Verbalized Sampling: A Training-Free Method to Mitigate Mode Collapse and Improve LLM Output Diversity

[Submitted on 1 Oct 2025 (v1), last revised 10 Oct 2025 (this version, v3)]

8d ago· 2 min readenInsight

technology science artificial intelligence machine learning research

Summary

This paper identifies a fundamental data-level cause of mode collapse in LLM post-training alignment: typicality bias in preference data, where annotators systematically favor familiar text due to cognitive psychology principles. The authors introduce Verbalized Sampling (VS), a training-free prompting strategy that asks models to verbalize a probability distribution over multiple responses. Experiments show VS improves diversity by 1.6-2.1x in creative writing tasks without sacrificing factual accuracy or safety, with more capable models benefiting more from the approach.

Source

Twitter / XVerbalized Sampling: A Training-Free Method to Mitigate Mode Collapse and Improve LLM Output Diversityarxiv.org

Key quotes

· 5 pulled

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 introduce Verbalized Sampling, a simple, training-free prompting strategy to circumvent mode collapse.

Comprehensive experiments show that VS significantly improves performance across creative writing (poems, stories, jokes), dialogue simulation, open-ended QA, and synthetic data generation, without sacrificing factual accuracy and safety.

In creative writing, VS increases diversity by 1.6-2.1x over direct prompting.

We further observe an emergent trend that more capable models benefit more from VS.

Snippet from the RSS feed

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 prefe

You might also wanna read

Study Finds Multimodal Training Provides Selective, Not Global, Benefits for Human-Like Language Processing

This research paper investigates whether vision-language models (VLMs) produce text representations that are more human-like than large lang

arxiv.org·1mo ago

Research: LLMs Encode Human-Labeled Problem Difficulty Better Than Model-Derived Difficulty

This research paper investigates whether large language models (LLMs) internally encode problem difficulty in alignment with human judgment.

arxiv.org·8mo ago

Comprehensive Survey of Reasoning Failures in Large Language Models

This article presents a comprehensive survey of reasoning failures in Large Language Models (LLMs), introducing a novel categorization frame

arxiv.org·4mo ago

Study Finds AI Discourse in Pretraining Data Creates Self-Fulfilling (Mis)alignment in LLMs

This research paper presents the first controlled study of how pretraining corpora containing discourse about AI systems causally influences

arxiv.org·1mo ago

RICP: A Teacher-Student Framework for Retrieved In-Context Principles from Mistakes in LLMs

This paper introduces Retrieved In-Context Principles (RICP), a novel teacher-student framework for improving Large Language Models (LLMs) t

arxiv.org·1mo ago

Human Conversations Display LLM-Like Failure Modes: Limited Context, Overgeneration, and Hallucination

This reflective essay explores how classic Large Language Model (LLM) failure modes—such as limited context, overgeneration, poor generaliza

embd.cc·5mo ago

Comments

No comments yet. Be the first.