Interleaved Cross-Block Quantization: A New Method for Compressing LLMs to Two Bits

A new paper on arXiv (2608.09595) introduces Interleaved Cross-Block Quantization (ICBQ), a scheduling modification for block-wise post-training quantization (PTQ) of large language models (LLMs). The method targets the fixed two-block setting where neighboring Transformer blocks are reconstructed within a moving window. The standard sequential baseline sweeps the window through the network once, leaving early errors unrevised. ICBQ revisits the boundary pair between consecutive chunks, refining each seam twice: once at the end of a chunk and again at the start of the next. This retains the local two-block objective and reuses existing calibration inputs. Under stated local contraction and smoothness assumptions, the authors derive a depth-wise upper-bound comparison showing that seam revisits multiply the propagated error term while the residual remains bounded. The paper is categorized as a new announcement on arXiv, with the abstract outlining the method's potential to improve compression quality for LLMs at two bits or fewer. The work is relevant to AI technology and model optimization, though it is not yet peer-reviewed.