An Academic Map of Research That Treats Hallucination as a Resource for Creativity

Among the outputs erased as defects, the seeds of ideation are mixed in

Hallucination has been treated as a defective product to be eliminated. A model states, with an air of confidence, a sentence that has no basis in fact. Lowering that rate is the goal that detection-and-mitigation research has been building toward.

Yet the model does not distinguish between an output that departs from fact and an idea that no one has written down yet. Both are continuations that were probabilistically possible within the training distribution, and there is, in principle, no easy way to erase only one of them. Here a doubt about single-minded suppression arises. If unrealistic outputs are pulled up by the roots, will the novel combinations mixed in among them not fall away as well?

What this map collects is the body of research that has taken up that doubt through empirical work and theory. It is bundled by four questions. Where is the basis for calling hallucination a resource for ideation? How are the creativity and output diversity of LLMs measured, and where do they thin out? What have implementations that turn deviant outputs into a catalyst for ideation made clear? And if diversity thins the more performance is raised, what room remains in the outputs of older generations or of high-temperature sampling?

The fourth question is the core of the request from which this collection originated. But let it be said in advance: research that has answered that question head-on is thin. Peer-reviewed research arguing that “the deviations of older, lower-precision models are precisely what work for ideation” was not found within the scope of this collection. What can be bundled instead is the empirical finding that the performance-raising process (RLHF and SFT) trims diversity, together with the genealogy of metrics for measuring diversity. The perspective of leveraging older generations can only be assembled as an implication of those. Whether that assembly is valid is treated in the final section and in the caveats at the end.

The basis for calling hallucination a resource

What is needed first is a name. Sui et al. renamed the factual deviation of LLMs confabulation¹. Confabulation is a term from clinical psychology, denoting the unwitting fabrication that fills gaps in memory. They showed that this fabrication raises narrativity and semantic coherence on a benchmark, and argued for a parallel with the human sense-making that builds a narrative out of fragments. The naive prediction that deviation lowers coherence is, here, overturned for once.

Once a name is attached, one can carve out what is useful. Chen and Wang formally defined valuable hallucination as a “proposition that is not real but realizable”². They report that, through ReAct-based prompting, they could raise the proportion of the useful side while suppressing overall hallucination (the figures are the source’s own claim). Yang et al. divided it further, constructing a benchmark that separately measures intelligent hallucination and defective hallucination³. They bring the creativity-test criteria of originality, feasibility, and value into the evaluation of factual deviation.

The one who developed this reframing most broadly is the survey by Jiang et al.⁴. Borrowing Guilford’s frame of divergent and convergent thinking, they typologized hallucination as a source of creativity. Attractive as a framework, it is a non-peer-reviewed survey, and one must discount the fact that it is an organizing effort rather than an empirical one.

The approach comes not only from theory but from the side of mechanism. He et al. looked at the decoding layers of a transformer one by one, and quantified hallucination and creativity simultaneously at the same layer⁵. The tradeoff between the two appears consistently regardless of layer depth, model type, or size, and the optimal layer lies further toward the front the larger the model. The suggestion is that deviation and ideation do not occur at separate places, but are two sides of the same computation.

On the applied side, there are cases where deviation directly pushed up accuracy. Yuan et al. observed, across multiple models, that LLM hallucination improves prediction in a molecular-property prediction task⁶. In scientific discovery, the very setting where accuracy would seem to be demanded most, deviation turns into a resource. So’s group extended this line further into a social application, proposing critical confabulation, which fills historical gaps with controlled fabrication⁷. These two, however, are both at the preprint stage and have not passed peer review.

How is creativity measured, and where does it thin out?

To call it a resource, it must be measurable. LLM creativity evaluation began by transplanting divergent-thinking tests from psychology. The wider body of research on using LLMs themselves as evaluators is mapped in Recent Currents in LLM-as-a-Judge: A Literature Map of 53 Core Studies and a Standalone Chapter on Creativity Evaluation. Stevenson et al. subjected GPT-3 to the Alternative Uses Test, making a pioneering attempt to measure originality by semantic distance⁸. Humans held the advantage at the time, but the authors predict a reversal. Bellemare-Pepin et al. compared the latest LLMs with human data on a scale of 100,000 people on the Divergent Association Task, showing that while LLMs surpass the average human, they fall short of the highly creative people at the top⁹.

