NoveltyBench: Evaluating Language Models for Humanlike Diversity
NoveltyBench evaluates language models' diversity; 20 models mostly underperform human baseline, larger models show less diversity, highlighting mode collapse issues.
Key Findings
Methodology
This paper introduces NoveltyBench, a benchmark designed to quantitatively assess the diversity of language model outputs. It employs 1,100 prompts, curated into two datasets: NB-CURATED, with prompts crafted to elicit multiple valid responses across four categories—randomness, factual knowledge, creative writing, and subjectivity—and NB-WILDCHAT, sourced from real user interactions. The evaluation framework integrates a functionally equivalent clustering classifier (trained on annotated pairs using DeBERTa-v3-large) to identify distinct output classes, and models user interaction behavior with a geometric patience parameter (p=0.8) to compute cumulative utility (utilityk). The quality of responses is scored via the RewardBench-based Gemma-2-27B reward model, ensuring a robust assessment of both diversity and content quality. The combined metrics (distinctk and utilityk) provide a comprehensive view of a model’s capacity to generate multiple, high-quality, and diverse responses, addressing the limitations of traditional single-output evaluation metrics.
Key Results
- Evaluation of 20 state-of-the-art models revealed that the average number of distinct responses (distinctk) out of 10 samples ranged from 1.76 (Llama-3.1-8B) to 2.88 (GPT-4o), significantly below the human average of approximately 5.5, indicating severe mode collapse. The maximum cumulative utility scores for these models were below 4 out of 10, with GPT-4o scoring 3.27, underscoring the gap between current models and ideal diversity-utility balance.
- A notable inverse correlation was observed between model size and diversity: larger models like Llama-3.3-70B and GPT-4o, despite excelling in traditional benchmarks, produced fewer unique outputs and lower utility scores when multiple responses were considered. Prompt strategies such as in-context regeneration markedly improved diversity, with models like GPT-4o surpassing human performance in cumulative utility under these conditions, demonstrating that diversity can be elicited through specific prompting techniques.
- The experiments highlight that current models, especially larger ones, tend to converge to high-probability modes, limiting their usefulness in subjective or creative tasks where multiple diverse responses are desired. The evaluation framework emphasizes the importance of considering both the number of meaningful alternatives (distinctk) and their quality (utilityk), advocating for a paradigm shift in model assessment that aligns more closely with human preferences and real-world applications.
Significance
This research exposes a critical shortcoming of state-of-the-art language models: their inability to produce sufficiently diverse outputs, which hampers their utility in many practical applications requiring creativity, personalization, and subjective judgment. By establishing NoveltyBench as a standardized, rigorous evaluation tool, the study provides a benchmark for future model development aimed at balancing quality with diversity. This shift is vital for advancing AI systems that better mimic human-like variability, thereby enhancing user engagement and satisfaction in areas like content creation, conversational agents, and personalized recommendation systems. Moreover, the findings challenge the assumption that larger models inherently generate more diverse outputs, prompting a reevaluation of current scaling and training strategies.
Technical Contribution
The paper introduces a novel evaluation paradigm that combines functionally-based output clustering with a user-centric utility model, enabling a comprehensive assessment of diversity and quality. The core technical innovation lies in training a DeBERTa-v3-large classifier to determine functional equivalence between responses, facilitating the partitioning of output space into meaningful classes. The metrics distinctk and utilityk are designed to quantify the number of unique, valuable responses and their cumulative utility, respectively. Additionally, the integration of a reward model (Gemma-2-27B) for quality scoring ensures that diversity is not achieved at the expense of content relevance. This framework advances beyond traditional metrics like BLEU or perplexity, providing a more nuanced and human-aligned evaluation of generative diversity.
Novelty
This work is the first to formalize the concept of functional equivalence-based diversity measurement in language models, moving beyond surface-level lexical or semantic overlap metrics. The combination of a classifier trained on annotated pairs and a user utility model introduces a new standard for evaluating multiple responses simultaneously, emphasizing meaningful variation over trivial paraphrasing. Unlike prior benchmarks that focus solely on single-best outputs, NoveltyBench captures the model’s ability to generate genuinely different, yet still relevant and high-quality, responses, thus addressing a long-standing gap in the evaluation of generative AI systems.
Limitations
- The reliance on manually curated prompts and annotated response pairs introduces potential biases and limits the generalizability of the evaluation framework. The functional equivalence classifier may not perfectly capture all nuances of human judgment, especially in highly subjective or creative tasks.
