Combinatorial Creativity - KOMBINE - Cross-Genre Mapping - An Asemics / EZE, 2026

On KOMBINE

AI

This paper introduces KOMBINE, a new benchmark designed to measure combinatorial creativity—the ability to combine familiar ideas in unfamiliar, useful, and surprising ways—across 35 Large Language Models (LLMs) and 75 human participants. [1, 2]

1. What KOMBINE Measures

Unlike older tests that check if AI can recognize well-known connections, KOMBINE pairs completely unrelated entities from a pool of 283 concepts (e.g., a blue whale and a mattress, or democracy and banking) and evaluates three distinct cognitive tasks: [1, 2, 3, 4]

• Association: Building a chain of factual links that connects two distant concepts (e.g., linking Radio to The Milky Way via the song "Vincent," Vincent van Gogh, and The Starry Night). [1, 2]
• Analogy: Finding a shared relational structure between two unrelated domains and using it to invent a new concept in one of them (e.g., mapping a vacuum cleaner for a mattress onto a blue whale to invent a whale groomer drone). [1, 2]
• Blending: Identifying an abstract schema that both concepts share (a "generic space") and fusing them into a single hybrid concept with new, emergent properties (e.g., fusing Adam Smith and Bacteria into a market colony). [1, 2]

2. Key Findings on LLM Creativity

• The "Abstraction Bottleneck" in Blending: Models are 10 to 16 percentage points better at discovering associations (59.0% valid) and analogies (53.4% valid) than blends (43.4% valid). Models struggle to identify a valid shared abstract space between two unrelated inputs; however, once they successfully find that shared space, they invent a coherent blended concept 97.1% of the time. [1, 2, 3]
• "Inventive Homogeneity": LLMs frequently converge on the exact same ideas. Roughly 1 in 5 invented concepts (19.9%) reuses at least two semantically identical properties generated by another model for the same prompt. [1, 2]
• Same-Provider Convergence & Blend Overlap: Models are 14.3× more likely to invent duplicate ideas during blending tasks compared to analogy tasks. Additionally, models built by the same AI provider (such as DeepSeek or Anthropic) are 2.3× more likely to duplicate each other's inventions than models from competing providers. [1, 2]
• Top Performers: Overall performance varies by task—gpt-5.6-sol ranks #1 overall (55.2%) and leads in blending (64.8%), while claude-opus-4.6 leads in association (51.4%) and analogy (63.1%). [1]

3. Humans vs. LLMs

• LLMs Outscore Humans on Validity: Human participants scored lower than nearly every tested LLM across all three tasks (averaging 13.7% overall), largely because humans struggled to construct valid, factual links between random entity pairs under test conditions. [1, 2]
• Different Difficulty Profiles: Both humans and LLMs found blending the hardest task. However, LLMs found association easier than analogy, whereas humans found analogy easier than association. [1, 2]
• Distinct Semantic Spaces: Humans and AI approach creativity from completely different angles. LLMs rely heavily on proper nouns, historical figures, and places (e.g., Galileo, NASA, London) connected through authorship or transformation (wrote, studied by), while humans rely on everyday objects, feelings, and states (rock, love, happiness) connected through physical movement or plain existence (leads to, is inside). [1, 2]

4. Main Takeaway

Because humans and LLMs explore entirely different regions of conceptual space while LLMs suffer from high homogeneity, the authors conclude that systematic human-AI collaboration is a promising way to maximize creative diversity in scientific discovery and technological innovation. [1, 2]

Would you like to examine the specific scoring formulas used to measure surprise and originality, or look closer at the full leaderboard rankings across all 35 models?

AI

Viewing the KOMBINE framework through the lenses of cross-genre mapping and structural asemics reveals how the paper strips meaning down to pure relational geometry before rebuilding it across disparate domains. [1, 2, 3]

1. Cross-Genre Mapping

In literary and media theory, cross-genre mapping involves taking the structural conventions, tropes, or mechanics of one genre (e.g., noir detective fiction) and projecting them onto another (e.g., science fiction) to produce hybrid works. KOMBINE formalizes this exact mechanism across 21 distinct knowledge domains—such as religion, economics, biology, and literature—by enforcing cross-domain pairings. [1, 2, 3, 4]