Why surpass the average yet fail to reach the top? One answer lies in the lack of diversity. Wenger and Kenett quantified that the creative outputs of multiple LLMs resemble one another strikingly, falling well below the diversity found among humans¹⁰. Whichever model one chooses, a similar answer comes back. The warning is that outsourcing creation to LLMs thins ideation as a collective.

At which stage, then, is that diversity lost? Kirk et al. showed that while RLHF raises generalization performance, it significantly lowers output diversity, attributing this chiefly to the mode-seeking nature of reward optimization¹¹. Karouzos et al. went finer, tracing at which stage of post-training diversity collapses, and pinned down across multiple OLMo3 series that the greater part of semantic diversity is lost at the SFT stage¹². The very process that raises performance is trimming diversity.

Here the question that is the starting point of this map connects. If the performance-raising process trims diversity, then outputs that have not gone through that process, or have gone through it only shallowly, should retain their diversity. The view that treats the deviations of older, lower-precision models as a resource stands as this implication. The collected research, however, does not compare diversity along the axis of a model’s newness or age itself. Wright et al. report that newer models have higher epistemic diversity (though lower than a basic web search)¹³, so it cannot be said that older generations are uniformly diverse. “The older the generation, the more it works for ideation” is not yet a proposition backed by empirical evidence; it remains an inference from tradeoff research.

Diversity moves not only with the process but with the sampling at generation time. The min-p sampling of Nguyen et al. is a method that raises diversity at high temperature while preventing quality collapse; it was accepted as an oral at ICLR and adopted into major OSS as well¹⁴. Raise the temperature and deviation increases, but conventionally this came in exchange for quality. Engineering that loosens that terms of exchange has come out of the creativity-support side.

What have implementations that make deviation a catalyst for ideation shown?

After measurement and engineering comes how people use it. Shaer et al. incorporated an LLM into both the divergent and convergent phases of brainwriting, showing in a small-scale group experiment that both the process and the outcomes of ideation improved¹⁵. They also report that the LLM’s evaluation of ideation agreed with the evaluations of experts and novices.

But reports that support instead narrows ideation came out in the same year. Wadinambiarachchi et al. observed in a between-subjects experiment that generative-AI images strengthen design fixation and lower the number, diversity, and originality of ideas¹⁶. Thinking is dragged along by the example shown first. The expectation that deviant outputs broaden ideation reverses if used wrongly.

Beyond short-term effects, an influence remains even after support is withdrawn. Kumar et al. showed, in a preregistered RCT on a scale of 1,100 people, that LLM support raises creativity in the short term but rather lowers autonomous divergent thinking once support is removed, a “hangover effect”¹⁷. Whether it can be used as a resource is not decided by the quality of the output in the moment alone.

Underlying such empirical work is the genealogy of Creativity Support Tools (CST). Shneiderman formulated early on the design principles of tools that support human creativity, placing the collaboration of exploration, variation, and refinement at the center¹⁸. Li et al. carried on this genealogy, reviewing 61 studies on LLM-assisted ideation and deriving a framework that narrows from divergence toward convergence¹⁹.

On the science side, research that supports hypothesis generation itself is emerging. Shahhosseini et al. organized LLM-based scientific ideation into five types and compiled a comprehensive survey that evaluates methods with Boden’s creativity taxonomy and Rhodes’s 4P²⁰. The idea of making deviation a catalyst for ideation extends not only to design and brainstorming but to scientific hypothesis generation.

Can homogenization be pushed back with prompts?

The diagnosis that diversity thins out comes with a prescription attached. Girotra et al. observed that while GPT-4 stands superior to human students in average quality and speed, it simultaneously shows a homogenization in which ideas concentrate in the top 10%²¹. Superiority and homogenization coexist in the same experiment.

