Synthetic pretraining with RL-trained generators: gradient-overlap reward explained

cephaloform · x · 2026-09-27

A new approach trains LMs on fully synthetic data: the learner uses standard next-token prediction on generator output bytes, while the generator itself is trained with RL. A simple difficulty reward fails (random bytes make output arbitrarily hard without structure), so the reward uses the AdamW-preconditioned gradient inner product with the learner's recent trajectory — matching programs to the learner's learnable frontier. Commenters link it to neural cellular automata training and the fully-synthetic pretraining thesis.

Original post →

More from Models

Models channel →