Large Language Models (LLMs) are pretrained on massive datasets and later instruction-tuned via supervised fine-tuning (SFT) or reinforcement learning (RL). Best practices emphasize large, diverse pretraining data, whereas post-training operates differently: SFT relies on smaller, high-quality datasets, while RL benefits more from scale, with larger amounts of feedback often outweighing label quality. Yet it remains unclear why pretraining and RL require large datasets, why SFT excels on smaller ones, and what defines high-quality SFT data. In this work, we theoretically analyze transformers trained on an in-context weight prediction task for linear regression. Our analysis reveals several key findings: $(i)$ balanced pretraining data can induce latent capabilities later activated during post-training, and $(ii)$ SFT learns best from a small set of examples challenging for the pretrained model, while excessively large SFT datasets may dilute informative pretraining signals. In contrast, RL is most effective on large-scale data that is not overly difficult for the pretrained model. We validate these theoretical insights with experiments on large nonlinear transformer architectures.
Semantic Mismatch and Perceptual Degradation: A New Perspective on Image Editing Immunity
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences (CAS)(人工智能安全国家重点实验室,计算技术研究所,中国科学院);University of China Academy of Sciences(中国科学院大学);Center for Machine Vision and Signal Analysis, University of Oulu(信号分析中心,奥卢大学);School of Computer Science, China University of Geosciences(计算机科学学院,中国地质大学)
Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target Interaction Prediction
Beijing University of Posts and Telecommunications(北京邮电大学);China Telecom Cloud Computing Research Institute(中国电信云计算研究院);College of Computer Science and Electronic Engineering, Hunan University(湖南大学计算机科学与电子工程学院)
Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations
Harvard University(哈佛大学)
事后可解释性领域的一个关键问题是方法之间缺乏共同的基本目标。例如,一些方法的动机是函数逼近,一些是博弈论概念,还有一些是获得清晰的可视化。这种目标的碎片化不仅导致对解释的概念理解不一致,而且还导致不知道何时使用哪种方法的实际挑战。在这项工作中,我们开始通过统一八种流行的事后解释方法(LIME,C-LIME,KernelSHAP,Occlusion,Vanilla Quantients,Quantients x Input,SmoothGrad和Integrated Quantients)来解决这些挑战。我们发现,这些方法都执行本地功能近似的黑盒模型,不同的只是在附近和损失函数用于执行近似。这种统一使我们能够(1)陈述解释方法的没有免费的午餐定理,证明没有任何方法可以在所有邻域中表现最佳,以及(2)提供基于对黑盒模型的忠诚度的方法选择的指导原则。我们使用各种真实世界的数据集,模型类和预测任务来实证验证这些理论结果。通过将不同的解释方法纳入一个共同的框架,这项工作(1)推进了对这些方法的概念理解,揭示了它们共享的局部函数逼近目标、属性和彼此之间的关系,(2)指导这些方法在实践中的使用,提供了一个原则性的方法来选择方法并为创建新方法铺平道路。
Active Learning in Physics: From 101, to Progress, and Perspective
University of the Basque Country UPV/EHU(巴斯克大学UPV/EHU);Shanghai University(上海大学);Universitat de València(瓦伦西亚大学);Valencian Graduate School and Research Network of Artificial Intelligence (ValgrAI)(瓦伦西亚人工智能研究生学院与研究网络(ValgrAI));EHU Quantum Center(EHU量子中心)
Why Does RL Generalize Better Than SFT? A Data-Centric Perspective on VLM Post-Training
College of Computer Science, Sichuan University(四川大学计算机科学学院);Department of Computer Science and Technology, Tsinghua University(清华大学计算机科学与技术系);College of Computer Science and Technology, Zhejiang University(浙江大学计算机科学与技术学院)