Abstract

Recent advances in LLMs have made automated scientific research the nextfrontline in the path to artificial superintelligence. However, these systemsare bound either to tasks of narrow scope or the limited creative capabilitiesof LLMs. We propose Spacer, a scientific discovery system that developscreative and factually grounded concepts without external intervention. Spacerattempts to achieve this via 'deliberate decontextualization,' an approach thatdisassembles information into atomic units - keywords - and draws creativityfrom unexplored connections between them. Spacer consists of (i) Nuri, aninspiration engine that builds keyword sets, and (ii) the Manifesting Pipelinethat refines these sets into elaborate scientific statements. Nuri extractsnovel, high-potential keyword sets from a keyword graph built with 180,000academic publications in biological fields. The Manifesting Pipeline findslinks between keywords, analyzes their logical structure, validates theirplausibility, and ultimately drafts original scientific concepts. According toour experiments, the evaluation metric of Nuri accurately classifieshigh-impact publications with an AUROC score of 0.737. Our Manifesting Pipelinealso successfully reconstructs core concepts from the latest top-journalarticles solely from their keyword sets. An LLM-based scoring system estimatesthat this reconstruction was sound for over 85% of the cases. Finally, ourembedding space analysis shows that outputs from Spacer are significantly moresimilar to leading publications compared with those from SOTA LLMs.