mlsanityData to 14 Oct 2026

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

arXiv 2510.01171 ↗First posted 1 Oct 2025cs.CL cs.AIcited by 105

Not on the map. Why?

Abstract (arXiv)

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 theoretically, verify it on preference datasets empirically, and show that it plays a central role in mode collapse. Motivated by this analysis, we introduce Verbalized Sampling, a simple, training-free prompting strategy to circumvent mode collapse. VS prompts the model to verbalize a probability distribution over a set of responses (e.g., "Generate 5 jokes about coffee and their corresponding probabilities"). 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. For instance, 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. In sum, our work provides a new data-centric perspective on mode collapse and a practical inference-time remedy that helps unlock pre-trained generative diversity.

Key facts. Cited by 105 papers in the map (data through 14 Oct 2026). First posted on arXiv 1 Oct 2025.

Builds on

0 of 0 references.

No references here yet.

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 6; 2026: 99.

105 papers.

Sort:

By area:

  1. 2025Cognitive Alpha Mining via LLM-Driven Code-Based Evolutioncited by 15
  2. 2026When Benign Inputs Lead to Severe Harms: Eliciting Unsafe Unintended Behaviors of Computer-Use Agentscited by 14
  3. 2026Shepherd: Enabling Programmable Meta-Agents via Reversible Agentic Execution Tracescited by 12
  4. 2026SPADE: Self-Play in Adaptive Synthetic Executable Environmentscited by 12
  5. 2026The Illusion of Stochasticity in LLMscited by 9
  6. 2026CREATE: Testing LLMs for Associative Creativitycited by 7
  7. 2026AdaPlanBench: Evaluating Adaptive Planning in Large Language Model Agents under World and User Constraintscited by 7
  8. 2026NOVA: NOise-aware Verbal Confidence CAlibration for Robust Large Language Models in RAG Systemscited by 6
  9. 2026DiscoverLLM: From Executing Intents to Discovering Themcited by 6
  10. 2026Reaching Beyond the Mode: RL for Distributional Reasoning in Language Modelscited by 6
  11. 2026LLMs Should Express Uncertainty Explicitlycited by 6
  12. 2026The Impact of AI-Generated Text on the Internetcited by 6
38 more
  1. 2026Unlocking LLM Creativity in Science through Analogical Reasoningcited by 6
  2. 2026One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RLcited by 6
  3. 2026Verbalizing LLMs' assumptions to explain and control sycophancycited by 5
  4. 2026Annotations Mitigate Post-Training Mode Collapsecited by 5
  5. 2025Reasoning about Intent for Ambiguous Requestscited by 4
  6. 2025MAR:Multi-Agent Reflexion Improves Reasoning Abilities in LLMscited by 4
  7. 2026LLMs Exhibit Significantly Lower Uncertainty in Creative Writing Than Professional Writerscited by 4
  8. 2026The Alignment Tax: Response Homogenization in Aligned LLMs and Its Implications for Uncertainty Estimationcited by 4
  9. 2026FlexSQL: Flexible Exploration and Execution Make Better Text-to-SQL Agentscited by 4
  10. 2025Let the Model Distribute Its Doubt: Confidence Estimation through Verbalized Probability Distributioncited by 3
  11. 2026PRISM: Pushing the Frontier of Deep Think via Process Reward Model-Guided Inferencecited by 3
  12. 2026TurboEvolve: Towards Fast and Robust LLM-Driven Program Evolutioncited by 3
  13. 2026Breaking the Solver Bottleneck: Training Task Generators at the Learnable Frontiercited by 3
  14. 2026Dutch Books for Language Modelscited by 3
  15. 2026An Empirical Study on Preference Tuning Generalization and Diversity Under Domain Shiftcited by 2
  16. 2026Decoding as Optimisation on the Probability Simplex: From Top-K to Top-P (Nucleus) to Best-of-K Samplerscited by 2
  17. 2026PowerFlow: Unlocking the Dual Nature of LLMs via Principled Distribution Matchingcited by 2
  18. 2026Inducing Sustained Creativity and Diversity in Large Language Modelscited by 2
  19. 2026No Single Best Model for Diversity: Learning a Router for Sample Diversitycited by 2
  20. 2026Sampling More, Getting Less: Calibration is the Diversity Bottleneck in LLMscited by 2
  21. 2026AI-Associated Lexical Shifts Across 34 Languages: Cross-Lingual Convergence and Diachronic Uptake in News Writingcited by 2
  22. 2026Simulating Human Memory with Language Modelscited by 2
  23. 2026The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignmentcited by 2
  24. 2026Where You Inject Diversity Matters: A Unified Framework for Diverse Generationcited by 2
  25. 2025QSTN: A Modular Framework for Robust Questionnaire Inference with Large Language Modelscited by 1
  26. 2026Finding RELIEF: Shaping Reasoning Behavior without Reasoning Supervision via Belief Engineeringcited by 1
  27. 2026DIVERGE: Diversity-Enhanced RAG for Open-Ended Information Seekingcited by 1
  28. 2026Not All Layers Need Tuning: Selective Layer Restoration Recovers Diversitycited by 1
  29. 2026Dataset Diversity Metrics and Impact on Classification Modelscited by 1
  30. 2026Auto Researching, not hyperparameter tuning: Convergence Analysis of 10,000 Experimentscited by 1
  31. 2026Self-Consistency for LLM-Based Motion Trajectory Generation and Verificationcited by 1
  32. 2026PromptEvolver: Prompt Inversion through Evolutionary Optimization in Natural-Language Spacecited by 1
  33. 2026Discourse Diversity in Multi-Turn Empathic Dialoguecited by 1
  34. 2026How AI Writing Assistance Distorts Perceptions of AI Users and Their Opinionscited by 1
  35. 2026Alignment Tuning for Large Language Models: A Data-Centric Lens on Alignment Data Pipelinescited by 1
  36. 2026Narrative Flattening: How Post-Training Compresses Thematic, Affective, and Stylistic Variation in LLM Fictioncited by 1
  37. 2026Recovering Diversity Without Losing Alignment: A DPO Recipe for Post-Trained LLMscited by 1
  38. 2026KnowledgeGain: Evaluating and Optimizing Science News Generation for Reader Learningcited by 1

50 most cited of 105 All 105 (JSON)