Reproduce “Universe of Thoughts: A Computational Framework for Creative Reasoning in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuto Suzuki, Farnoush Banaei-Kashani
Universe of Thoughts (UoT) turns three theories of creativity—combining ideas across domains, exploring within an existing solution space, and changing the rules that define it—into distinct LLM reasoning procedures. Results across designed problem-solving tasks and scientific hypothesis generation suggest that the best strategy depends on task structure and model: transformational search led on two designed tasks, while exploratory search ranked first on HypoArena with Qwen3-14B.
Key insights
- Creative search is task-dependent rather than governed by one universally strongest strategy: changing assumptions can help when goals are clear, while cross-domain combination can suit broader objectives.
- On more constrained hypothesis-generation tasks, local exploration can be effective, but rankings also vary with the underlying language model.
- A promising analogy or rule change may be lost during final synthesis; generating an intermediate insight is not enough unless the system carries it through to the answer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.