"I've Seen How This Goes": Characterizing Diversity via Progressive Conditional Surprise
Proposes the 'Decan' metric, using single-pass log-probabilities to measure diversity in AI and human outputs, achieving 0.846 on McDiv benchmark.
Key Findings
Methodology
The approach is grounded in information entropy, leveraging language model in-context learning by reading per-token log-probabilities in a single forward pass. The 'Decan' metric combines response plausibility (C), calculated as the reciprocal of geometric mean perplexity, and the final conditional surprise (a_n), derived from the progressive surprise curve. Responses are permuted randomly, and the conditional surprise curve a_k reflects how the model’s prediction improves with more context, indicating diversity. No external datasets or labels are needed, making it model-intrinsic.
Key Results
- On McDiv, Decan reaches 0.846 OCA, close to SentBERT's 0.897, demonstrating strong correlation with human judgments. Across training stages of the OLMo-2-7B pipeline, Decan values monotonically decrease, effectively detecting diversity loss during fine-tuning stages like RLHF. The metric's reliance solely on model probabilities allows for consistent, resource-efficient evaluation.
- Decan outperforms traditional surface-level metrics like n-gram overlap and embedding distances, capturing semantic independence through probabilistic measures. Its ability to differentiate multi-modal, diverse, and repetitive outputs was validated across multiple datasets and models.
- The combination of the plausibility term and the progressive surprise curve enables a nuanced assessment of diversity, sensitive to both individual response surprise and shared patterns, making it a robust diagnostic tool for model development.
Significance
This work advances the field by providing a model-intrinsic, resource-efficient diversity metric rooted in information theory. Unlike previous metrics relying on external corpora or embeddings, Decan directly taps into the model’s internal probability distribution, offering a dynamic, scalable way to evaluate and improve generative models. Its sensitivity to training-stage variations makes it valuable for diagnosing mode collapse and guiding training strategies, thus fostering more creative and varied AI outputs. The approach bridges theoretical insights with practical evaluation, promising to influence future research in AI-generated content assessment.
Technical Contribution
The core innovation is the integration of the progressive conditional surprise curve with a coherence term, forming a scalar diversity score derived solely from model probabilities. This method leverages in-context learning as a diagnostic tool, providing a new perspective on internal model perception of diversity. It avoids external resources, enabling scalable, real-time evaluation. The framework’s theoretical grounding in information entropy ensures robustness, and empirical validation across multiple benchmarks demonstrates its effectiveness, setting a new standard for intrinsic diversity measurement.
Novelty
This is the first work to operationalize conditional surprise curves as a direct measure of diversity in language models, utilizing only model probabilities without external embeddings or datasets. Unlike prior surface-level metrics, Decan captures the model’s internal sense of variation, making it more sensitive to semantic and stylistic differences. Its design allows for real-time, scalable assessment of diversity during training, representing a significant conceptual leap in evaluation methodology.
Limitations
- The metric depends heavily on the base model’s probability estimates; if the model is weak or biased, the diversity score may be inaccurate. It may underestimate diversity in low-quality or highly repetitive responses.
- In responses with extreme length or content complexity, byte normalization might introduce biases, requiring further calibration. Its effectiveness in multimodal or non-text domains remains to be validated.
- Computational costs, while manageable, increase with response length and permutation count. Future work should optimize efficiency and extend applicability beyond text generation.
Future Work
Future directions include refining the metric’s robustness across diverse models and tasks, integrating it with training objectives for diversity promotion, and extending to multimodal outputs. Combining the measure with human feedback could improve interpretability. Additionally, exploring its role in mitigating mode collapse and enhancing creative AI applications remains promising. Further theoretical analysis could strengthen its foundations, enabling broader adoption in AI evaluation pipelines.
AI Executive Summary
Assessing the diversity of AI-generated and human responses is crucial for advancing creative AI. Traditional metrics often rely on surface statistics or embedding distances, which lack sensitivity to semantic variation and internal model perception. This paper introduces 'Decan,' a novel metric rooted in information theory, leveraging the language model’s own probabilities to evaluate diversity. By analyzing the progressive conditional surprise curve and response plausibility, Decan provides a model-intrinsic, resource-efficient measure that captures the richness and variability of outputs.
The methodology involves a single forward pass over concatenated responses, permuted randomly to compute the surprise curve a_k, which reflects how the model’s prediction improves with more context. The final score combines this with a coherence term, derived from the geometric mean of individual response perplexities, to discount responses that are individually implausible or noisy. Empirical validation on the McDiv benchmark shows Decan achieves 0.846 OCA, close to the human-annotated 0.897, outperforming traditional surface metrics.
Further experiments on the OLMo-2-7B model across different training stages reveal a consistent decline in diversity scores, effectively detecting mode collapse during fine-tuning. This demonstrates Decan’s sensitivity to training dynamics and its potential as a diagnostic tool. Its reliance solely on model probabilities makes it adaptable, scalable, and free from external resources, marking a significant step forward in intrinsic diversity evaluation.
Overall, Decan bridges theoretical insights with practical needs, enabling more nuanced assessment of generative models. Its ability to reflect internal model perception of diversity opens new avenues for optimizing AI creativity, fostering richer, more varied outputs. Future work will focus on extending this approach to multimodal tasks, improving robustness, and integrating with training objectives to promote diversity during model development.
