3.5–7×
Broader effective atom vocabulary
Compared with frontier-LLM ideation baselines
The Alien Space of Science: Sampling Coherent but Cognitively Unavailable Research Directions
A way to search for scientifically coherent ideas that existing research communities are unlikely to propose.
The question
Why we’re working on this
When we ask a language model for research ideas, how often does it return to the same familiar combinations? We wanted to search beyond the directions that a field is already equipped to imagine.
Alien Science scores two things separately: whether an idea is scientifically coherent, and whether an existing research community is likely to propose it. We search for ideas that score well on the first and low on the second.
Approach
How it works
The method decomposes papers into reusable idea atoms, learns whether combinations of those atoms form coherent research directions, and separately estimates whether an existing author community is positioned to produce them. It then ranks combinations that are coherent but comparatively unavailable.
-
01
Atomize
Decompose papers into a shared vocabulary of reusable scientific ideas.
-
02
Score coherence
Estimate whether a combination could form a viable research direction.
-
03
Score availability
Estimate whether an existing author community is positioned to propose it.
-
04
Sample
Search for directions that remain coherent while being comparatively unavailable.
What to keep in mind
- Community availability is inferred from publication and authorship patterns, which are imperfect proxies for what researchers know or could imagine.
- The reported study focuses on the LLM literature; extending the method to other scientific fields requires new vocabularies and evaluation.
Evidence
Results at a glance
16,068
Peer-reviewed LLM papers
Across major ML and NLP venues
3
Blind evaluation modes
LLM, human, and downstream experiments
The study covers LLM research. A broader idea vocabulary does not establish that every proposed direction is novel or feasible; domain experts still need to assess the ideas.
Source: Alien Science v2, evaluationAuthors
People behind the work
Listed in the paper’s author order.
Alejandro H. Artiles
Tiptree intern
Tiptree Systems · Max Planck Institute for Human Development · Max Planck Institute for Intelligent Systems
First author of The Alien Space of Science. Now studies cognition–action tradeoffs in language agents through TextCraft.
Contributed to
- Lacuna
- The Alien Space of Science
- TextCraft
Martin Weiss
Tiptree staff
Tiptree Systems · Mila · Polytechnique Montréal
Builds tools that connect the literature, research questions, and the people working on them. Also studies peer-review verification and agent decision-making.
Contributed to
- Lacuna
- The Alien Space of Science
- AI Meta-Reviewing
- LLM Micro-Rationality
Levin Brinkmann
Research collaborator
Max Planck Institute for Human Development
Research collaborator at the Max Planck Institute for Human Development. Coauthor of The Alien Space of Science.
Contributed to
Iyad Rahwan
Research collaborator
Max Planck Institute for Human Development
Coauthor of Alien Science at the Max Planck Institute for Human Development, studying how AI can expand scientific exploration.
Contributed to
Bernhard Schölkopf
Scientific advisor
Max Planck Institute for Intelligent Systems · ELLIS Institute Tübingen
Scientific advisor and coauthor of Alien Science, exploring research directions beyond familiar combinations of ideas.
Contributed to
Christopher Pal
Scientific advisor
Polytechnique Montréal · Mila · Canada CIFAR AI Chair
Canada CIFAR AI Chair at Polytechnique Montréal. Scientific advisor and coauthor of Lacuna and The Alien Space of Science.
Contributed to
Hugo Larochelle
Scientific advisor
Scientific Director, Mila · Université de Montréal · McGill University
Scientific advisor and coauthor of Lacuna and Alien Science. Also collaborates on the behavioural economics of LLM agents.
Contributed to
- Lacuna
- The Alien Space of Science
- LLM Micro-Rationality
Anirudh Goyal
Research collaborator
Meta Superintelligence Labs · Formerly Google DeepMind
Research collaborator at Meta Superintelligence Labs and coauthor of The Alien Space of Science.
Contributed to
Nasim Rahaman
Tiptree staff
Tiptree Systems
Brings a background in theoretical physics and machine learning to scientific discovery. Coauthor of Lacuna and The Alien Space of Science.
Contributed to
Artifacts
Paper and project resources
- Artificial Intelligence
- Machine Learning
- AI for Science
- Paper PDF Full manuscript · v2
- arXiv record Abstract, authors, and version history
- DOI Persistent identifier for citing this paper
- Code Implementation of the method · MIT license
- Data and checkpoints The study’s datasets, model checkpoints, and evaluation artifacts
Paper: CC BY 4.0 · Code: MIT
Citation
Cite this work
Copy this entry into your .bib file or import it into your reference manager.
Show BibTeX
@misc{artiles2026alienspace,
title = {The Alien Space of Science: Sampling Coherent but Cognitively Unavailable Research Directions},
author = {Artiles, Alejandro H. and Weiss, Martin and Brinkmann, Levin and Rahwan, Iyad and Sch{\"o}lkopf, Bernhard and Pal, Christopher and Larochelle, Hugo and Goyal, Anirudh and Rahaman, Nasim},
year = {2026},
publisher = {arXiv},
doi = {10.48550/arXiv.2603.01092},
url = {https://arxiv.org/abs/2603.01092}
}Full paper
Read the full paper
Open the PDF or preview it here.
Work with us
Have a question worth working on?
We work with researchers on scientific discovery, agent evaluation, and tools for navigating the literature. Tell us what you’re exploring.