💻 computer science

Seeing Differently: Modeling Interpretive Perspectives in Computational Creativity using a Four-World Framework

This paper proposes a computational framework that models creativity as a perspective-dependent construct by employing three distinct evaluative personas (formalist, social-historical, and iconographic) to demonstrate how interpretive viewpoints systematically shape the assessment of artistic merit and reveal corresponding orientation vectors in visual representation space.

Original authors: Prerna Luthra

Published 2026-08-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Prerna Luthra

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in a massive art gallery. In one corner, a robot is trying to decide which painting is the "most creative." For a long time, scientists building these robots have acted like strict judges with a single, rigid checklist. They ask: "Is it new? Is it surprising? Does it look good?" They assume that if the robot is smart enough, it can find the one true score for every piece of art, just like a thermometer measures the exact temperature of a room.

But here is the twist: humans don't see art that way. When you look at a painting, your brain doesn't just see paint on canvas; it sees a story. If you are a historian, you might see a protest against a king. If you are a poet, you might see a sad love letter. If you are a scientist, you might see a study of light and shadow. The "meaning" of the art isn't stuck inside the frame; it happens in the space between the painting and the person looking at it. This paper asks a big question: If a robot wants to understand creativity, should it try to find one single score, or should it learn to wear different "hats" to see the art through different eyes?

This study, titled "Seeing Differently," dives right into that question. The researchers didn't just ask a robot to guess; they gave it a specific set of tools to test how perspective changes the score. They used a framework called the "Four-World Framework," which breaks creativity down into twelve different traits. Think of these traits as four different rooms in a house: the Inner World (feelings and memories), the Outer World (society and culture), the Imaginative World (surreal dreams and play), and the Moral World (ethics and justice).

To test their theory, the researchers took 1,069 famous European paintings from a digital collection called SemArt. They then programmed a powerful AI (GPT-4.1) to look at these paintings three different times, wearing a different "persona" or hat each time:

The Formalist: This hat focuses only on the visual stuff—colors, shapes, and how the paint is applied. It ignores the history or the story.
The Social-Historian: This hat looks at the painting as a window into the past. It asks, "What does this say about the society, politics, or people who made it?"
The Iconographer: This hat is a detective for symbols. It looks for hidden meanings, religious stories, or secret messages hidden in the objects in the painting.

The AI scored each of the twelve traits for every single painting under each of these three hats. The result? The robot didn't just give slightly different numbers; it gave completely different stories.

The most surprising finding was that the "hat" the robot wore changed the score more than the painting itself did. For example, the trait called Social Reflexivity (how much the art talks about society) got a huge score of 2.60 when the robot wore the Social-Historian hat. But when the same robot looked at the exact same painting wearing the Formalist hat, it gave that same trait a score of only 1.35. That is a massive difference! The robot wasn't confused; it was simply following the rules of the hat it was wearing. The Formalist hat told it to ignore social stories, so it didn't see them. The Social-Historian hat told it to look for them, so it found them.

Some traits were stubborn and didn't change much no matter what hat was worn. Things like Cultural Situatedness (how much the art fits into its culture) got high scores from everyone, ranging only from 3.83 to 3.98. But other traits, like Symbolic Density (how many hidden symbols are packed in), swung wildly depending on who was looking. The Iconographer hat saw a treasure trove of symbols (scoring 3.18), while the Formalist hat saw very few (scoring 2.20).

The researchers also looked under the hood of the AI's brain to see how it was seeing these differences. They found that when the AI switched hats, it didn't just change its mind; it literally looked at different parts of the image. In the mathematical "space" where the AI stores pictures, the Formalist hat pointed in one direction, while the Social-Historian hat pointed in a totally different direction. It's as if the Formalist was looking at the brushstrokes, while the Social-Historian was looking at the shadows to guess the time of day.

The paper suggests that creativity isn't a fixed property of the painting, like its weight or its size. Instead, creativity is a relationship. It emerges when a specific way of looking meets a specific piece of art. The study doesn't claim to have solved the mystery of creativity forever, but it proves that trying to measure creativity with just one ruler is a mistake. If we want computers to be true creative partners to humans, they need to learn to see the world through many different pairs of glasses, not just one. By modeling these different perspectives, we might build systems that can help us understand art—and each other—in richer, more diverse ways.

Technical Summary: Modeling Interpretive Perspectives in Computational Creativity

Problem Statement
Current Computational Creativity (CC) frameworks typically evaluate creative merit as an objective property of artifacts or systems, assuming that a consistent rubric yields stable evaluations across observers. This approach contrasts with longstanding insights from aesthetics and art history, which posit that the meaning of a creative work is not fixed but emerges from the relationship between the artifact and the interpretive perspective applied to it. The central problem addressed is whether computational evaluation systems should model this interpretive diversity rather than collapsing it into a single metric, particularly when human evaluators applying different analytical frameworks systematically disagree on the creative properties of the same artifact.