- The current metrics primarily focus on response diversity and quality, but do not explicitly evaluate other important aspects such as factual consistency, stylistic variation, or contextual appropriateness. Additionally, the evaluation is conducted under idealized sampling conditions (temperature=1), which may differ from real-world deployment scenarios.
- While the framework effectively measures diversity in controlled settings, its scalability to larger, more complex models and diverse application domains remains to be validated. Future work should incorporate broader datasets and real-world user feedback to enhance robustness.
Future Work
Future directions include developing adaptive, multi-dimensional diversity metrics that incorporate stylistic, factual, and contextual factors, as well as integrating reinforcement learning techniques to explicitly optimize for diversity during training. Expanding the dataset to include more real-world prompts and user interactions will improve the ecological validity of the benchmark. Additionally, exploring methods to dynamically balance diversity and relevance based on user preferences could lead to more personalized and engaging AI systems. Finally, extending the evaluation framework to multilingual and multimodal settings will broaden its applicability and impact across diverse AI applications.
AI Executive Summary
The rapid advancement of large language models (LLMs) such as GPT-4 and Llama-3 has revolutionized natural language processing, enabling machines to perform complex reasoning, knowledge retrieval, and conversational tasks with unprecedented fluency. However, despite these impressive capabilities, a persistent challenge remains: the models tend to generate highly repetitive, mode-focused outputs, lacking the diversity and novelty characteristic of human responses. This phenomenon, often termed 'mode collapse,' severely limits the utility of LLMs in applications that demand creativity, personalization, and subjective judgment.
Traditional evaluation metrics—like BLEU, ROUGE, and perplexity—primarily assess the relevance and correctness of a single best output, neglecting the richness of the entire output distribution. Consequently, models optimized for these metrics often produce monotonous responses, failing to meet user expectations for variety and novelty. Recognizing this gap, the authors introduce NoveltyBench, a comprehensive benchmark designed explicitly to measure a model’s capacity to generate multiple, diverse, and high-quality responses.
NoveltyBench employs a carefully curated set of 1,100 prompts, divided into two datasets: NB-CURATED, with prompts crafted to evoke multiple valid answers across four categories—randomness, factual knowledge, creative writing, and subjectivity—and NB-WILDCHAT, derived from real user interactions with ChatGPT. The evaluation framework combines a functionally-based clustering classifier trained on annotated response pairs (using DeBERTa-v3-large) to identify meaningful response classes, with a user model that simulates interaction patience (p=0.8) to compute a cumulative utility score (utilityk). Responses are scored for quality using the RewardBench-based Gemma-2-27B reward model, ensuring that diversity does not come at the expense of content relevance.
Experimental results across 20 state-of-the-art models reveal a concerning trend: most models generate fewer than three distinct responses out of ten samples, with the maximum average distinctk around 2.88 for GPT-4o and 1.76 for Llama-3.1-8B. Interestingly, larger models tend to produce less diverse outputs, contradicting the common belief that bigger models are inherently more capable. The use of prompt engineering techniques like in-context regeneration significantly boosts diversity, even surpassing human performance in some cases, demonstrating that diversity can be elicited through proper guidance.
These findings have profound implications. They highlight that current models, despite their scale and sophistication, are fundamentally limited in their ability to produce varied responses, especially in subjective or creative tasks. This limitation hampers their practical utility in personalized content creation, conversational AI, and other applications where diversity is crucial. The study advocates for a paradigm shift in model training and evaluation—moving beyond single-output relevance metrics to embrace diversity as a core objective.
In conclusion, this research provides a critical lens on the current state of generative AI, emphasizing the need for new benchmarks, training strategies, and prompting techniques that prioritize diversity alongside quality. By doing so, it paves the way for more human-like, engaging, and versatile AI systems capable of truly understanding and reflecting the multifaceted nature of human language and thought.