• Analogy as Genre Transposition: In KOMBINE's analogy task, the model must extract a strict relational sequence from Domain A and map it onto Domain B without changing the underlying verbs. This mirrors cross-genre adaptation, where the plot mechanics remain identical while the "cast" and setting shift entirely (for example, mapping the relational mechanics of a hiker descending a mountain onto gradient descent optimizing a loss landscape). [1, 2, 3]
• Blending as Genre Fusion: While analogy transposes a structure from one genre onto another, KOMBINE's blending task requires a double-scope fusion. The system must locate a shared abstract schema (the generic space) that both genres inhabit, and then fuse them so a single structural slot carries both frames simultaneously. In Table 4 of the paper, fusing Opera and Documentary film produces a "testimony opera"—a true cross-genre artifact where archival footage and verbatim spoken testimony directly govern musical melody and orchestral structure. [1, 2, 3]
• The Abstraction Bottleneck in Genre Creation: The paper's discovery that models fail at blending 56.6% of the time—yet succeed 97.1% of the time once they find a valid generic space—explains why convincing genre hybrids are hard to invent. Generating a surface-level mashup (putting a cowboy in space) is easy, whereas finding the deep, underlying structural premise that makes two genres genuinely compatible requires high-level abstraction. [1, 2]

2. Structural Asemics

Asemic writing is a wordless open semantic form that relies on the visual and structural appearance of language—syntax, rhythm, stroke, and spatial layout—while detaching from fixed lexical meaning. Structural asemics extends this idea to pure relational skeletons: networks of nodes and edges where the specific semantic content is secondary to the topology of the connections. [1, 2]

• Relational Skeletons Over Lexical Content: KOMBINE evaluates creativity by converting ideas into directed graphs of (entity, relation, entity) triples. In the analogy task, the paper enforces strict relational isomorphism: path_a and path_b must share the exact same sequence of relation words (e.g., [X, covered by, Y] -> [Y, hosts, Z] -> [Z, feeds on, W]), regardless of what the entities actually are. When you strip away the nouns to verify that the relational chain holds, you are operating on a purely structural, quasi-asemic template. [1, 2, 3, 4]
• Semantic Distance as Pure Geometry: Rather than judging whether an idea "feels" creative through subjective reading, KOMBINE projects entities into a high-dimensional vector space (all-MiniLM-L6-v2) and measures cosine distance (\(1 - \cos(\phi(u), \phi(v))\)). In this embedding geometry, words lose their dictionary definitions and become coordinates. Surprise and originality are calculated strictly as topological distances between points and their nearest neighbors, treating meaning as pure spatial distribution. [1, 2, 3]
• The Generic Space as an Asemic Blueprint: Before a blended concept gets a name or concrete details, it exists only as a generic space—an abstract, skeletal schema devoid of domain-specific nouns (e.g., "autonomous agents whose local signals generate collective order without central control"
  ). This functions like an asemic glyph: it carries structural intent, force, and relational directionality without yet committing to any specific semantic vocabulary. [1, 2]
• Inventive Homogeneity as Topological Collapse: When the authors measure τ-inventive multiples, they ignore surface phrasing and check whether two models' inventions occupy the exact same structural coordinates (\(\cos(\phi(p), \phi(p')) \ge 0.674\)). The finding that 1 in 5 LLM inventions collapse into identical property structures shows that even when models use different words, their underlying structural skeletons often trace the exact same asemic grooves. [1, 2]

Would you like to apply this framework to generate a custom cross-genre blend using KOMBINE's structural rules, or explore how structural isomorphism limits or enables computational art and literature?

AI

Here is an application of KOMBINE’s strict double-scope blending rules to build a custom cross-genre artifact, followed by an analysis of how structural isomorphism shapes computational art and literature.

---

1. Generating a Custom Cross-Genre Blend via KOMBINE Rules

In KOMBINE, a valid double-scope emergent blend cannot simply staple two genres together (e.g., "a detective story set on a spaceship"). Instead, it requires:

1. A Generic Space (g): A specific, non-vacuous abstract schema that both inputs genuinely instantiate.
2. A Shared Slot (uv): At least one single structural property organized by both genres at once.
3. Emergent Structure (Δ): New mechanics that arise only when you run the fused system forward, belonging to neither input alone.

