Inducing Sustained Creativity and Diversity in Large Language Models
Proposes Recoding-Decoding (RD) to enhance LLMs' sustained creativity and diversity, surpassing modal decoding limits.
Key Findings
Methodology
This paper introduces a simple yet effective Recoding-Decoding (RD) algorithm that injects randomness—such as random phrases and diverting tokens—during decoding. RD leverages the model’s positional bias to explore less traversed knowledge regions, activating latent heterodox, contrarian, and creative knowledge. It involves: random priming, insertion of diverting tokens, and API-based continuous generation, avoiding retraining or internal modifications. Experiments across GPT-5.1, Gemini-3, and various tasks demonstrate significant improvements in content diversity (e.g., geographic coverage expanding from Europe to global, design ideas from conventional to culturally fused) while maintaining high relevance (average scores above 0.98).
Key Results
- In the 'world history battlefields' task, RD generated 1307 unique locations, vastly exceeding OD’s 19, covering Asia, Africa, Australia, and beyond.
- In image-based tasks, RD increased design diversity (e.g., gender-neutral gowns, Mongolian motifs), enhancing collective diversity and maintaining relevance scores of 0.99.
- Across 50 topics, RD outperformed traditional methods by over 3x in diversity metrics, especially with newer models like GPT-5.1 and Gemini-3, showing superior exploration of non-modal knowledge spaces.
Significance
This work breaks through the limitations of modal decoding, unlocking the vast latent knowledge in large models. By guiding models with randomness, it enables continuous, diverse content generation, vastly improving exploratory search efficiency and user satisfaction. It holds promise for personalized recommendation, creative design, and academic discovery, providing a practical tool for expanding AI’s creative potential without internal model modifications.
Technical Contribution
The RD algorithm introduces a novel approach of injecting randomness during decoding, exploiting positional biases to activate non-modal knowledge. Unlike existing diversity methods (top-k, nucleus sampling), RD does not require fine-tuning or internal vector manipulations. It combines API-based simulated continuation with post-processing, ensuring scalable, compatible, and diverse outputs, representing a significant step forward in exploration-oriented language generation.
Novelty
This is the first method to harness model positional bias via random perturbations to explore non-modal, heterodox knowledge without internal model changes. It differs fundamentally from prior work focused on post-training fine-tuning or prompt engineering, offering a lightweight, adaptable solution that unlocks latent creativity and sustains diversity over many outputs.
Limitations
- RD relies on API simulation, which may be limited by API constraints; randomness can sometimes produce irrelevant or erroneous outputs.
- In highly specialized domains, random perturbations might lead to off-topic or inaccurate results, requiring additional filtering.
- Scaling to very large outputs increases computational cost; optimizing efficiency remains an area for future work.
Future Work
Future directions include adaptive perturbation strategies, integrating reinforcement learning for quality control, extending to multi-modal models, and developing more efficient sampling techniques to further enhance exploration and creativity.
AI Executive Summary
This study addresses a fundamental challenge in large language models: their tendency to produce homogeneous, conservative answers due to modal decoding strategies like top-k and nucleus sampling. While effective for factual accuracy, these methods limit the models’ ability to explore the vast, complex knowledge space, especially in exploratory search tasks where diversity and creativity are paramount. To overcome this, the authors propose Recoding-Decoding (RD), a novel algorithm that injects randomness at strategic points during decoding—such as random phrases and diverting tokens—exploiting the model’s positional bias to activate latent, heterodox knowledge. This approach guides the model away from the dominant mode, uncovering unconventional, contrarian, and creative responses.
Extensive experiments across multiple models (GPT-5.1, Gemini-3) and tasks demonstrate RD’s effectiveness. In a 'world history battlefield' task, RD generated 1307 unique locations, covering a broad geographic spectrum, compared to only 19 from traditional methods. In creative design tasks, RD produced a richer variety of wedding dresses, bouquets, and Halloween themes, significantly enhancing collective diversity. Quantitative metrics show that RD increases the number of distinct content clusters by over three times, with newer models benefiting even more. Importantly, relevance scores remain high (above 0.98), indicating that diversity does not come at the expense of pertinence.
This work offers a practical, model-agnostic tool for expanding the exploratory capabilities of large language models. By unlocking the hidden heterodox knowledge, RD paves the way for AI systems that can support more innovative, personalized, and comprehensive search experiences. Future research will refine adaptive perturbation techniques, extend to multi-modal applications, and optimize computational efficiency, further pushing the boundaries of AI-driven exploration and creativity.
