NLA results surprising, token-level attribution may help

aryaman2020 · x · 2026-08-22

The author found the NLA results unexpected. A suggestion was made to use simple token-level gradient attribution, citing relevant work by @fjzzq2002.

Related event: Studies Question Interpretability Tools: NLA Found Impractical and Misleading(3 posts)→

Original post →

More from Research

Research channel →