Review of NeurIPS 2025

I added the NeurIPS workshops this year (not the main conference).¹ Here are my takeaways. I'll follow up with predictions for 2026 in a later post (EDIT: here is that post).

(NOTE: I mostly attended bio/drug discovery/science workshops, so compared to last year my takeaways are much more focused)

• Nothing really surprised me this year. Perhaps this is because I focused on things related to drug discovery, and by virtue of being with Valence I've kept up with this subfield pretty well throughout the year.
• Overall breakdown of papers seemed to be ~50% LLM stuff, ~30% diffusion models for science, ~20% "other".
• LLMs keep getting better, no publishing from the frontier labs, no big conceptual shifts as far as I can tell.
• Lots of hype around "virtual cells", there was even a dedicated workshop for them. However, conceptually I don't think much changed: people are trying to fit models to gene expression data, sometimes other data too.
• Some trends in AI for science:
  • AI scientists, lots of exploring AI for literature search / reasoning / tool use
  • ML people recognizing that drug discovery is hard and can't be solved by making public benchmarks go up.
  • Slight preference towards fine-tuned open weight models instead of frontier models?
  • Lots of cool data development in structure-based drug design, eg CryoEM, more stuff using MD data
• Some talks I liked:
  • Eric Xing's talk about AIDO (AI digital organism, basically like virtual cell). Basically said "virtual cell won't be achieved by doing better on benchmarks every 3 months" and "we really need multi-modal data"
  • Yoshua Bengio's vision for a "Scientist AI" (basically a non-agentic system whose goal is to learn and say true things, you can query it to get actions but it doesn't actually have any goals itself).
• Best part of the conference was the people, good to meet people from other labs working on the same problems.
• I met a lot of BO researchers, and through conversations it finally became clear to me how I think the field of BO should re-orient itself so its work is more practical. I'll post a longer follow-up with my exact recommendations around the start of next year.
• PFNs (prior-fitted networks) are getting a lot of attention. I still have mixed feelings about them: their performance compared to exact Bayesian inference depends on generalization of the underlying transformer, especially in non-standard and misspecified settings. I'll see if people find successful use cases for them.