A Reddit post says SAOD could shrink a 744B model from 1.5TB to under 100GB
pmttyji · reddit · 2026-07-23
A Reddit post highlights Session-Adaptive Orthogonal Distillation (SAOD), claiming it can compress a 744B model from 1.5 TB down to under 100 GB.
The attached image shows a reproduction-style writeup with plots comparing GRPO, RLSD, and related training behavior. It reports that entropy collapses under GRPO while RLSD stays stable, that the credit-clip ratio lands in the paper’s 3%–6% range, that leakage stays at zero across runs, and that a decay schedule choice meaningfully affects performance. The poster suggests the method could make very large MoE models much more feasible on modest VRAM.
More from Research
- Researcher flags unresolved data transparency issues in compositionality evals — najoungkim · 2026-10-27
- Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27
- EMNLP paper probes why some LLM biases survive rewording while others don't — radamihalcea · 2026-10-27
- Datology webinar: Curation Studio matched top data baselines with under 20% compute — josh_wills · 2026-10-27
- KANLab: an interactive browser playground for Kolmogorov-Arnold Networks — Lukas Niehaus · 2026-10-27
- CMU's LLM2Cert synthesizes stability certificates verified via SMT and Lean — CarnegieMellonU · 2026-10-27