CellFill fine-tunes 4-bit models without changing the checkpoint

CellFill updates a deployed 4-bit model's facts by writing into a frozen per-weight residual, then re-quantizes back to the exact original checkpoint, the CellFill paper reports.

That matters for anyone running quantized models in production. A shipped INT4 checkpoint is normally treated as fixed until someone re-trains and re-releases it. CellFill freezes the integer codes and scales inside every quantization decision cell and writes new information only into that cell's residual, so re-quantizing the patched model returns the released weights bit for bit, a property the paper checks directly rather than assumes. Dropping the residual reverses the update completely.

The method still has to earn its accuracy. Across three paired seeds, it matched an unconstrained fine-tuning baseline on fact recall, 58.9% against the baseline's 59.3%, and it beat that baseline on held-out cross-domain perplexity.

Reversible by design

Most fine-tuning changes every weight it touches and leaves no clean way back. CellFill's residual write means an operator can apply a factual patch, verify the checkpoint hasn't drifted, and pull the patch out later without redeploying from scratch.

Bet on this becoming standard practice for patching quantized checkpoints once it's tested past three seeds and one paper's numbers. Right now, it's the only method here that makes an update to a deployed 4-bit model both machine-checkable and fully reversible.

The paper posted to arXiv in August 2026, as more teams ship quantized models they can't easily re-release just to fix one stale fact.