Generalization of large models: Methods, limitations and applications

Generalization is a cornerstone capability of Large Models (LMs), encompassing the transferability of pre-trained knowledge to novel tasks, unseen domains, or unprecedented combinatorial settings. It critically underpins LM performance in terms of robustness, reliability, and fairness, making it a persistent research hotspot. In recent years, LMs have attracted extensive attention due to their remarkable generalization prowess, zero-shot performance, and latent cross-domain value, progressively evolving into Industrial Large Models (ILMs). While these capabilities are primarily derived from large-scale unsupervised pre-training, significant limitations remain unaddressed. This paper presents a comprehensive survey of 150 recent studies on LM generalization, systematically categorizing them into 7 meta-tasks, and explores the role of efficient parameterization techniques, such as tensor decomposition. Building upon this foundation, we conduct 170 extensive empirical studies across 8 major industrial sectors and 4 industrial data formats, spanning the period from 2018 to 2026, to delineate the current landscape of ILMs. Conclusively, we propose the CS² (from Computer Science to Control Science) framework, which offers a holistic discussion on the effective integration of LM generalization methodologies into real-world industrial applications. The key contributions of this paper are the innovative establishment of a systematic classification paradigm for LM generalization methods, ensuring structured organization of the field; the proposal of a bridging framework that aligns generalization strategies from computer science with the stringent control requirements of industrial scenarios; and the synthesis of mainstream ILM industrial practices, inclusively covering non-Western LMs. We also discuss the limitations and provide directions for future research.