Sampling-Based LLM Hypothesis Search Reduces Diversity Collapse

Framing scientific hypothesis generation as a sampling problem rather than evolutionary optimization, this method maintains diversity across generated hypotheses where standard selection-pressure approaches collapse to near-identical outputs. Targeted at settings where validation is expensive and multiple high-quality candidates are needed.