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

• URL: http://arxiv.org/abs/2510.01171v3
• Date: Fri, 10 Oct 2025 17:38:52 GMT
• Title: Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
• Abstract summary: We identify a fundamental, pervasive data-level driver: typicality bias in preference data.<n>We formalize this bias theoretically, verify it on preference datasets empirically, and show that it plays a central role in mode collapse.<n>We introduce Verbalized Sampling, a training-free prompting strategy to circumvent mode collapse.
• Score: 47.31915660949334
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: 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.

Related papers

• On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity [5.35852583397955]
  On-policy self-distillation achieves strong pass@1 accuracy by using a single model as both teacher and student.<n>We show that this could come at a hidden cost: rollout diversity decreases and pass@k curves flatten.
  arXiv  Detail & Related papers  (2026-06-24T17:59:02Z)
• Bi-directional Bias Attribution: Debiasing Large Language Models without Modifying Prompts [29.864293711943038]
  We propose a framework for detecting stereotype-inducing words and attributing neuron-level bias in large language models.<n>Our framework first identifies stereotype-inducing adjectives and nouns via comparative analysis across demographic groups.<n> Experiments on three widely used LLMs demonstrate that our method effectively reduces bias while preserving overall model performance.
  arXiv  Detail & Related papers  (2026-02-04T10:27:36Z)
• Debiasing Reward Models by Representation Learning with Guarantees [20.170532007501883]
  We propose a principled framework that mitigates spurious correlation issues in reward models.<n>We show that non-spurious latent variables can be theoretically identified from data, regardless of whether a surrogate for the spurious latent variables is available.<n>Experiments on synthetic and real-world datasets demonstrate that our method effectively mitigates spurious correlation issues and yields more robust reward models.
  arXiv  Detail & Related papers  (2025-10-27T18:37:57Z)
• Towards Minimal Causal Representations for Human Multimodal Language Understanding [20.44307628909198]
  We introduce a Causal Multimodal Information Bottleneck (CaMIB) model that leverages causal principles rather than traditional likelihood.<n>To ensure global consistency of causal features, we incorporate an instrumental variable constraint.<n>Experiments on multimodal sentiment analysis, humor detection, and sarcasm detection, along with OOD test sets, demonstrate the effectiveness of CaMIB.
  arXiv  Detail & Related papers  (2025-09-26T03:04:23Z)
• A Unified Virtual Mixture-of-Experts Framework:Enhanced Inference and Hallucination Mitigation in Single-Model System [9.764336669208394]
  Generative models, such as GPT and BERT, have significantly improved performance in tasks like text generation and summarization.<n>However, hallucinations "where models generate non-factual or misleading content" are especially problematic in smaller-scale architectures.<n>We propose a unified Virtual Mixture-of-Experts (MoE) fusion strategy that enhances inference performance and mitigates hallucinations in a single Qwen 1.5 0.5B model.
  arXiv  Detail & Related papers  (2025-04-01T11:38:01Z)
• Chain of Thought Still Thinks Fast: APriCoT Helps with Thinking Slow [0.0]
  We introduce Counterfactual Prompting with Agnostically Primed CoT (APriCoT)<n>APriCoT effectively reduces the influence of base-rate probabilities while improving overall accuracy.<n>Our results suggest that mitigating bias requires a slow thinking process which CoT alone may not provide.
  arXiv  Detail & Related papers  (2024-08-16T10:34:50Z)
• Debiasing Multimodal Large Language Models via Penalization of Language Priors [38.97645845493758]
  Multimodal Large Language Models (MLLMs) have become indispensable tools in computer vision and natural language processing.<n>Despite their advancements, our investigation reveals a noteworthy bias: the generated content is often driven more by the inherent priors of the underlying Large Language Models (LLMs) than by the input image.<n>We propose two simple, training-free strategies to rectify these biases and redirect the model's focus toward visual information.
  arXiv  Detail & Related papers  (2024-03-08T12:35:07Z)
• SA-Attack: Improving Adversarial Transferability of Vision-Language Pre-training Models via Self-Augmentation [56.622250514119294]
  In contrast to white-box adversarial attacks, transfer attacks are more reflective of real-world scenarios. We propose a self-augment-based transfer attack method, termed SA-Attack.
  arXiv  Detail & Related papers  (2023-12-08T09:08:50Z)
• Fairness-guided Few-shot Prompting for Large Language Models [93.05624064699965]
  In-context learning can suffer from high instability due to variations in training examples, example order, and prompt formats. We introduce a metric to evaluate the predictive bias of a fixed prompt against labels or a given attributes. We propose a novel search strategy based on the greedy search to identify the near-optimal prompt for improving the performance of in-context learning.
  arXiv  Detail & Related papers  (2023-03-23T12:28:25Z)
• Delving into Identify-Emphasize Paradigm for Combating Unknown Bias [52.76758938921129]
  We propose an effective bias-conflicting scoring method (ECS) to boost the identification accuracy. We also propose gradient alignment (GA) to balance the contributions of the mined bias-aligned and bias-conflicting samples. Experiments are conducted on multiple datasets in various settings, demonstrating that the proposed solution can mitigate the impact of unknown biases.
  arXiv  Detail & Related papers  (2023-02-22T14:50:24Z)
• Look Beyond Bias with Entropic Adversarial Data Augmentation [4.893694715581673]
  Deep neural networks do not discriminate between spurious and causal patterns, and will only learn the most predictive ones while ignoring the others. Debiasing methods were developed to make networks robust to such spurious biases but require to know in advance if a dataset is biased. In this paper, we argue that such samples should not be necessarily needed because the ''hidden'' causal information is often also contained in biased images.
  arXiv  Detail & Related papers  (2023-01-10T08:25:24Z)
• AES Systems Are Both Overstable And Oversensitive: Explaining Why And Proposing Defenses [66.49753193098356]
  We investigate the reason behind the surprising adversarial brittleness of scoring models. Our results indicate that autoscoring models, despite getting trained as "end-to-end" models, behave like bag-of-words models. We propose detection-based protection models that can detect oversensitivity and overstability causing samples with high accuracies.
  arXiv  Detail & Related papers  (2021-09-24T03:49:38Z)
• Evading the Simplicity Bias: Training a Diverse Set of Models Discovers Solutions with Superior OOD Generalization [93.8373619657239]
  Neural networks trained with SGD were recently shown to rely preferentially on linearly-predictive features. This simplicity bias can explain their lack of robustness out of distribution (OOD) We demonstrate that the simplicity bias can be mitigated and OOD generalization improved.
  arXiv  Detail & Related papers  (2021-05-12T12:12:24Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.