Cultural Alien Sampler: Open-ended art generation balancing originality and coherence
Alejandro H. Artiles, Hiromu Yakura ∗, Levin Brinkmann ∗, Mar Canet Sola, Hassan Abu Alhaija, Ignacio Serna, Nasim Rahaman, Bernhard Schölkopf, Iyad Rahwan · 2025
Read Cultural Alien Sampler: Open-ended art generation balancing originality and coherence by Alejandro H. Artiles, Hiromu Yakura ∗, Levin Brinkmann ∗, Mar Canet Sola, Hassan Abu Alhaija, Ignacio Serna, Nasim Rahaman, Bernhard Schölkopf, Iyad Rahwan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-20849.txt.
Introduction
Alejandro H. Artiles and eight co-authors present a method for making art ideas. Open ended art needs ideas that are fresh yet still hang together. Large language models tend to fall back on familiar cultural patterns. When pushed for novelty, they often lose inner coherence. Coherence here means the parts of an idea fit together.
Their method, the Cultural Alien Sampler, splits fit from typicality. Typicality means how common an idea is in a culture. One small model scores whether concepts plausibly occur together in art. A second scores how typical those mixes are for one artist. The sampler picks mixes that fit well but seem uncommon. In a human test with one hundred people, it did well. It beat random choice and a GPT-4o baseline. It matched art students on freshness and harmony. A second study found its outputs more varied and broader.
Catalog record
| CREATOR | Alejandro H. Artiles, Hiromu Yakura ∗, Levin Brinkmann ∗, Mar Canet Sola, Hassan Abu Alhaija, Ignacio Serna, Nasim Rahaman, Bernhard Schölkopf, Iyad Rahwan |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 58 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-20849
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Cultural Alien Sampler: Open-ended art generation balancing originality and coherence Alejandro H. Artiles, Hiromu Yakura ∗, Levin Brinkmann ∗, Mar Canet Sola, Hassan Abu Alhaija, Ignacio Serna, Nasim Rahaman, Bernhard Schölkopf, Iyad Rahwan Abstract In open-ended domains like art, autonomous agents must generate ideas that are both original and internally coherent, yet current Large Language Models (LLMs) either default to familiar cultural patterns or sacrifice coherence when pushed toward novelty. We address this by introducing the Cultural Alien Sampler (CAS), a concept-selection method that explicitly separates compositional fit from cultural typicality. CAS uses two GPT-2 models fine-tuned on WikiArt concepts: a Concept Coherence Model that scores whether concepts plausibly co-occur within artworks, and a Cultural Context Model that estimates how typical those combinations are within individual artists’ bodies of work. CAS targets combinations that are high in coherence and low in typicality, yielding ideas that maintain internal consistency while deviating from learned conventions and embedded cultural context. In a human evaluation (N = 100), our approach outperforms random selection and GPT-4o baselines and achieves performance comparable to human art students in both perceived originality and harmony. Additionally, a quantitative study shows that our method produces more diverse outputs and explores a broader conceptual space than its GPT-4o counterpart, demonstrating that artificial cultural alienness can unlock creative potential in autonomous agents. * * footnotetext: Equal contribution Figure 1 : Evolution of generated images across generations 2, 5, and 10 using different sampling methods in the Inspiration Module. Concepts listed below each image were selected by the Prompt Compositor; purple concepts indicate newly introduced elements. 1 Introduction Generative Large Language Models (LLMs) have rapidly evolved into foundational tools for general-purpose problem solving. Paired with agent-based workflows, LLMs can deconstruct complex tasks and integrate solutions. However, their prominent successes remain confined to fields where goals are explicit and correctness is easily measurable such as programming, games, and mathematics novikov2025alphaevolve ; lu2024ai . In contrast, ambiguous, open-ended, or culturally situated tasks, such as creative writing or art, remain underexplored, despite their centrality to human cognition. These domains lack fixed endpoints or universally agreed-upon criteria for success. Instead, they thrive on iterative exploration, contextual sensitivity, and continual generation of ideas. In this work, we address this challenge by reframing art generation as the combinatorial problem of navigating a discrete space of concepts. This approach is inspired by conceptual art theory, which prioritizes abstract ideas over material execution or aesthetic qualities alberro2000conceptual , treating art as the development of an internally consistent system of concepts kosuth1966art . While the notion of conceptual art is broad alberro2000conceptual , we operationalize a narrow definition to leverage a key insight: artistic meaning can arise from combining abstract concepts in deliberate ways. By treating art generation as the strategic recombination of atomic concepts (e.g., "ukiyo-e + machinery + hunter"), we transform an intractable problem into a discrete navigation