Discovery under Hypothesis Redundancy: A Geometric Theory of Discovery Bottlenecks

The article presents a study on hybrid discovery systems that integrate structured local search with LLM-generated non-local proposals, introducing the Search Compression Hypothesis. It identifies three geometric conditions necessary for effective non-local exploration: spectral compression, orthogonal escape from the explored span, and residual signal alignment with the target. The findings suggest that while novelty is important, predictive alignment is crucial for improving yield in discovery tasks, impacting how practitioners can optimize LLM-guided discovery processes in various applications.