Elamur Research⌕
☰
home / papers / 2025 / Shakespearean Sparks: The Dance of Hallucination and Creativity in LLMs' Decoding Layers

Shakespearean Sparks: The Dance of Hallucination and Creativity in LLMs' Decoding Layers

Paper · 2025 · Zicong He; Boxuan Zhang; and Lu Cheng

Summary

Hallucination in LLMs is often said to be linked to creativity, but earlier work treated this link mostly in theory or through qualitative examples. This paper measures it. The authors define creativity narrowly: among sampled answers to a question that has several valid answers, creativity is how many semantically distinct correct answers appear. Hallucination is the share of sampled answers that are factually wrong. Their framework, HCL, uses LayerSkip-style early exit to produce answers from each decoder layer. It samples repeatedly per layer, scores both quantities, and merges them into a hallucination-creativity balanced (HCB) score. Tests on LLaMA 2-7B, LLaMA 2-13B and LLaMA 3-8B with TriviaQA and NQ show a consistent tradeoff: more creativity comes with more hallucination. Each model has an intermediate layer with a favorable balance, and in larger models it tends to sit earlier. Self-rated confidence also peaks at that layer. This suggests that reading the final layer is not always best, and that early exit could save compute while keeping answers both varied and accurate.

Key points

  • HCL has three stages: sampling answers at each layer through early exit (speculative decoding for earlier layers, standard autoregressive decoding for the final one), scoring each answer for correctness and semantic type, and combining the creativity score S_C with the hallucination score S_H into HCB.
  • Confidence per layer is measured with P(True) self-evaluation. In LLaMA 2-13B, earlier layers show lower confidence and more diverse outputs. In LLaMA 2-7B on TriviaQA, confidence spikes sharply at layer 8.
  • LLaMA 2-7B: layer 8 balances the tradeoff best, stable at temperatures 0.6 and 1.0 and on both TriviaQA and NQ, although it is not the top scorer at 0.6.
  • LLaMA 2-13B: layer 4 is the stable choice. LLaMA 3-8B: layer 30 scores highest, but the authors recommend layer 6 for efficient early exit.
  • Code and data are at github.com/ZicongHe2002/HCL-Spark. Stated limitations are the narrow creativity definition (no novelty or originality), the need for questions with known ground-truth answers, and coverage of only a few model sizes.

Interesting ideas

  • I1. Count diversity only among correct outputs, grouped into semantic clusters, so that diversity and accuracy become two separate, measurable axes instead of rewarding random variation.
  • I2. Choose an operating point by combining an error rate and a diversity score into one balance metric, then check that the chosen point is stable across settings (temperature, dataset) rather than picking the single best score.
  • I3. Probe every intermediate layer with early exit; a middle layer may give a better tradeoff than the last one at lower cost.
  • I4. A sharp rise in model self-confidence at one layer can mark where a decision is effectively taken, giving an inexpensive way to locate that layer.
  • I5. Build evaluation sets around questions that have several valid answers, so the benchmark itself separates variety from error.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (answer): School-test and fable-QA readers could report correctness and "number of distinct valid readings" separately, so that an answer set covering several legitimate interpretations is credited without hiding errors.
  • I2 (learn): When tuning silver-data agreement thresholds among witness parsers, pick the threshold that stays near-optimal across treebanks, not the one that peaks on a single dev set.
  • I3 (tree): The gparse transformer parser could be evaluated with arc/label heads attached to intermediate layers, to see whether an earlier exit gives near-equal LAS with faster inference.
  • I4 (learn): Tracking per-layer confidence when querying an LLM teacher could flag uncertain teacher labels, which could then be routed to a higher teacher tier or excluded from silver data.
  • I5 (back): The text → graph → text compressor and generator could be tested on inputs with multiple acceptable paraphrases, scoring how many distinct faithful realizations appear versus how many distort the meaning graph.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Zicong He; Boxuan Zhang; and Lu Cheng. 2025. Shakespearean Sparks: The Dance of Hallucination and Creativity in LLMs' Decoding Layers. arXiv:2503.02851.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2503.02851 · arXiv 2503.02851

@misc{zhang2025shakespearean,
  title = {{Shakespearean Sparks: The Dance of Hallucination and Creativity in LLMs' Decoding Layers}},
  author = {Zicong He; Boxuan Zhang; and Lu Cheng},
  year = {2025},
  doi = {10.48550/arxiv.2503.02851},
  eprint = {2503.02851},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2503.02851}
}

