A novel approach for scientific idea generation based on large language models (LLMs) has been presented in a recent research paper.

GYWI: A Hybrid System for Scientific Inspiration

The system, named GYWI, integrates author knowledge graphs with retrieval-augmented generation (RAG) techniques to provide a richer academic context and clearer inspiration pathways. The goal is to overcome the limitations of traditional LLMs, which often produce results lacking a solid academic foundation.

GYWI uses an author-centered knowledge graph construction method and inspiration source sampling algorithms to create an external knowledge base. A hybrid retrieval mechanism, combining RAG and GraphRAG, is employed to retrieve content with both depth and breadth knowledge. Finally, a prompt optimization strategy, incorporating reinforcement learning principles, automatically guides LLMs in optimizing the results.

Evaluation and Results

The system was evaluated using a dataset created from arXiv articles (2018-2023). The generated ideas were evaluated based on novelty, feasibility, clarity, relevance, and significance. Experiments were conducted on different LLMs, including GPT-4o, DeepSeek-V3, Qwen3-8B, and Gemini 2.5. Experimental results show that GYWI significantly outperforms mainstream LLMs in several metrics, including novelty, reliability, and relevance.

For those evaluating on-premise deployments, there are trade-offs to consider. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these aspects.