Latent Space Optimization
Overview
Latent space optimization maps complex or discrete objects into compact representations where searching for high-performing solutions can be easier and cheaper. Common approaches learn these spaces with variational autoencoders or pretrained language models, then use Bayesian optimization, gradient-based search, evolutionary methods, or constrained and active optimization. Applications include molecular and protein design, robot and physical-system design, neural architecture search, and steering language-model outputs. A central challenge is ensuring that the latent geometry supports useful search while decoded solutions remain valid: representation gaps, imperfect surrogates, and trade-offs between objective quality, diversity, and evaluation cost can limit results.
Latest papers
41 papers9 papers in the last four weeks, up 50% on the four weeks before. 0.1% of all new papers.