# Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

**Authors:** Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael R. Tomz, Christopher D. Manning, Weiyan Shi  
**arXiv:** [2510.01171](https://arxiv.org/abs/2510.01171)  
**Date:** 2025-10-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2510.01171v3/

## Key points

- Authors address the issue of post-training alignment reducing diversity in Large Language Models (LLMs), leading to mode collapse
- Verbalized Sampling is introduced as a strategy to combat mode collapse by prompting the model to verbalize a probability distribution over a set of responses
- Bias in preference data, where annotators favor familiar text, is identified as a fundamental data-level driver of mode collapse
- Extensive experiments across various tasks show that Verbalized Sampling significantly enhances performance without compromising factual accuracy or safety
- More advanced models benefit more from Verbalized Sampling
- This work offers a fresh data-centric perspective on mode collapse and presents an effective inference-time solution through Verbalized Sampling

## Summary

In their paper titled "Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity," authors Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael R. Tomz, Christopher D. Manning, and Weiyan Shi address the issue of post-training alignment reducing diversity in Large Language Models (LLMs), leading to mode collapse. is a straightforward prompting strategy introduced by the authors to combat this issue. It prompts the model to verbalize a probability distribution over a set of responses, such as generating jokes about coffee with corresponding probabilities. While previous studies attribute mode collapse to algorithmic limitations, the authors identify  in preference data as a fundamental data-level driver. This bias stems from annotators favoring familiar text due to . The authors formalize this bias theoretically and validate it on preference datasets empirically, demonstrating its central role in mode collapse. Extensive experiments across various tasks like creative writing (poems, stories, jokes), dialogue simulation, open-ended QA, and synthetic data generation show that VS significantly enhances performance without compromising factual accuracy or safety. For instance,in creative writing tasks,

The authors also observe that more advanced models benefit more from VS. In conclusion,this work offers a fresh data-centric perspective on mode collapse and presents an effective inference-time solution through Verbalized Sampling.By addressing typicality bias in preference data,this approach helps unlock generative diversity in pre-trained models for improved performance across multiple applications.