LiftQuant: Continuous Bit-Width LLM via Dimensional Lifting and Projection

In three linesLiftQuant introduces continuous bit-width quantization for LLMs via a "lift-then-project" mechanism that projects a 1-bit lattice from a higher-dimensional lifted space. Effective bit-width is controlled by the dimension ratio, enabling quasi-continuous tuning. A 70B model compressed to 2.4 bits on 24GB GPU outperforms existing 2-bit models.

Summary generated by Claude — human-verified