USC and Yale propose KronQ for state-of-the-art 2-bit LLaMA-3-70B quantization
burkov · x · 2026-07-24
Researchers from USC and Yale introduce KronQ, a post-training quantization framework for large language models.
What it does
- Targets 2-bit weight-only quantization on LLaMA-3-70B
- Incorporates gradient covariance via a Kronecker-factored Hessian
- The authors say it reaches state-of-the-art results for this setting
Why it matters
- Existing methods reportedly fail to converge in this regime
- KronQ is positioned as a more stable way to push LLM compression to very low bit-widths without collapsing optimization
The post links to an AI tutor for reading the work.
More from Research
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Terence Tao's SAIR Foundation revamps Mathathon with $1.5M compute per team — _sathvikr · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27
- PixGR: Pixel-space generative decoding for fast feed-forward 3D reconstruction — kwangmoo_yi · 2026-10-27
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27