Study probes whether causal neuron circuits transfer across prompts
A paper by Ömer Tekin Yavuz examines whether intervening on neuron groups changes language-model answers in a necessary or transferable way. Across Qwen3, Llama-3.2-1B, and Gemma-3-1B, the experiments find that a few neurons can often suffice, while group effects transfer to rephrasings of the same reasoning chain but weaken on other chains.
The results show that neuron-level causal effects can persist across paraphrases without generalizing broadly to other tested reasoning chains, offering a more qualified view of circuit uniqueness and transferability.
Computer Science > Machine Learning
arXiv:2611.13885 (cs)
[Submitted on 10 Nov 2026]
Title:Are Causal Neuron Circuits Unique and Transferable?
Authors:Ömer Tekin Yavuz (İletişim Yazılım)
View a PDF of the paper titled Are Causal Neuron Circuits Unique and Transferable?, by \"Omer Tekin Yavuz (\.Ileti\c{s}im Yaz{\i}l{\i}m)