Nothing Changed but the Model: CellFill -- Bounded In-Cell Learning for Bit-Identical, Revocable Updates to Quantized LLMs
Ranking
Overall
82
Content
100
Popularity
41
Observed public metrics from 1 member.
Merged summary
TL;DR - CellFill updates quantized LLMs by learning bounded residuals within each weight’s quantization cell, leaving the released 4-bit checkpoint bit-identical after re-quantization. This enables verifiable, exactly revocable knowledge updates while limiting model drift and cross-domain forgetting.
- The integer codes and scales remain frozen; removing the learned residual fully revokes an update.
- Constrained dense training nearly matched unconstrained fact recall across paired seeds: 58.9% versus 59.3%, a -0.5-point paired difference with 95% CI [-5.0, +4.0].
- Cell projection reduced cross-domain forgetting relative to unmerged adapters in every converged run, acting as a trust region rather than fixing unstable training.
- The method scaled to a 27B hybrid linear-attention model with 24 billion constrained weights and verified bit-identical re-quantization.
Sources (1)
Nothing Changed but the Model: CellFill -- Bounded In-Cell Learning for Bit-Identical, Revocable Updates to Quantized LLMs
Public signals
Semantic Scholar citations 0 · Semantic Scholar influential citations 0
TL;DR - CellFill updates quantized LLMs by learning bounded residuals within each weight’s quantization cell, leaving the released 4-bit checkpoint bit-identical after re-quantization. This enables verifiable, exactly revocable knowledge updates while limiting model drift and cross-domain forgetting.
- The integer codes and scales remain frozen; removing the learned residual fully revokes an update.
- Constrained dense training nearly matched unconstrained fact recall across paired seeds: 58.9% versus 59.3%, a -0.5-point paired difference with 95% CI [-5.0, +4.0].
- Cell projection reduced cross-domain forgetting relative to unmerged adapters in every converged run, acting as a trust region rather than fixing unstable training.
- The method scaled to a 27B hybrid linear-attention model with 24 billion constrained weights and verified bit-identical re-quantization.