research
Two-Sided Curvature Lost a Power of Work
BaKron keeps Kronecker-factored Hessian information while reducing adaptive-rounding cost.
Published Updated Story ID: mp-2026-08-07-016
Summary
BaKron keeps Kronecker-factored Hessian information while reducing adaptive-rounding cost.
For an m-by-n weight matrix, BaKron reduces total work from O(m^2 n^2) to O(mn(m+n)) with O(m+n) sequential steps, matching GPTQ's cubic scaling while retaining richer two-sided curvature information.
Why it matters
BaKron keeps Kronecker-factored Hessian information while reducing adaptive-rounding cost.
Limits and context
No additional limitation was separately recorded.
Key claims
BaKron keeps Kronecker-factored Hessian information while reducing adaptive-rounding cost.
Evidence: source-2026-08-07-018
Sources
- arXiv preprint 2608.06291arXiv · primary research
Corrections
No corrections have been recorded for this story.