Alien Space of Science: AI finds coherent research ideas no lab would propose, accepted at NeurIPS

anirudhg9119 · x · 2026-10-22

"The Alien Space of Science," an internship project at TipTree Systems with authors including Hugo Larochelle, Bernhard Schölkopf, Iyad Rahwan and Anirudh Goyal, is accepted to the NeurIPS 2026 main track.

The premise: scientific discovery is constrained by cognitive availability, not just truth — many directions are coherent given the literature but unlikely to be proposed because no existing community combines the needed concepts. LLMs inherit this bias, recombining high-density regions when asked for novel ideas.

Method and results:

Papers are decomposed into conceptual units clustered into a vocabulary of "idea atoms";
A coherence model scores whether atom combinations form viable research directions, while an availability model scores whether any existing author community could produce them; sampling ranks for high coherence and low availability;
On 16,068 peer-reviewed LLM papers from NeurIPS, ICLR, ICML and major NLP venues, the sampler yields more diverse research portfolios than frontier LLM baselines (including batch-prompted variants that see all outputs), while matching or beating per-idea quality;
Built entirely from pre-2025 data, availability weighting recovers 2025 combinations at 35.9× chance and strictly novel ones at 3.24×.

The authors argue AI ideation should complement rather than merely accelerate human science.

Original post →

Alien Space of Science: AI finds coherent research ideas no lab would propose, accepted at NeurIPS

More from Research