Methodology
The study employs a mixed-methods approach combining a structured trait framework, large-scale persona-conditioned evaluation, and linear probing of visual embeddings.

Framework and Traits: The research adopts a "Four-World" creativity framework (Luthra 2025), organizing twelve traits across four conceptual domains: Inner World (e.g., Emotional Intensity, Memory Imprint), Outer World (e.g., Cultural Situatedness, Social Reflexivity), Imaginative World (e.g., Surreal Divergence, Symbolic Density), and Moral World (e.g., Ethical Provocation, Collective Resonance). This framework serves as a structured vocabulary for examining interpretive divergence rather than a predictive model of intrinsic creativity.
Dataset: The analysis utilizes 1,069 artworks from the standardized test partition of the SemArt dataset, a corpus of European fine-art paintings containing metadata (title, artist, date, technique, school) and textual descriptions.
Persona-Conditioned Evaluation: Three interpretive perspectives, derived from established art-critical traditions, are operationalized as personas:
- Formalist: Focuses on visual composition, form, color, and stylistic execution.
- Social–Historical: Interprets works through cultural, political, and historical conditions.
- Iconographic: Focuses on symbolic content, narrative meaning, and visual motifs.
  Using GPT-4.1 as a vision–language model, each artwork is evaluated under all three personas. The model assigns scores (0–4) for each of the twelve traits based on the image and accompanying metadata. This process generates 38,484 trait scores.
Statistical Analysis: Non-parametric repeated-measures tests (Friedman rank-sum and Wilcoxon signed-rank) are used to compare trait scores across personas. Kendall's coefficient of concordance ( $W$ ) measures the effect size of persona-conditioned differences.
Representation Space Analysis: To determine if perspectives correspond to distinct visual feature weightings, the study trains ridge regression probes on CLIP (ViT-B/32) image embeddings. These probes map embeddings to persona-specific trait scores. Cosine similarity between the resulting weight vectors quantifies whether different personas prioritize different visual features for the same trait.

Key Results

Systematic Divergence: Creativity assessments vary systematically across perspectives. The Friedman omnibus test is significant for all twelve traits ( $p < 10^{-49}$ ), indicating that the choice of interpretive lens significantly alters scoring.
Trait Sensitivity: Not all traits are equally perspective-sensitive.
- High Sensitivity: Social Reflexivity shows the largest divergence (Kendall's $W = 0.79$ ), with the Social–Historical persona scoring it significantly higher (mean 2.60) than the Formalist (1.35). Symbolic Density and Collective Resonance also show high dependence on perspective.
- Low Sensitivity: Traits like Cultural Situatedness, Environmental Dialogicity, and Emotional Intensity exhibit smaller effect sizes, suggesting they rely more on shared visual cues or conventions.
Boundary Objects: Specific artworks, such as Chardin's Water Glass and Jug, function as "boundary objects" where interpretive communities (simulated by personas) assign radically different scores to the same artifact (e.g., Social Reflexivity scores of 0, 1, and 3 across personas) while agreeing on others (e.g., Emotional Intensity scores of 1 across all).
Orientation in Representation Space: Linear probing reveals that persona-conditioned evaluations correspond to distinct orientation vectors in the CLIP embedding space.
- For high-divergence traits like Social Reflexivity, the cosine similarity between persona vectors is low (average 0.48), indicating that different personas weight distinct visual features to predict the trait.
- For low-divergence traits like Surreal Divergence, vector similarity is high (0.82), suggesting overlapping visual feature reliance.

Significance and Claims
The paper claims to provide empirical support for a relational view of creativity, arguing that creative meaning emerges from the interaction between an artifact, an interpretive framework, and the cultural context, rather than being an intrinsic property of the artifact alone.

The study demonstrates that:

Interpretive Plurality is Tractable: Modeling multiple evaluative perspectives is computationally feasible using persona-conditioned prompts and large-scale datasets.
Divergence is Structured: Disagreements in creativity assessment are not random noise but reflect stable, systematic differences in how different analytical lenses prioritize evidence (e.g., social context vs. visual form).
Computational Mechanism: Interpretive perspectives can be understood as distinct orientations within a shared visual representation space. Different personas do not necessarily alter the underlying representation of the image but apply different weighting directions to predict trait scores.

The author concludes that incorporating multiple evaluative perspectives could enable co-creative systems to better support diverse human collaborators by moving beyond a single objective metric toward a model that acknowledges the role of the interpretive context in defining creative merit. The work positions itself as a step toward operationalizing the "press" or "environment" component of creativity frameworks (e.g., Rhodes's 4Ps, Gl˘aveanu's 5A's) specifically through the dimension of the evaluative perspective.

Technical Summary: Modeling Interpretive Perspectives in Computational Creativity

More like this