The prescription is directed not at changing the model but at changing how one asks. Meincke et al. showed that adding Chain-of-Thought and personas greatly widens the variance of ideas, reaching a diversity close to that of a human group²². The claim is that homogenization is not a fixed property of the model but a variable that moves with how it is elicited. Azad and Baten proposed a framework for evaluating group-level diversity collapse from AI use ex ante against a human baseline, identifying the problem that the latest LLMs fall below human parity²³.

The cognitive-science foundation supporting these lies in the genealogy that grasps creativity as a problem of association. Mednick defined creation as “the combining of remote associations”²⁴, and Guilford proposed divergent thinking, opening modern creativity research²⁵²⁶. Beaty and Kenett showed in an integrative review that highly creative people move more widely through semantic space and bridge more distant concepts²⁷. The body of research that rereads hallucination as “the unconstrained combining of remote concepts” takes this associative theory as its tacit underlay.

Gaps and caveats

What the collected research together illuminates is the tradeoff between performance and diversity. Detection and mitigation trim hallucination; RLHF and SFT trim diversity. On that trimmed-away side were the narrativity of confabulation and the realizable deviation of valuable hallucination: that is the map up to this point.

But the proposition at the core of the request, that “the deviations of older, lower-precision models are precisely what work for ideation,” is not directly supported by this map. Let the reasons be carved out. First, the research on diversity collapse (Kirk, Karouzos) measures along the axis of the training process rather than a model’s newness or age, and research comparing the ideational advantage of older-generation models themselves was not within the scope of the collection. Second, the epistemic diversity of Wright et al. is rather higher in newer models, so the premise that older generations are uniformly diverse does not hold. Third, the research showing the usefulness of deviation (Sui, Chen, Yuan) targets the outputs of current models, not older generations.

The leveraging of older generations, therefore, can only be set up not as a deductive consequence of existing research but as a leap of framing that extends the implications of tradeoff research into a new question. To test that leap, a comparison is needed that varies generation, scale, and temperature on the same task and measures the diversity of ideation and the rate of useful deviation. This map leaves it as a gap that such a comparison has not yet been written. The sequel that reframes this gap through the two lenses of recombination and diversity is the essay on recombination and non-human reasoning.

The references carry full bibliographies (DOIs/URLs) in the footnotes of the main text. Reachability of the bibliographies has been confirmed, but preprints prior to peer review make up much of it, and some items have unconfirmed formal acceptance. These are aggregated at the end.

---

References

Unverified Items

• Yang et al. 2025 (KDD 2026): formal acceptance at the conference is unconfirmed at the preprint stage ([primary-source verification needed]).
• Sui et al. 2025 (Critical Confabulation, ICLR 2026): formal acceptance is unconfirmed ([primary-source verification needed]).
• Girotra et al. 2023: the DOI for journal publication of the SSRN working paper is unconfirmed ([primary-source verification needed]).
• Guilford 1967 (book): DOI/URL reachability is unconfirmed, ISBN only ([primary-source verification needed]).
• Many of the preprints referenced in the main text (Chen & Wang, He et al., Yuan et al., Kirk et al., Karouzos et al., Wright et al., the Chen line, Jiang et al., Li et al., Shahhosseini et al., Meincke et al., Azad & Baten) are pre-peer-review, and their figures and conclusions are treated as the sources’ own claims.
• Of the References above, ²⁸, ²⁹, ³⁰, and ³¹ are related works included in the corpus but not used directly in the argument of the main text (they are positioned at the periphery of the map as reference).

Footnotes

1. Sui, P., Duede, E., Wu, S., & So, R. J. (2024). Confabulation: The Surprising Value of Large Language Model Hallucinations. Proceedings of ACL 2024. https://arxiv.org/abs/2406.04175 (reachability confirmed 2026-07-21) ↩

2. Chen, Q., & Wang, B. (2025). Valuable Hallucinations: Realizable Non-realistic Propositions. arXiv:2502.11113. https://arxiv.org/abs/2502.11113 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

3. Yang, C., Yuan, J., Cai, S., Jiang, J., & Hu, C. (2025). Heaven-Sent or Hell-Bent? Benchmarking the Intelligence and Defectiveness of LLM Hallucinations. arXiv:2512.21635. https://arxiv.org/abs/2512.21635 (reachability confirmed 2026-07-21; formal acceptance at KDD 2026 is [primary-source verification needed]) ↩

