Research
[Paper] Graph Machine: Towards Better Pretraining via Edges
GM-2: Ultra-sparse 2–4/4096 dynamic routing replaces most Qwen3-0.6B layers, preserving pretraining loss over 15B tokens.
2026-09-02
[Blog] Graph Machine: Intuitions
Two intuitions for Graph Machine (2), one algorithmic and one biological.
2026-08-24
[Paper] Graph Machine: Exploring Edge Mechanisms as an Inductive Bias
GM-1: Investigating whether explicit, dynamically updatable edges improve relational reasoning on Sudoku.
2026-08-07