In-depth Analysis of the Impact of RLVR Training on the Internal Representations of Large Language Models

Introduction / Main Floor: In-depth Analysis of the Impact of RLVR Training on the Internal Representations of Large Language Models

A groundbreaking open-source research project uses mechanistic interpretability techniques to systematically compare and analyze the differences in internal representations between base models, supervised fine-tuned models, and RLVR reinforcement learning models, providing a new perspective for understanding the formation mechanism of LLM reasoning capabilities.