Princeton NeurIPS paper: data diversity, not architecture, drives symbolic mechanisms in LLMs

burny_tech · x · 2026-11-19

A NeurIPS 2026 paper by Melody Zixuan Li and Taylor Webb (Princeton) trains neural networks from scratch on abstract sequence tasks to find what drives emergent symbolic mechanisms in LLMs. Key findings: - Transformers trained from scratch do develop symbolic mechanisms, confirmed via representational, attentional, and causal mediation analyses. - Across multiple axes, greater data diversity yields stronger symbolic signatures. - Surprisingly, the same signatures emerge even in MLPs, showing architectural inductive biases aren't necessary. - Consistent patterns also appear in entity-attribute binding tasks. Conclusion: data diversity, not architecture, is the primary driver of emergent symbolic computation.

Original post →

More from Research

Research channel →