Cursor | The Hidden Bug in Every Large-Scale RL Run
Numerical mismatch in inference log-probabilities is a hidden bug plaguing large-scale asynchronous RL runs.
“Hopefully next composer versions are going to be our own model instead of basing it off an open source base.”
A Cursor engineer explains how asynchronous RL training forces a re-run of the forward pass to recompute log probabilities, which often diverge from the original even on the same model version—a phenomenon known as inference numerical mismatch. The conversation also signals that future Cursor Composer versions will be trained as Cursor's own model rather than fine-tuned from an open-source base, marking a shift toward proprietary model development.