References

38 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901. • Bias & Fairness: LLMs may inherit biases related to gender, ethnicity, and other social factors. Since our framework evaluates hallucination and creativity within existing models, it does not explicitly mitigate bias. Future research should consider fairness-aware approaches to ensure responsible AI deployment.
  2. Tuhin Chakrabarty, Philippe Laban, Divyansh Agarwal, Smaranda Muresan, and Chien-Sheng Wu. 2024. Art or artifice? large language models and the false promise of creativity. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1–34. find
  3. Daixuan Cheng, Shaohan Huang, Junyu Bi, Yuefeng Zhan, Jianfeng Liu, Yujing Wang, Hao Sun, Furu Wei, Denvy Deng, and Qi Zhang. 2023. Uprise: Universal prompt retrieval for improving zero-shot evaluation. arXiv preprint arXiv:2303.08518. • Computational Impact & Efficiency: Our layer-wise analysis and early exit strategies aim to optimize computational efficiency, potentially reducing energy consumption in large-scale model inference. However, running extensive experiments with multiple models still requires substantial computational resources. ↗
  4. Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. 2023. Dola: Decoding by contrasting layers improves factuality in large language models. arXiv preprint arXiv:2309.03883. ↗
  5. Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. 2023. Improving factuality and reasoning in language models through multiagent debate. arXiv preprint arXiv:2305.14325. ↗
  6. Minhyeok Lee. 2023. A mathematical investigation of hallucination and creativity in gpt models. Mathematics, 11(10):2320. ↗
  7. Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625–630. ↗
  8. Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691. ↗
  9. Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Y Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, et al. 2022. Rarr: Researching and revising what language models say, using language models. arXiv preprint arXiv:2210.08726. ↗
  10. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459–9474.
  11. Carlos Gómez-Rodríguez and Paul Williams. 2023. A confederacy of models: A comprehensive evaluation of llms on creative writing. arXiv preprint arXiv:2310.08433. ↗
  12. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36.
  13. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. ↗
  14. Li-Chun Lu, Shou-Jen Chen, Tsung-Min Pai, Chan-Hung Yu, Hung-yi Lee, and Shao-Hua Sun. 2024. Llm discussion: Enhancing the creativity of large language models via discussion framework and role-play. arXiv preprint arXiv:2405.06373. ↗
  15. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276. ↗
  16. Xiaoliang Luo, Akilles Rechardt, Guangzhi Sun, Kevin K Nejad, Felipe Yáñez, Bati Yilmaz, Kangjoo Lee, Alexandra O Cohen, Valentina Borghesani, Anton Pashkov, et al. 2024. Large language models surpass human experts in predicting neuroscience results. Nature human behaviour, pages 1–11. ↗
  17. Ziwei Ji, Tiezheng Yu, Yan Xu, Nayeon Lee, Etsuko Ishii, and Pascale Fung. 2023. Towards mitigating hallucination in large language models via self-reflection. arXiv preprint arXiv:2310.06271. ↗
  18. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2024. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36. find
  19. Xuhui Jiang, Yuxing Tian, Fengrui Hua, Chengjin Xu, Yuanzhuo Wang, and Jian Guo. 2024. A survey on large language model hallucination via a creativity perspective. arXiv preprint arXiv:2402.06647. ↗
  20. Potsawee Manakul, Adian Liusie, and Mark JF Gales. 2023. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896. ↗
  21. Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551. ↗
  22. Niels Mündler, Jingxuan He, Slobodan Jenko, and Martin Vechev. 2023. Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. arXiv preprint arXiv:2305.15852. ↗
  23. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221. ↗
  24. OpenAI. 2024. Learning to reason with llms. https://openai.com/index/ learning-to-reason-with-llms/. Accessed: [02/15/2015]. ↗
  25. Haoqiang Kang, Juntong Ni, and Huaxiu Yao. 2023. Ever: Mitigating hallucination in large language models through real-time verification and rectification. arXiv preprint arXiv:2311.09114. ↗
  26. Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, and Yonatan Belinkov. 2024. Llms know more than they show: On the intrinsic representation of llm hallucinations. arXiv preprint arXiv:2410.02707. ↗
  27. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453– 466. ↗
  28. Max Peeperkorn, Tom Kouwenhoven, Dan Brown, and Anna Jordanous. 2024. Is temperature the creativity parameter of large language models? arXiv preprint arXiv:2405.00492. ↗
  29. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. find
  30. Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821–8831. Pmlr. ↗
  31. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018. Semantically equivalent adversarial rules for debugging nlp models. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (volume 1: long papers), pages 856–865. find
  32. Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Mohamed Amin, Le Hou, Kevin Clark, Stephen R Pfohl, Heather Cole-Lewis, et al. 2025. Toward expert-level medical question answering with large language models. Nature Medicine, pages 1–8. ↗
  33. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288. ↗
  34. Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong. 2024. Solving olympiad geometry without human demonstrations. Nature, 625(7995):476–482. ↗
  35. Elena Vergou, Ioanna Pagouni, Marios Nanos, and Katia Lida Kermanidis. 2023. Readability classification with wikipedia data and all-minilm embeddings. In IFIP International Conference on Artificial Intelligence Applications and Innovations, pages 369–380. Springer. find
  36. Yuxia Wang, Minghan Wang, Hasan Iqbal, Georgi Georgiev, Jiahui Geng, and Preslav Nakov. 2024. Openfactcheck: A unified framework for factuality evaluation of llms. arXiv preprint arXiv:2405.05583. ↗
  37. Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. 2023. Bloomberggpt: A large language model for finance. arXiv preprint arXiv:2303.17564. ↗
  38. Yue Zhou, Barbara Di Eugenio, and Lu Cheng. 2025. Unveiling performance challenges of large language models in low-resource healthcare: A demographic fairness perspective. In Proceedings of the 31st International Conference on Computational Linguistics, pages 7266–7278. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/he-2025-shakespearean-sparks-dance-hallucination-creativity/