This thesis invites you to explore one of the most exciting open problems in modern AI: How can we make LLMs truly creative? While models like GPT-4 can write stories or brainstorm ideas, they mostly recombine patterns from their training data. True creativity, however, requires novelty, surprise, and value—properties current models often lack.

In this thesis, you will investigate how creativity can be defined, measured, and engineered in large language models. You will explore how LLMs generate ideas, where their limitations come from, and how we can design data-driven strategies to push them beyond the obvious. The work includes both conceptual analysis and hands-on ML engineering. The aim is to develop a working prototype or method that meaningfully increases creativity – with clear, realistic evaluation.

The thesis will be supervised at TU Dresden. With strong results, a publication is possible.

What are the tasks?

Understand creativity in AI & analyze current limitations

• Familiarize yourself with research on creativity (e.g., novelty, divergent thinking, recombination).
• Analyze how current LLMs fail or succeed when asked to produce creative content (e.g., low diversity, predictable patterns).
• Select appropriate creativity tasks (e.g., idea generation, concept blending, creative writing, scientific ideation).

Develop methods to enhance LLM creativity

• Implement data-driven techniques to increase creativity, such as:
  ◦ prompt-chaining, persona prompting, or multi-step generation
  ◦ sampling strategies or diversity-encouraging decoding
  ◦ “collective creativity” setups (multi-agent LLM brainstorming)
  ◦ adding external memory or knowledge for richer concept recombination
• Optionally experiment with architectural ideas (e.g., retrieval-augmented creativity, self-critique loops).
• Design and implement a small prototype system or pipeline that operationalizes creativity.

Design realistic evaluation for creativity

• Define concrete evaluation criteria (e.g., novelty, diversity, usefulness, human-rated creativity).
• Compare baseline LLM behavior and your enhanced model.
• Conduct small-scale human or expert assessments.

Document & analyze findings

• Analyze strengths, weaknesses, and failure modes.
• Discuss what your results mean for the future of creative AI.

What prerequisites do you need?

• Strong motivation for working on foundational capabilities of LLMs.
• Good programming skills in Python; experience with deep learning or transformer models.
• Curiosity about creativity, cognition, and idea generation.
• Very good English skills (for reading and writing).

Why this thesis is special

• Frontier problem: Creativity is one of the last major gaps between human and machine intelligence.
• Ambitious but feasible: Mix of theory and practice with clear, achievable outcomes.
• High relevance: Creative LLMs could transform design, science, writing, and innovation.
• Publication potential: Well-scoped experiments can yield publishable insights.