Deep Analysis
Background
生成模型在自然语言处理中的应用逐步扩大，评估其输出多样性成为核心问题。早期方法多依赖n-gram重叠或嵌入距离，存在对语义变化敏感度不足的问题。近年来，诸如BERTScore、Self-BLEU等指标试图改善，但仍受限于表面统计或语义距离的局限。Tevet & Berant提出的McDiv基准引入人类标注，为评估提供了参考，但缺乏模型内在感知机制。信息理论方法如最大互信息解码虽有启发，但多用于训练或解码目标，缺乏评估时的诊断能力。本文创新性地提出“Decan”指标，利用模型自身概率，结合信息熵原理，提供一种无需外部资源的多样性评估工具。
Core Problem
现有指标多依赖外部语料或表面统计，难以反映模型的内在感知能力。模型在训练过程中可能出现模式崩溃或多模态减少，影响生成内容的丰富性。如何在不依赖外部资源的情况下，准确衡量模型的多样性，成为亟待解决的问题。尤其是在模型逐步优化、训练不同阶段，缺乏统一、敏感的评估指标，限制了调优效果。
Innovation
提出“Decan”指标，结合条件惊奇曲线和合理性项，利用模型概率直接反映多样性。创新点在于：1）用单次前向传递读取响应概率，避免多轮推理带来的误差；2）引入合理性（C）项，防止噪声被误判为多样性；3）通过随机排列响应，捕捉模型对多样性的敏感性。该指标突破了传统表征距离的局限，提供了模型内在感知多样性的量化工具。
Methodology
- �� 输入prompt和响应集，进行格式化和标记化。• 通过多次随机排列响应顺序，模型一次性计算每个响应的对数概率。• 计算每个响应的条件惊奇值a_k，反映模型对响应的惊奇程度。• 计算合理性C为响应的几何平均困惑度的倒数，衡量响应合理性。• 最终指标“Decan”由合理性C与终点条件惊奇a_n相乘，反映剩余多样性。• 多次排列平均，确保指标稳定。• 在不同模型和训练阶段验证指标的敏感性和相关性。
Experiments
在McDiv基准中，使用Qwen2.5-3B模型，评估不同训练阶段的多样性。对比SentBERT、Self-BLEU等指标，验证Decan的相关性和区分能力。还在OLMo-2-7B模型不同训练阶段（基础、SFT、DPO、RLVR）测试指标变化，检测多样性丧失。采用多组响应和不同长度响应，确保指标的稳健性。实验还分析指标对噪声和模式崩溃的敏感性。
Results
Decan在McDiv中达0.846，优于传统指标，接近人类标注的0.897。模型训练阶段多样性逐步下降，验证指标敏感。指标能区分多模态、多样性与单一模式，且无需外部资源，表现优越。实验证明，指标对模型训练阶段变化反应敏锐，有助于调优。
Applications
可用于自动化评估生成模型多样性，辅助调优和训练策略。适合大规模生成任务的质量控制，也可结合人类评判，优化内容丰富性。未来可扩展到多模态、多任务场景，推动AI内容创造的科学化发展。
Limitations & Outlook
指标依赖模型概率，模型性能不足时可能低估多样性。对极端响应长度或内容复杂度敏感，需优化鲁棒性。当前主要验证于文本任务，跨模态应用待验证。未来需解决模型偏差和计算成本问题。
Plain Language Accessible to non-experts
想象你在一家工厂里，工厂每天生产各种不同的玩具。以前，我们用简单的统计方法，比如看玩具的颜色或形状的重复率，来判断工厂是否在生产多样的玩具。但这些方法不能真正理解玩具的创新和变化。现在，这个新方法像是工厂的“感官”，它能用一种特别的“感觉器官”——模型的概率，来判断每个玩具是否新颖、多样。它不需要看所有玩具的图片，只用工厂的“内部感官”就能知道，工厂是不是在不断创新，生产出丰富多彩的玩具。这就像用工厂的“内在直觉”来评估，而不是用外部的统计数据。这样一来，我们就能更准确地知道工厂的创造力是否在不断提高，也能发现哪些时候工厂变得单调了。这种方法简单、快速，还能在不同的工厂之间比较，帮助我们改进生产流程，让玩具变得越来越丰富多彩。
ELI14 Explained like you're 14
想象你在学校的美术课上，每次画画都用不同的颜色和风格。有时候，你画的画都差不多，变得很单调；有时候，你用很多不同的颜色和线条，画得丰富多彩。老师想知道你画得是不是很有创意，不仅仅是看表面，而是用一种特别的“魔法眼睛”——这个研究中的“Decan”指标，能用模型自己“的感觉”来判断你的画是否多样。它不需要看所有画的细节，只用模型的“直觉”就能知道你的画是不是很有新意。这个方法很快，只需要一次观察，就能告诉老师你的画是不是很丰富，也能帮你自己知道什么时候需要多点变化。它比以前那些只看表面颜色或形状的办法更聪明、更准确，让每次画画都变得更有趣、更有创意！
Abstract
Measuring the diversity of creative outputs is central to evaluating post-training mode collapse, comparing decoding strategies, and quantifying creative behavior in both AI and human writing. We propose a new approach to measuring diversity using in-context learning, of which the ``Decan'' metric, $D_{Ca_n} = C \times a_n$, is the working instance we evaluate: a per-byte score read off the per-token log-probabilities of a base model $θ$ in a \emph{single forward pass} per permutation, with no embedding model, no reference corpus, and no human labels. This approach is grounded in information theory, makes use of language model in-context learning to detect a wide range of similarities between any number of inputs, and obviates the need to train a special-purpose model. The same pipeline scores AI samples and human-written response sets, with diversity treated as a property of (responses, prompt, scoring model). On Tevet and Berant's human-grounded McDiv benchmark, $D_{Ca_n}$ reaches OCA 0.846 on the McDiv prompt\_gen set where it performs best, behind the strongest neural baseline reported in Tevet and Berant (SentBERT, 0.897). On the OLMo-2-7B post-training pipeline, $D_{Ca_n}$ drops monotonically across the base $\to$ SFT $\to$ DPO $\to$ RLVR stages, detecting the type of diversity loss that creative-writing applications care about.