Announcement
Tracing language-model outputs to training data in one forward pass
A COLM poster-session presenter says changing model training can impose structure that post hoc attribution methods usually fail to satisfy.
TLDR
A presenter says a language model’s outputs can be traced to its training data in a single forward pass. They plan to discuss the work at COLM Poster Session 1 and say changes to model training can impose structure that post hoc training-data attribution methods usually fail to satisfy.
Combined views
6.9K
7 Sources, first seen ago