4. Jiang, X., Tian, Y., Hua, F., Xu, C., Wang, Y., & Guo, J. (2024). A Survey on Large Language Model Hallucination via a Creativity Perspective. arXiv:2402.06647. https://arxiv.org/abs/2402.06647 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

5. He, Z., Zhang, B., & Cheng, L. (2025). Shakespearean Sparks: The Dance of Hallucination and Creativity in LLMs’ Decoding Layers. arXiv:2503.02851. https://arxiv.org/abs/2503.02851 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

6. Yuan, S., Qu, Z., Kangen, A. Y., & Färber, M. (2025). Can Hallucinations Help? Boosting LLMs for Drug Discovery. arXiv:2501.13824. https://arxiv.org/abs/2501.13824 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

7. Sui, P., Duede, E., Long, H., & So, R. J. (2025). Critical Confabulation: Can LLMs Hallucinate for Social Good? arXiv:2511.07722. https://arxiv.org/abs/2511.07722 (reachability confirmed 2026-07-21; formal acceptance at ICLR 2026 is [primary-source verification needed]) ↩

8. Stevenson, C., Smal, I., Baas, M., Grasman, R., & van der Maas, H. (2022). Putting GPT-3’s Creativity to the (Alternative Uses) Test. Proceedings of ICCC 2022. https://arxiv.org/abs/2206.08932 (reachability confirmed 2026-07-21) ↩

9. Bellemare-Pepin, A., Lespinasse, F., Thölke, P., Harel, Y., Mathewson, K., Olson, J. A., Bengio, Y., & Jerbi, K. (2026). Divergent creativity in humans and large language models. Scientific Reports, 16(1), Article 1279. https://doi.org/10.1038/s41598-025-25157-3 (DOI resolution confirmed 2026-07-21) ↩

10. Wenger, E., & Kenett, Y. N. (2026). Large language models are homogeneously creative. PNAS Nexus, 5(3), pgag042. https://doi.org/10.1093/pnasnexus/pgag042 (DOI resolution confirmed 2026-07-21. Preprint version: arXiv:2501.19361 “We’re Different, We’re the Same: Creative Homogeneity Across LLMs”) ↩

11. Kirk, R., Mediratta, I., Nalmpantis, C., Luketina, J., Hambro, E., Grefenstette, E., & Raileanu, R. (2023). Understanding the Effects of RLHF on LLM Generalisation and Diversity. arXiv:2310.06452. https://arxiv.org/abs/2310.06452 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

12. Karouzos, C., Tan, X., & Aletras, N. (2026). Where does output diversity collapse in post-training? arXiv:2604.16027. https://arxiv.org/abs/2604.16027 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

13. Wright, D., Masud, S., Moore, J., Yadav, S., Antoniak, M., Christensen, P. E., Park, C. Y., & Augenstein, I. (2025). Epistemic Diversity and Knowledge Collapse in Large Language Models. arXiv:2510.04226. https://arxiv.org/abs/2510.04226 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

14. Nguyen, M. N., Baker, A., Neo, C., Roush, A., Kirsch, A., & Shwartz-Ziv, R. (2025). Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs. ICLR 2025 (Oral). https://arxiv.org/abs/2407.01082 (reachability confirmed 2026-07-21) ↩

15. Shaer, O., Cooper, A., Mokryn, O., Kun, A. L., & Ben Shoshan, H. (2024). AI-Augmented Brainwriting: Investigating the use of LLMs in group ideation. Proceedings of CHI 2024. https://doi.org/10.1145/3613904.3642414 (DOI reachability confirmed 2026-07-21) ↩

16. Wadinambiarachchi, S., Kelly, R. M., Pareek, S., Zhou, Q., & Velloso, E. (2024). The Effects of Generative AI on Design Fixation and Divergent Thinking. Proceedings of CHI 2024. https://doi.org/10.1145/3613904.3642919 (DOI reachability confirmed 2026-07-21) ↩