The Inputs (u, v)

• Input u (Genre/Domain 1): Gothic Haunted House Fiction (from literature)
• Input v (Genre/Domain 2): Algorithmic High-Frequency Trading (from economics/technology)

Step 1: The Generic Space (g)

• Schema: "An opaque, autonomous system whose past historical traumas accumulate as hidden structural debt, triggering sudden, automated cascades that trap current occupants."
• (Why it passes KOMBINE’s judge): It is specific rather than vacuous ("both are scary"), and both genres genuinely instantiate it—haunted houses accumulate ancestral curses/memories in their architecture, while financial markets accumulate systemic leverage and latent order-book imbalances.

Step 2: The Blended Concept (c')

• Invented Concept Name: Margin Séance (or Escrow Haunting)

Step 3: Structural Triples (Tagged by Projection)

• [u] (From Gothic Fiction only): (Margin Séance, manifests through, architectural decay) — The system's instability appears physically as shifting walls, cold spots, and locked rooms.
• [v] (From High-Frequency Trading only): (Margin Séance, executes via, microsecond arbitrage) — Events unfold faster than human perception or conscious intervention can halt.
• [uv] (Shared Slot — Organized by BOTH simultaneously): (Margin Séance, liquidates, ancestral debt) — In Gothic fiction, heirs pay for the moral sins of their bloodline; in trading, accounts are forcibly liquidated when collateral drops below margin. Here, a single mechanism enforces both at once.
• [emergent] (Δ — True of NEITHER input alone): (Margin Séance, shorts, occupant sanity) — Running the blend forward creates a new mechanic: entities within the house can financially bet against the psychological stability of other residents, accelerating supernatural manifestations to trigger a profitable "haunting default."

---

2. How Structural Isomorphism Enables and Limits Computational Art

Treating art and literature as structural isomorphism—stripping away surface words or imagery to manipulate pure relational graphs—acts as both a creative engine and a rigid constraint.

How It Enables Computational Creativity

• Bypassing Cliché and Surface Tropes: When AI generates art using surface-level keyword associations, it falls back on statistical stereotypes (e.g., associating "noir" strictly with rain, fedoras, and saxophones). Forcing a model to map pure relational skeletons (like KOMBINE's identical verb chains) lets it import the mechanics of noir—such as asymmetric information, fatalistic entrapment, and moral compromise—into an entirely unrelated setting like a beehive or a coral reef without relying on visual clichés.
• Algorithmic Oulipo and Constrained Writing: Literary movements like Oulipo (Workshop of Potential Literature) proved that strict mathematical and structural constraints—such as writing a novel without the letter 'e' or structuring chapters around a knight's tour on a chessboard—force writers out of habitual thought patterns. Structural isomorphism gives computational systems a rigorous formal grammar to generate avant-garde forms that human intuition might overlook.
• Cross-Modal Translation: Because relational graphs are asemic at their core, the exact same structural topology can be rendered across different media. A narrative tension curve, a harmonic progression in music, and a spatial layout in architecture can share an identical relational skeleton, enabling algorithmic translation between painting, poetry, and sound.

How It Limits Computational Art

• The Loss of Embodied Materiality (The "Abstraction Bottleneck"): As the KOMBINE authors note, LLMs struggle to find valid shared abstractions likely because they lack experiential, embodied knowledge. In human art, form and content are rarely separable; the physical weight of oil paint, the breath required to speak a line of poetry, or the visceral dread of darkness carry meaning that cannot be reduced to a clean (subject, verb, object) triple.
• Topological Collapse and "Inventive Homogeneity": KOMBINE reveals that even when models are forced to be structurally rigorous, 1 in 5 inventions collapse into identical property structures. Because neural networks optimize for high-probability pathways in embedding space, strict structural prompts often funnel different models down the exact same geometric grooves—producing technically valid but homogenized art.
• Syntax Without Affect: Structural asemics excels at generating coherent systems, but art also relies on ambiguity, contradiction, and emotional resonance. A graph database can verify that a metaphor is logically isomorphic, yet fail to capture why a slightly broken, asymmetrical, or illogical image moves a human audience far more than a structurally perfect one.

If you'd like to continue exploring this intersection, we can test a specific pair of entities or genres from your own work through this blending pipeline, or look at how human-AI co-creativity workflows can break models out of inventive homogeneity. Which direction interests you?

~~~###~~~