A few recent papers including Generating Literature-Driven Scientific Theories at Scale, https://arxiv.org/abs/2601.16282, and Heuristic-Based Ideation for Guiding LLMs Toward Structured Creativity
https://cichicago.substack.com/p/heuristic-based-ideation-for-guiding worked on scientific idea generation. However, LLMs are still widely believed not able to generate good ideas. It relates to the the fact that novelty is hard to define, either LLMs are criticized as generating incremental ideas or too crazy ideas. The hypothesis is that LLM can handle the abstract notion of novelty better when grounded. The ground can be both deduction and induction, writing out a proposal and imagining the best possible outcome, running a small experiment, and making comparisons in existing literature. Try these approaches and see if we can have better idea generation.
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{tan-scientific-idea-generation-2026,
author = {Tan, Chenhao},
title = {Scientific Idea Generation with Grounded Novelty},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/b79u7nCRP6kCV4Y3ORs8}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!