17. Kumar, H., Vincentius, J., Jordan, E., & Anderson, A. (2024). Human Creativity in the Age of LLMs: Randomized Experiments on Divergent and Convergent Thinking. Proceedings of CHI 2025. https://arxiv.org/abs/2410.03703 (reachability confirmed 2026-07-21) ↩

18. Shneiderman, B. (2007). Creativity Support Tools: Accelerating Discovery and Innovation. Communications of the ACM, 50(12), 20–32. https://doi.org/10.1145/1323688.1323689 (DOI reachability confirmed 2026-07-21) ↩

19. Li, S., Padilla, S., Le Bras, P., Dong, J., & Chantler, M. (2025). A Review of LLM-Assisted Ideation. arXiv:2503.00946. https://arxiv.org/abs/2503.00946 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

20. Shahhosseini, F., Marioriyad, A., Momen, A., Baghshah, M. S., Rohban, M. H., & Javanmard, S. H. (2025). Large Language Models for Scientific Idea Generation: A Creativity-Centered Survey. arXiv:2511.07448. https://arxiv.org/abs/2511.07448 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

21. Girotra, K., Meincke, L., Nave, G., Terwiesch, C., & Ulrich, K. T. (2023). Ideas are Dimes a Dozen: Large Language Models for Idea Generation in Innovation. SSRN Working Paper 4526071. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4526071 (reachability confirmed 2026-07-21; DOI for journal publication is [primary-source verification needed]) ↩

22. Meincke, L., Mollick, E. R., & Terwiesch, C. (2024). Prompting Diverse Ideas: Increasing AI Idea Variance. arXiv:2402.01727. https://arxiv.org/abs/2402.01727 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

23. Azad, N. S., & Baten, R. A. (2026). Ex Ante Evaluation of AI-Induced Idea Diversity Collapse. arXiv:2605.06540. https://arxiv.org/abs/2605.06540 (reachability confirmed 2026-07-21; not peer-reviewed) ↩

24. Mednick, S. A. (1962). The Associative Basis of the Creative Process. Psychological Review, 69(3), 220–232. https://doi.org/10.1037/h0048850 (DOI resolution confirmed 2026-07-21) ↩

25. Guilford, J. P. (1950). Creativity. American Psychologist, 5(9), 444–454. https://doi.org/10.1037/h0063487 (DOI resolution confirmed 2026-07-21) ↩

26. Guilford, J. P. (1967). The Nature of Human Intelligence. McGraw-Hill. ISBN 0070251355 (DOI/URL reachability is [primary-source verification needed]) ↩

27. Beaty, R. E., & Kenett, Y. N. (2023). Associative Thinking at the Core of Creativity. Trends in Cognitive Sciences, 27(7), 671–683. https://doi.org/10.1016/j.tics.2023.04.004 (DOI reachability confirmed 2026-07-21) ↩

28. Zhao, Y., et al. (2024). Assessing and Understanding Creativity in Large Language Models. Machine Intelligence Research (2025). https://arxiv.org/abs/2401.12491 (reachability confirmed 2026-07-21) ↩

29. Banerjee, M., Wangsajaya, N. Y., Alsagoff, S. A. R., Tan, M. S., Chun, Z. C. K., & Wei, A. C. G. (2025). Does Less Hallucination Mean Less Creativity? An Empirical Investigation in LLMs. AAAI 2026 Workshop (AI4Research). https://arxiv.org/abs/2512.11509 (reachability confirmed 2026-07-21; workshop paper) ↩

30. Shen, H., Shen, L., Wu, W., & Zhang, K. (2025). IdeationWeb: Tracking the Evolution of Design Ideas in Human-AI Co-Creation. Proceedings of CHI 2025. https://doi.org/10.1145/3706598.3713375 (DOI reachability confirmed 2026-07-21) ↩

31. Guo, S., Shariatmadari, A. H., Xiong, G., Huang, A., Xie, E., Bekiranov, S., & Zhang, A. (2024). IdeaBench: Benchmarking Large Language Models for Research Idea Generation. arXiv:2411.02429. https://arxiv.org/abs/2411.02429 (reachability confirmed 2026-07-21; not peer-reviewed) ↩