NeurIPS 2026 paper: neuron universality and selectivity scale systematically up to 30B

CSProfKGD · x · 2026-09-25

A NeurIPS 2026 paper, "Neuron Populations Exhibit Divergent Selectivity with Scale," asks: scaling laws describe how loss changes with scale — but do neurons inside models change predictably too?

Studying vision and language models up to 30B parameters (a collaboration involving Yasaman Bahri, Yunzhi Gandelsman, and Alyosha Efros), the authors find systematic scaling in neuron universality, specialization, and selectivity. Paper and code are public; topics span representation alignment/universality, scaling laws, and superposition.

Original post →

More from Research

Research channel →