Particle.news
Download on the App Store

NeurIPS 2025 Ends With Record Attendance, Deep Split on Interpretability, and a Surge in AI for Biology

Divergent strategies on explainability underscored the field’s measurement gap.

Overview

  • Organizers reported about 26,000 attendees and more than 5,000 accepted papers from over 21,000 submissions, marking the conference’s largest year to date.
  • Google’s interpretability team said it is pivoting toward practical, impact-focused methods, while OpenAI’s lead signaled a push for deeper reverse‑engineering to fully understand neural networks.
  • Martian announced a $1 million prize to spur progress in a field its co‑founder described as still in its infancy.
  • Experts warned that current tools fail to assess higher‑level behaviors like reasoning, and biology benchmarks proved tricky as the Virtual Cell Challenge named two grand‑prize winners and a Generalist Prize for Altos Labs, which also presented scGeneScope and PerturBench.
  • AI‑for‑biology took center stage with company‑reported advances such as Genesis Therapeutics’ Pearl model said to exceed AlphaFold3 in atomic accuracy, Chai‑2’s 16% de novo antibody hit rate, and Insilico’s accelerated pipeline, with researchers calling for independent validation.