Sparse Weight Atlas

Every nonzero MLP weight of a weight-sparse transformer — ablated, profiled, and explained by an executable predicate. Metric of arXiv 2607.02964.

model

–nonzero weights

–scored

–cone robust · headline – passing verdicts

–robust · beats null mean– passing verdicts

–robust · strict (null mean + 2 SD)– passing verdicts

–noise floor

–unverified candidate rules

InterpA is discovery: predicates are directly searched and selected on slice 00, so its score is inflated by selection, not held-out evidence. InterpB and all three robust rules use held-out slices.

cone robust pass scored · cone fail / unavailable noise floor no candidate unverified candidate rule unscored click a cell for its card · T=0.75, c=0.5