CellFill Teaches a Deployed 4-Bit Model New Facts and Still Re-Quantizes to the Bit-Identical Released Checkpoint

arXiv 2608.20873·medium signalXBlueskyLinkedInCopy link

CellFill (arXiv 2608.20873, Aug 21) writes new knowledge only into the per-weight residual inside each quantization decision cell, with the integer codes and scales of a 4-bit release frozen, so re-quantization returns the released artifact bit for bit as a machine-checkable guarantee and updates are exactly revocable by dropping the residual. Across three paired seeds the constrained path matched an unconstrained reference on fact recall (58.9 versus 59.3 percent, paired difference -0.5 points, 95% CI -5.0 to +4.0) and was better on held-out cross-domain perplexity. It transfers to a 27B hybrid linear-attention model with 2.4e10 constrained weights verified bit-identical, where matched recall costs about half as much cross-domain perplexity as at 1.7B.