In-Cell Learning Updates LLMs Without Bit Changes.

A novel paradigm called "in-cell learning" has been introduced, demonstrating that deployed language models can acquire new knowledge without modifying their fundamental stored bits. This approach is particularly relevant for 4-bit quantized models, where each weight is stored as an integer code and a shared scale, leaving a discarded interval between stored values. The core idea is to write new information exclusively into this "in-cell" interval. This ensures that when the served weights are re-quantized, they precisely reproduce the original released codes and scales. The update itself is a separate, reversible "fill" file, allowing for easy withdrawal by subtraction, and drift is strictly bounded by predefined grid radii. The CellFill method implements this paradigm, training a low-rank position within the cell using the vendor's 4-bit release. Experiments on various 4-bit quantized models (NF4, QAT, GPTQ-style grids, Qwen3, Gemma, 1.7B to 31B parameters) showed CellFill successfully injected 83% to 97% of new facts with zero violations across billions of constrained weights. The injected knowledge was demonstrably usable, improving answers to previously unknown questions by 82% to 90% against retrieval methods, with significantly fewer tokens.