Paper says localized gradients improve language-model data attribution
On the synthetic CATT benchmark, top-token negative log-odds scoring lifts ground-truth recall by 35 to 72 points.
Stories per day · 30 days
1 stories, newest first · one card per story