Creativity as Variation — The Essence of Human Creativity Revealed by LLMs and the Design of Controllable Creativity —

Introduction

One of the most fascinating applications of large language models (LLMs) is:

Creative Generation

For example:

• Music generation
• Story generation
• Design ideation

Here, we notice a simple but important fact:

From the same prompt, multiple different outputs can be generated.

In other words:

From a single theme, diverse variations emerge.

This raises a fundamental question:

Is this the same structure as human creativity?

---

Creativity Is Not “Pure Novelty”

Creativity is often understood as:

“Producing something that has never existed before.”

However, when we look at actual creative processes:

■ Music

• Rhythm
• Harmony
• Scales

→ Clearly structured systems

■ Storytelling

• Narrative arcs (introduction, development, twist, conclusion)
• The hero’s journey
• Templated plot structures

→ Pattern-based composition

■ Design

• Layout
• Grids
• Usability constraints

→ Designed within constraints

In other words:

What changes is not the structure.
What changes is the variation within that structure.

---

What LLMs Reveal About Creativity

When LLMs generate music or text:

• The theme remains consistent
• The structure is preserved
• Yet the output differs each time

The key insight:

This is not randomness.

Its essence is:

Controlled variation within a structured space

In other words:

Creativity = Controlled Variation

---

Human Creativity Has the Same Structure

Human creativity follows the same pattern:

■ Reuse

• Using existing ideas
• Leveraging past experiences

■ Iteration

• Trial and error
• Fine-tuning

■ Recomposition

• Combining elements
• Shifting perspectives

Even great creators:

• Do not create in a single attempt
• Continuously revise
• Accumulate small differences

Conclusion:

Creativity is not invention from nothing.
It is transformation and variation within structure.

---

From Variation to System Design

This understanding has major implications for AI system design.

If:

Creativity = Variation

Then AI systems can be designed as follows:

■ Fix the structure

• DSL (Domain-Specific Language)
• Rules
• Constraints

■ Allow variation

• LLM-based generation
• Multiple candidate outputs

■ Evaluate

• Agent-based filtering
• Scoring

■ Select

• Decision-making
• Optimization

As a result:

Creativity becomes a designable process

---

Creative Architecture with Multi-Agent Systems

In the Decision Trace Model × Multi-Agent framework,
creativity is decomposed into layers:

■ Structure Layer

• Constraint definition via DSL
• Decision rules

■ Generation Layer

• Variation generation by LLMs
• Divergent exploration

■ Evaluation Layer

• Multi-agent evaluation
• Noise reduction
• Quality assessment

■ Selection Layer

• Choosing the optimal solution
• Determining direction

Only at this point:

Creativity becomes controllable

---

Variation in the Decision Trace Model

Structurally, it can be expressed as:

The critical perspective here:

LLM outputs are not answers.
They are candidates.

In other words:

Creativity is the process of generating and selecting candidates.

---

Why This Perspective Matters

Most AI applications aim to:

• Find a single correct answer
• Produce an optimal output

However, in creative domains:

There is no single correct answer.

What is needed instead:

• Multiple candidates
• Comparison
• Selection

Thus:

Creativity is a continuous process of exploration and selection

---

Key Insight

The most important point is:

Creativity does not emerge from nothing.

It emerges through:

Exploring variations within structure

---

Conclusion

LLMs do more than generate content.

They reveal something deeper:

• Creativity depends on structure
• Variation is its essence
• Selection gives it meaning

Final conclusion:

Creativity emerges from
Structure × Variation × Selection