Research ideas

Latent Bayesian Discovery of Symbolic Optimizers

Search for a compact symbolic optimizer instead of selecting among fixed AdamW-like formulas. Encode optimizer programs as token sequences, learn a continuous variational representation of those sequences, and use a Gaussian-process Bayesian optimizer to propose promising update rules based on short neural-network training rollouts.