Unbiased activation compression is safe for linear operators in LLMs and preserves convergence rates under L-smoothness, enabling a new co-compression method that reuses low-rank factors for both activations and gradients.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Activation Compression in LLMs: Theoretical Analysis and Efficient Algorithm
Unbiased activation compression is safe for linear operators in LLMs and preserves convergence rates under L-smoothness, enabling a new co-compression method that reuses low-rank factors for both activations and gradients.