Large language models trained with Reinforcement Learning (RL) with verifiable rewards exhibit strong reasoning ability and broad generalization, whereas models trained with Supervised Fine-Tuning (SFT) are often viewed as more prone to memorization and limited transfer. This paper rethinks this distinction through the lens of SFT training data on widely used Qwen2.5 and Qwen3 series models. First, we study the role of data source and show that it is critical: a carefully mixed SFT dataset substantially outperforms data generated solely by a larger model. Second, we study the role of data scale and show that matching the number of correct rollouts between SFT and RL greatly improves SFT generalization, while matching the total rollout budget enables SFT to generalize as well as RL. Combining these two factors further enables SFT to generalize even better than RL. Third, by using LLM annotations to characterize the solution methods in training rollouts, we show that larger datasets cover more tail methods and that these tail methods provide generalizable reasoning signals. Finally, we support these empirical findings theoretically by analyzing the training dynamics of shallow transformers under both RL and SFT.
Xingyao Li, Hao Zhang, Fengzhuo Zhang, Jiachun Pan, Yunlong Hou, Vincent Y. F. Tan, Zhuoran Yang
原文摘要
Despite the substantial progress in long video generation, multi-prompt synthesis often encounters inconsistencies resulting from the training-inference gap caused by length extension techniques and coarse prompt interpolation. To overcome these issues, we propose the Video Diffusion with hidden states Editing (ViDE) framework, which consists of two key components. The first is the Time-frequency based Temporal Attention Reweighting (TiTAR) algorithm, which leverages the relationship between inconsistencies and diagonal elements of temporal attention. By reweighting the attention scores via the Discrete Short-Time Fourier Transform (DSTFT), TiTAR effectively reduces frame inconsistencies, a capability further corroborated by a Fourier-based analysis. The second component, PromptBlend, reduces inconsistencies in multi-prompt settings through fine-grained prompt alignment and adaptive interpolation, enabling smooth semantic transitions. Extensive experiments demonstrate the effectiveness of ViDE, demonstrating consistent and significant improvements over multiple baselines.