkChr4Q

And Sheridan Feucht investigates the "implicit vocabulary" of LLMs via token erasure: arxiv.org/abs/2406.20086 (w/David Atkinson and @davidbau.bsky.social)

Somin Wadhwa has some intriguing findings on distillation with "chain of thought" sequences (e.g., this works better when "reasoning" follows labels, and individual tokens seem to be sufficient): arxiv.org/abs/2406.14511 (w/@Silvio Amir)