An Investigation of Neuron Activation as a Unified Lens to Explain Chain-of-Thought Eliciting Arithmetic Reasoning of LLMs

Large language models (LLMs) have shown 001 strong arithmetic reasoning capabilities when 002 prompted with Chain-of-Thought (CoT) 003 prompts. However, we have only a limited 004 understanding of how they are processed 005 by LLMs. To demystify it, prior work 006 has primarily focused on ablating different 007 components in the CoT prompt and empirically 008 observing their resulting LLM performance 009 change (Madaan and Yazdanbakhsh, 2022; 010