Zhijie Deng

In this paper, we propose a fine-grained influence function framework to trace how training data on SFT phase shapes LLM reasoning in math and code tasks.