Deep Dive
Abstract
Language models have demonstrated remarkable capabilities on standard benchmarks, yet they struggle increasingly from mode collapse, the inability to generate diverse and novel outputs. Our work introduces NoveltyBench, a benchmark specifically designed to evaluate the ability of language models to produce multiple distinct and high-quality outputs. NoveltyBench utilizes prompts curated to elicit diverse answers and filtered real-world user queries. Evaluating 20 leading language models, we find that current state-of-the-art systems generate significantly less diversity than human writers. Notably, larger models within a family often exhibit less diversity than their smaller counterparts, challenging the notion that capability on standard benchmarks translates directly to generative utility. While prompting strategies like in-context regeneration can elicit diversity, our findings highlight a fundamental lack of distributional diversity in current models, reducing their utility for users seeking varied responses and suggesting the need for new training and evaluation paradigms that prioritize diversity alongside quality.
References (20)
Evaluating the State-of-the-Art of End-to-End Natural Language Generation: The E2E NLG Challenge
Ondrej Dusek, Jekaterina Novikova, Verena Rieser
Judging LLM-as-a-judge with MT-Bench and Chatbot Arena
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng et al.
Whose Opinions Do Language Models Reflect?
Shibani Santurkar, Esin Durmus, Faisal Ladhak et al.
Simultaneous Translation and Paraphrase for Language Education
Stephen Mayhew, K. Bicknell, Chris Brust et al.
DeBERTa: Decoding-enhanced BERT with Disentangled Attention
Pengcheng He, Xiaodong Liu, Jianfeng Gao et al.
Evaluating the Evaluation of Diversity in Natural Language Generation
Guy Tevet, Jonathan Berant
Generating Diverse Translations with Sentence Codes
Raphael Shu, Hideki Nakayama, Kyunghyun Cho
Boosting Dialog Response Generation
Wenchao Du, A. Black
BERTScore: Evaluating Text Generation with BERT
Tianyi Zhang, Varsha Kishore, Felix Wu et al.
Towards Measuring the Representation of Subjective Global Opinions in Language Models
Esin Durmus, Karina Nyugen, Thomas Liao et al.
(Preprint)
Sarah Verschueren, J. van Aalst, A. Bangels et al.
Hierarchical Neural Story Generation
Angela Fan, M. Lewis, Yann Dauphin
Texygen: A Benchmarking Platform for Text Generation Models
Yaoming Zhu, Sidi Lu, Lei Zheng et al.
Deal or No Deal? End-to-End Learning of Negotiation Dialogues
M. Lewis, Denis Yarats, Yann Dauphin et al.
A Diversity-Promoting Objective Function for Neural Conversation Models
Jiwei Li, Michel Galley, Chris Brockett et al.
Rank-biased precision for measurement of retrieval effectiveness
Alistair Moffat, J. Zobel
ROUGE: A Package for Automatic Evaluation of Summaries
Chin-Yew Lin
Bleu: a Method for Automatic Evaluation of Machine Translation
Kishore Papineni, Salim Roukos, T. Ward et al.
Modeling Human Subjectivity in LLMs Using Explicit and Implicit Human Factors in Personas
Salvatore Giorgi, Tingting Liu, A. Aich et al.
Cited By (20)
The One-Word Census: Answer-Choice Conformity Across 44 Language Models
Jointly Reinforcing Diversity and Quality in Language Model Generations
Inducing Sustained Creativity and Diversity in Large Language Models
Random Policy Valuation is Enough for LLM Reasoning with Verifiable Rewards
"I've Seen How This Goes": Characterizing Diversity via Progressive Conditional Surprise
No Single Best Model for Diversity: Learning a Router for Sample Diversity
G2: Guided Generation for Enhanced Output Diversity in LLMs
Optimizing Diversity and Quality through Base-Aligned Model Collaboration
Enhancing Diversity of LLM-Generated Educational Tasks
What Shapes a Creative Machine Mind? Comprehensively Benchmarking Creativity in Foundation Models
Polychromic Objectives for Reinforcement Learning
LLM-NAS: LLM-driven Hardware-Aware Neural Architecture Search
Death of the Novel(ty): Beyond n-Gram Novelty as a Metric for Textual Creativity
Mode-Conditioning Unlocks Superior Test-Time Scaling
Beware of Reasoning Overconfidence: Pitfalls in the Reasoning Process for Multi-solution Tasks
MindGPT-4ov: An Enhanced MLLM via a Multi-Stage Post-Training Paradigm
String Seed of Thought: Prompting LLMs for Distribution-Faithful and Diverse Generation
The Homogenization Problem in LLMs: Towards Meaningful Diversity in AI Safety
DPWriter: Reinforcement Learning with Diverse Planning Branching for Creative Writing
Language of Thought Shapes Output Diversity in Large Language Models