Shanghai AI Lab Study: SFT Generalizes Under Three Conditions
上海AI Lab新研究:SFT能泛化,只要满足这三个条件
Shanghai AI Lab, Shanghai Jiao Tong University, and USTC tested Long-CoT SFT on Qwen3-14B-Base and found that cross-domain performance recovered and improved after 8 epochs, with generalization conditioned on optimization depth, data quality and structure, and base-model capability.
Why it matters: Shanghai AI Lab and other teams fine-tuned Qwen3-14B-Base on long chains of thought, observed cross-domain performance dropping first and then rising, and summarized three conditions under which the generalization holds.