Novelty Search

Novelty search guides exploration toward outcomes unlike those already found, often using behavioral descriptors and an archive of past discoveries rather than a hand-designed reward. Evolutionary algorithms and quality-diversity methods use distance-based novelty scores, while Bayesian surrogates add uncertainty and noise awareness; extensions include multitask search, learned novelty rewards, and novelty-guided branching for language-model reasoning. These methods are tested in deceptive maze and control tasks, generative systems, optimization landscapes, and scientific design settings such as materials discovery. A central challenge is that novelty can produce diverse but low-quality results, so current work seeks reliable ways to balance novelty with task performance and verify that purportedly new ideas are genuinely useful or distinct.