The alien space of science: Coherent but cognitively unavailable directions

The work tackles the gap between what is scientifically plausible and what is cognitively available to research communities, arguing that many coherent research directions remain unexplored because no existing group combines the relevant concepts and methods. It operationalizes the "alien space" by decomposing papers into conceptual units, clustering those into a vocabulary of idea atoms, and training two scorers: a coherence model that assesses whether atom combinations form viable research directions and an availability model that estimates whether any current author community is positioned to produce them. Candidate combinations are ranked to maximize coherence while minimizing community availability.

Applied to 16,068 peer-reviewed LLM papers from NeurIPS, ICLR, ICML and major NLP venues, the sampler generates more diverse research portfolios than strong LLM baselines (including batch-prompted variants) while matching or exceeding per-idea quality under blind human judgments, LLM evaluation, and downstream experiments. Using only pre-2025 data for temporal validation, selecting under 1% of candidates recovers 2025 combinations at 35.9× chance, and emphasizing unavailability raises recovery of strictly novel combinations to 3.24× chance. By separating scientific plausibility from community availability, the framework produces AI-driven ideation that complements human science by expanding exploration into coherent directions the community is likely to overlook.