Curvature-guided compression avoids rotating model weights
In plain words
This paper introduces CurveTQ, a way to shrink the stored weights of large language models to about two or three bits each.
Instead of rotating the weights before compressing them, CurveTQ uses information about which weight errors matter more to guide its search, and stores a small starting-state marker for each block.
On three tested smaller Instruct models, it scored higher on average than two competing codecs given the same kind of marker. The result applies to those models and tests, not every model.
The research
At very low bit rates, compressing model weights can reduce storage, but the compared trellis codecs rotate the weights and must undo that transform during generation.
Can a codec use information about which weight errors matter more to avoid rotation without giving up accuracy on language-model tasks?
CurveTQ weights its trellis search by curvature, aiming to compress model weights without the rotation used by competing codecs.
The authors compared CurveTQ with QTIP and Proteus after giving those codecs the same stored block-start information, which itself helps accuracy.
At two bits per weight, CurveTQ scored 0.9 to 3.1 percentage points higher in mean task accuracy across three 4–8B Instruct models; five of six paired comparisons excluded zero.
A separate rotation-by-search comparison found that curvature weighting helped in native coordinates but added little after rotation. This supports the view that weighting can replace rotation in the tested conditions, while the full codec comparison also reflects other design choices.
These comparisons cover the tested checkpoints, calibration procedure and task suites, with no post-quantization fine-tuning. The boundary-state control was run only on Qwen3-4B and also changed the coding rate, so it does not isolate the state’s effect at equal rate.