Teaching LLMs to innovate by finding hidden connections between scientific discoveries
A framework implementing combinatorial creativity theory using LLMs for generating scientific ideas, featuring generalization-level retrieval and structured combinatorial processes for cross-domain knowledge discovery.
-----
https://arxiv.org/abs/2412.14141
🤔 Original Problem:
Recent LLM approaches for research idea generation lack grounding in established creativity theories, focusing only on novelty while neglecting value aspects of creative solutions.
-----
🔧 Solution in this Paper:
→ The framework uses a semi-structured "ideation format" to capture innovations at multiple abstraction levels (L1-L4)
→ A two-stage retrieval pipeline maps concepts across generalization levels using OpenAI's text-embedding model
→ The combinatorial process features parallel processing for component analysis, cross-domain application, and building block assessment
→ Integration phase reviews analyses from all levels to generate solutions with problem structure, design rationale, and key mechanisms
-----
💡 Key Insights:
→ LLMs can effectively implement combinatorial creativity when guided by theoretical frameworks
→ Multi-level retrieval enables meaningful connections between disparate domains
→ Structured combinatorial process generates ideas aligning with real research developments
-----
📊 Results:
→ Tested on OAG-Bench dataset with 87 papers
→ Outperformed baselines: Design Rationale (0.85 vs 0.78), Universal Principle (0.83 vs 0.75), Key Mechanism (0.87 vs 0.77)
→ Improved similarity scores by 7%-10% across metrics
------
Are you into AI and LLMs❓ Join me on X/Twitter with 51K+ others, to remain on the bleeding-edge of AI every day.
𝕏/🐦 https://x.com/rohanpaul_ai
Share this post