AI sampler surfaces research ideas that fit the literature but that no author community is positioned to pursue
- The method scores coherence and community availability separately, so it can rank research directions that are scientifically plausible but that no existing author group is positioned to propose.
- Selecting under 1% of the candidate space recovers 2025 idea combinations at 35.9 times chance, and weighting unavailability raises recovery of strictly novel combinations to 3.24 times chance.
- The pipeline splits 16,068 peer-reviewed LLM papers from NeurIPS, ICLR, ICML, and major NLP venues into 273 idea atoms, then trains a 42.8M-parameter coherence transformer and a dual-encoder availability model on publication histories of 39,791 researchers.
- In blind human, LLM, and downstream experimental evaluation, the sampler matches or exceeds the per-idea quality of frontier LLM baselines, including batch-prompted variants that see their full output, while producing more diverse research portfolios.
Hacker News opinions
Co-author here. This is my first paper from my postdoc, accepted to NeurIPS main. We split 16,068 LLM papers into 273 idea atoms, trained a 42.8M-parameter coherence model on papers as sequences of atoms, and trained an availability model on publication histories of 39,791 researchers to predict where each person could go next. Then we rank sampled combos by coherence and low availability.
Thanks, great idea and execution. Why does typing 'medicine' only give 6 proposals while 'physics' gives 447?
Congrats, this is a cool approach. I lead a research team and care a lot about tooling for human and agent collaboration. Do you have any evidence this scales to cross-domain research?
I think Velcro is the canonical example of this. There's a whole blog post about it:
I was impressed by the scale of the lacuna tool, but when I dug into steganography and covert communication, my PhD area, my expectations got tempered. Lots of papers add almost nothing. Still, a big cluster of papers that don't move the needle usually means the field needs a new approach, not more incremental work in the same local minimum.
This reminds me of Abbott's Flatland, where the inhabitants can't perceive the 3D space above them.
I'm reminded more of a Lem sub-story in A Perfect Vacuum about geniuses whose discoveries were too far from the mainstream for anyone to follow up on.
Random tangent: what if handwashing had spread thousands of years earlier? That one is an easy win that some cultures adopted and others refused for ages. Looking back from 2126, I'd bet historians will facepalm at something we're doing now that's easy to fix.
Quick fun read on this: Harry Turtledove's The Road Not Taken. I wonder if smartphones and LLM investment have pulled attention away from other technologies.
I filtered the generated biology proposals since I know that field a little. Mostly nonsense, and the confident language is what really gaslights me.
I think we can conceive of less than 1% of what's true about the universe, and understand 1% of that 1%.
If you can't conceive of it, you can't take a percentage of it.
I'm with the 1% idea. Take the Mona Lisa: an art historian, a doctor spotting a health issue, a geologist, a pigment chemist, and someone studying what the image does to the viewer all see different things. Good sci-fi tries to capture that, but writing it means actually building the understanding.
Humans could also use idea atoms like this to deliberately explore the alien space of a field.
Damn, this is the killer app for semantic spaces I've been waiting for since word2vec.