A mysterious behavior in how AI learns has been turned into a predictable law of physics using a concept called "edge coupling."
April 24, 2026
Original Paper
The Origin of Edge of Stability
arXiv · 2604.20446
The Takeaway
Training a neural network often leads to a weird, unstable state that researchers couldn't explain for years. This paper proves that the training process naturally forces the model toward a specific threshold regardless of how it started. By understanding this mathematical limit, engineers can now predict and control the stability of their models. It removes the guesswork from setting learning rates and speeds up the entire development cycle. This turns a dark art of AI training into a precise engineering discipline. Stability is no longer a matter of luck but a matter of calculation.
From the abstract
Full-batch gradient descent on neural networks drives the largest Hessian eigenvalue to the threshold $2/\eta$, where $\eta$ is the learning rate. This phenomenon, the Edge of Stability, has resisted a unified explanation: existing accounts establish self-regulation near the edge but do not explain why the trajectory is forced toward $2/\eta$ from arbitrary initialization. We introduce the edge coupling, a functional on consecutive iterate pairs whose coefficient is uniquely fixed by the gradien