task that aligns naturally with the associative and combinatorial strengths of LLMs. The emphasis shifts from rendering pixels to exploring an unbounded space of conceptual combinations, where meaning emerges from subjective assessments of consistency rather than objective measures klingemann2022botto . To operate meaningfully within this space, an autonomous agent must balance two often competing criteria of creativity: originality and appropriateness amabile1996creativity ; runco2012standard ; boden2004creative . Originality entails deviating from expectations and introducing novel or surprising combinations. Appropriateness concerns the degree to which an output aligns with its intended context, whether through relevance or utility boyd2005origin ; henrich2015secret , or through intrinsic coherence and harmony arnheim1954art ; kosuth1966art . In our combinatorial framework, appropriateness can be understood as conceptual coherence, the perception that the elements of a composition fit together according to an internal logic arnheim1954art . Importantly, reflecting art’s dialectical nature adorno1997aesthetic , originality and appropriateness exist in productive tension, underscoring its open-endedness. This raises a central question: To what extent can LLMs continually generate ideas that are simultaneously original and intrinsically coherent? On the surface, LLMs seem well-equipped for such a task. Trained on large corpora of human text and culture, they internalize a wide spectrum of concepts, styles, and associations, and frequently produce outputs that appear fluent, relevant, and culturally informed bojic2023gpt ; chiang2024chatbot ; veselovsky2025localized . As a result, in open-ended exploratory tasks within verifiable domains, LLMs have been shown to possess a useful sense of what humans consider novel or interesting zhang2023omni ; faldor2024omni ; lu2024intelligent . However, this strength is also a liability. Because their training data reflects dominant cultural norms and conventions, LLMs are inherently biased toward reproducing familiar patterns bender2021dangers . This cultural anchoring can limit their ability to produce truly original and diverse outputs doshi2024generative ; zhang2025noveltybench , to challenge assumptions and subverting norms. To address this limitation, we introduce the Cultural Alien Sampler (CAS) , a novel method to generate concept combinations designed to enhance the creative exploration capabilities of LLM-based open-ended art agents. CAS comprises two pre-trained GPT-2 models, fine-tuned on a large pool of visual concepts extracted from the WikiArt dataset wikiart . The Concept Coherence Model is trained to predict combinations of concepts that tend to co-occur in artworks. In contrast, the Cultural Context Model estimates the likelihood that specific concept combinations appear within the body of work of individual artists, serving as a proxy for their cultural familiarity or typicality. By encouraging combinations that are compositionally sound but culturally rare, i.e highly ranked but the first model but low by the second, it effectively debiases proposed concept combinations from their cultural context baxandall1988painting . On the one hand, this approach can be situated within the literature on bias mitigation niu2021counterfactual ; ferrara2024fairness . On the other hand, its design aligns with methods that explicitly model an art distribution in order to sample beyond it lisi2020modelling ; elgammal2017can ; wyse2019mechanisms . CAS operates within an iterative open-ended art agent workflow lu2024discovering ; kumar2024automating (see Section 2.1 ). It expands human-provided seed concepts, which GPT-4o converts into prompts for a state-of-the-art image generator. Novelty scores from the resulting images feed back to refine the system’s concept selection in subsequent iterations. To evaluate the effectiveness of the Cultural Alien Sampler, we compare agents using CAS against those using vanilla GPT-4o, randomly sampled concepts, and concepts proposed by human art students. We perform both subjective human evaluations and objective computational analyses. Human raters assess the resulting images on two key criteria: originality and harmony. Our results show that agents guided by CAS consistently outperform GPT-4o and random baselines on both metrics. Moreover, CAS-based agents surpass those guided by human-provided concepts in harmony, and nearly match them in originality. A text-based embedding analysis of the generated ideas further confirms that CAS encourages broader and more diverse exploration across the concept space. In summary, our contributions are as follows: • We propose Cultural Alien Sampler (CAS), a novel selection mechanism that promotes culturally rare but compositionally coherent ideas. • We integrate CAS into a four-stage LLM-based open-ended art agent. • We evaluate CAS against multiple baselines, GPT-4o, random selection, and human-provided inspiration, showing it improves originality and harmony according to human evaluators, while generating more diverse outputs and exploring a broader conceptual space than GPT-4o. 2 Method To simplify the task of generating conceptual