arXiv 2510.01171
Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
By Jiayi Zhang, Simon Yu, et al.
Published 2025-10-01
Mindmap
Browse the paper's core ideas, clusters, and relationships in a structured outline.
Post-training alignment often reduces LLM diversity, leading to a phenomenon known as mode collapse. Unlike prior work that attributes this effect to algorithmic limitations, we identify a fundamental, pervasive data-level driver: typicality bias in preference data, whereby annotators systematically favor familiar text as a result of well-established findings in cognitive psychology. We formalize this bias theoretic…