CellFill Enables Bit-Identical, Revocable Updates for Quantized LLMs.

When updating deployed language models, traditional methods like fine-tuning or adapter merging replace the original checkpoint, invalidating previous evaluations and caches. This paper introduces CellFill, a new paradigm for "in-cell learning" that addresses this issue by performing updates within the dequantization gap of 4-bit quantized LLMs. The core idea is to write new knowledge only into the per-weight residual, which remains strictly within each quantization decision cell, leaving the integer codes and scales frozen. This method offers several critical advantages. Firstly, re-quantization returns the model to its original bit-for-bit state, providing a machine-checkable guarantee of invariance. Secondly, updates are exactly revocable by simply dropping the residual. Thirdly, model drift is inherently bounded. The research presents six propositions and three training paths, with CellFill being a bounded reparameterization that structurally ensures this invariance. Experiments show that this constrained approach nearly matches the performance of unconstrained methods in fact recall while significantly reducing cross-domain forgetting. It also transfers effectively to larger models, such as a 27B hybrid linear-attention model, demonstrating that matched recall can be achieved with about half the cross-domain perplexity cost compared to smaller models.