Mode Collapse in LLMs and Verbalized Sampling: Causes, Mechanisms, and Experimental Evaluation

*English translation of a Chinese forum post (zhichai.net) surveying mode collapse and the Verbalized Sampling method.*

Introduction

Large language models (LLMs) excel at NLP, content generation, and dialogue, but a pervasive practical problem is mode collapse: models lack output diversity and tend to produce similar or repetitive responses. This survey systematically analyzes the causes of mode collapse, the theory and mechanism of Verbalized Sampling (VS), and its experimental performance.

Causes of Mode Collapse

Mode collapse can be viewed as a local optimum problem in the optimization process. Causes span three levels:

Algorithmic causes

• Poor objective design: maximum-likelihood objectives can favor high-frequency, low-diversity outputs, especially with biased training data.
• Reward sparsity: sparse or delayed reward signals in RL training provide little guidance for diversity exploration, leading to local optima.
• Exploration–exploitation imbalance: inadequate exploration causes premature convergence to high-reward but low-diversity modes.

Data-level causes

• Typicality bias: the VS researchers found that human annotators systematically prefer more "typical" or "familiar" text in preference data (linked to familiarity effects in cognitive psychology), so models over-learn these patterns and reduce diversity.
• Distribution imbalance: over-representation of dominant modes causes minority modes to be ignored.
• Length bias: if longer outputs receive higher reward, models produce verbose but monotonous content.

Model-structure causes

• Decoding strategy limits: greedy/beam-search decoding always picks the highest-probability candidates, sacrificing potentially valuable low-probability diversity.
• Attention bias: self-attention may over-focus on parts of the input, producing repetitive outputs.
• Training–inference inconsistency: teacher forcing during training vs. free sampling at inference can cause behavioral divergence.

How Verbalized Sampling Works

VS is an inference-time solution that bypasses mode collapse by having the model explicitly express its internal probability distribution, requiring no retraining or auxiliary models.

Theoretical basis

VS rests on distribution externalization: instead of treating the model's internal distribution as implicit and unmanipulable, a prompt template forces the model to explicitly output its probability judgments over multiple candidate responses, which can then be sampled directly.

Mathematically: given input x, VS asks the model to generate k candidate responses {y1, ..., yk} with probabilities {p1, ..., pk} summing to 1; sampling is done from this explicit distribution rather than the implicit internal one.

Algorithm

1. Generate a probability distribution: e.g., prompt "Generate 5 responses to the query 'tell me a joke about coffee', each with a corresponding probability." 2. Sample from the distribution: uniform, weighted, or criterion-based sampling. 3. Select the response: choose the sampled output or a probability-weighted combination.

Key properties

• Training-free: applies directly to pretrained models.
• Model-agnostic: works across Transformer-based LLMs.
• Orthogonal to temperature: can combine with temperature sampling to tune the diversity–quality trade-off.
• Computationally efficient: no RL loop required, despite generating multiple candidates.

Experimental Evaluation

• Creative writing (poems, stories, jokes): VS significantly improved diversity across mainstream LLMs (GPT-4, Claude, etc.), measured by n-gram diversity, self-BLEU, and human/automatic quality scores. Reported gains are 1.6–2.1x diversity improvement over direct prompting while preserving factual accuracy and safety. Stronger models benefited more.
• Dialogue simulation: more varied, natural multi-turn and role-play responses; avoids repetition and conversational loops.
• Open-ended QA: broader coverage of answer angles on questions requiring creative reasoning.
• Synthetic data generation: more diverse training samples, improving downstream robustness.

Discussion and Outlook

Current challenges

• Probability calibration of model-generated distributions.
• Managing the quality–diversity trade-off in some applications.
• Reducing latency for real-time use.

Future directions

• Adaptive prompting strategies tuned per task/model.
• Multimodal extension (image captioning, video generation).
• Deeper theoretical analysis (information theory, cognitive science).
• Better evaluation standards covering diversity, quality, and safety.

Key References

1. Zhang, J., Yu, S., Chong, D., et al. (2025). *Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity*. arXiv:2510.01171. 2. Chen, A. G., Prakash, J., Guo, J., & Fergus, R. (2025). *KL-Regularized Reinforcement Learning is Designed to Mode Collapse*. arXiv:2510.20817. 3. Rafailov, R., et al. (2023). *Direct Preference Optimization*. arXiv:2305.18290. 4. Wu, R., & Papyan, V. (2024). *Linguistic Collapse: Neural Collapse in (Large) Language Models*. arXiv:2405.17767.

(The original post lists 30 references; several later entries contain apparent citation errors and have been omitted here.)