Deep Analysis
Background
近年来，随着GPT、BERT等大模型的出现，模态解码策略如top-k和nucleus采样成为主流，确保生成内容的准确性和流畅性。然而，这些方法在探索性任务中表现不足，因其倾向于生成概率最高的答案，忽略了长尾信息，导致内容缺乏多样性和创新性。已有研究尝试微调模型或通过提示工程增强多样性，但难以实现持续、多层次的探索。探索性搜索（如寻找创新点、未被发现的知识空间）需要模型突破模态路径限制，激发潜藏的异端知识，现有方法尚未充分解决这一难题。
Core Problem
当前的解码策略在探索性搜索中存在内容同质化、创新不足的问题。模态解码偏向于生成最可能的答案，忽略长尾信息，限制了内容的多样性和创新性。用户在进行“搜索探索”时，难以获得丰富、异端的答案，影响探索效率。如何在保证内容相关性的同时，激发模型潜藏的非模态知识，成为亟待解决的核心难题。
Innovation
本研究提出RD算法，具有三大创新：1）在解码过程中引入随机扰动（随机短语和偏移符号），引导模型偏离模态路径；2）利用模型偏置特性，激发潜藏的异端、创新知识；3）无需微调模型参数，结合API模拟连续生成，兼容多模型。该方法突破传统多样性增强技术的局限，实现持续、多层次的探索性内容生成，为大模型的创新能力提供新路径。
Methodology
- �� 设计随机引导：在输入前加入随机短语，利用模型偏置引导探索非模态空间；• 插入偏移符号：在每句开始随机插入偏移符号，激发多样性；• API模拟连续生成：结合API接口模拟连续生成，确保内容连贯；• 结合后处理：进行语法校正，确保输出质量；• 多模型验证：在GPT-5.1、Gemini-3等模型上测试，验证多样性提升效果。
Experiments
采用多任务、多模型评估，涵盖历史、设计、创意等领域。使用50个不同话题和500个公开数据集，比较RD与传统OD的多样性、相关性指标。指标包括地理分布、设计多样性、集体差异性等。通过用户评分和自动指标验证，确保内容相关性高（平均0.98以上），同时多样性显著提升（如战场地点从19个增至1307个）。
Results
RD在多任务中显著优于OD，地理覆盖从欧洲扩展到全球，设计多样性提升，集体内容差异性增强。具体表现为：在“世界历史战场”任务中，RD生成1307个不同地点，OD仅19个；在图像生成中，RD带来丰富多彩的设计方案，相关性保持高水平。多模型评估显示，RD在50个话题中，探索空间扩大3倍以上，模型性能越高，优势越明显。
Applications
该算法适用于个性化推荐、创新设计、学术研究等探索性任务。无需模型内部调整，易于集成，提升内容丰富性和创新性。未来可结合多模态、强化学习，推动AI在复杂探索任务中的应用，满足个性化和创新的需求。
Limitations & Outlook
RD依赖API模拟连续生成，可能受API限制影响；随机扰动可能引入偏差或错误；在极端专业领域，内容偏离主题风险增加；计算成本较高，需优化效率。未来将结合自适应扰动策略和多模态扩展，提升稳定性与效率。
Plain Language Accessible to non-experts
想象你在一家厨房里做菜，传统的方法就像用最常用的调料，做出来的菜味道都差不多，缺乏新意。现在，厨师突然加入一些奇怪但合理的调料，比如柠檬汁或辣椒粉，菜肴变得丰富多彩，令人惊喜。这就像RD算法，在模型生成内容时加入随机元素，激发出平时不容易想到的创意和新颖点。这样，模型就像一个有趣的厨师，不仅能做出美味的菜，还能不断创新，带来新鲜感。它打破了传统的套路，让AI变得更有趣、更有创造力，满足人们探索未知的需求。
ELI14 Explained like you're 14
想象你在玩一个超级厉害的机器人，它平时只会说最常见的话，就像每次都说“你好”或“谢谢”。但如果你告诉它加入一些随机的词，比如“火星”或“恐龙”，它就会说出一些奇怪又有趣的句子，比如“火星上的恐龙在跳舞”。这就是RD的魔法：它在生成内容时加入一些随机元素，让机器人说出更多新奇、不同寻常的话。这样，你每次问它问题，都能得到不一样的答案，有趣又有用。它就像一个喜欢冒险的朋友，总是带你去发现新世界！
Abstract
We address a not-widely-recognized subset of exploratory search, where a user sets out on a typically long "search quest" for the perfect wedding dress, overlooked research topic, killer company idea, etc. The first few outputs of current large language models (LLMs) may be helpful but only as a start, since the quest requires learning the search space and evaluating many diverse and creative alternatives along the way. Although LLMs encode an impressive fraction of the world's knowledge, common decoding methods are narrowly optimized for prompts with correct answers and thus return mostly homogeneous and conventional results. Other approaches, including those designed to increase diversity across a small set of answers, start to repeat themselves long before search quest users learn enough to make final choices, or offer a uniform type of "creativity" to every user asking similar questions. We develop a novel, easy-to-implement decoding scheme that induces sustained creativity and diversity in LLMs, producing as many conceptually unique results as desired, even without access to the inner workings of an LLM's vector space. The algorithm unlocks an LLM's vast knowledge, both orthodox and heterodox, well beyond modal decoding paths. With this approach, search quest users can more quickly explore the search space and find satisfying answers.