Large Language Models (LLMs) frequently exhibit strong translation abilities, even without task-specific fine-tuning. However, the internal mechanisms governing this innate capability remain largely opaque. To demystify this process, we leverage Sparse Autoencoders (SAEs) and introduce a novel framework for identifying task-specific features. Our method first recalls features that are frequently co-activated on translation inputs and then filters them for functional coherence using a PCA-based consistency metric. This framework successfully isolates a small set of "translation initiation" features. Causal interventions demonstrate that amplifying these features steers the model towards correct translation, while ablating them induces hallucinations and off-task outputs, confirming they represent a core component of the model's innate translation competency. Moving from analysis to application, we leverage this mechanistic insight to propose a new data selection strategy for efficient fine-tuning. Specifically, we prioritize training on "mechanistically hard" samples—those that fail to naturally activate the translation initiation features. Experiments show this approach significantly improves data efficiency and suppresses hallucinations. Furthermore, we find these mechanisms are transferable to larger models of the same family. Our work not only decodes a core component of the translation mechanism in LLMs but also provides a blueprint for using internal model mechanism to create more robust and efficient models.

Ling Shi, Xinwei Wu, Xiaohu Zhao, Hao Wang, Heng Liu, Yangyang Liu, Linlong Xu, Longyue Wang, Deyi Xiong, Weihua Luo. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Hao Wang, Linlong Xu, Heng Liu, Yangyang Liu, Xiaohu Zhao, Bo Zeng, Liangying Shao, Yichen Dong, Xinwei Wu, Jiang Zhou, Tianyu Dong, Xiangxiang Zeng, Longyue Wang, Weihua Luo. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Jiang Zhou, Xiaohu Zhao, Xinwei Wu, Tianyu Dong, Hao Wang, Yangyang Liu, Heng Liu, Linlong Xu, Longyue Wang, Weihua Luo, Deyi Xiong. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

This paper investigates the practical prescribed-time (PT) synchronization and control for the multiweighted complex networks (MCNs) with or without external disturbances. First, due to the possible out of memory and practical error tolerance, we define a class of general time-varying regulatory functions for practical PT stability, which can contain many previous special forms. The classic PT regulatory function is replaced with a positive constant after the practical settling time according to the tolerant precision. Dynamical systems are proved to achieve practical PT stability at the practical settling time, which can be calculated under a given condition, then exponentially converge after this time. Both the settling time and the tolerance precision can be prescribed arbitrarily. Next, taking the external disturbances into account, both PT and practical PT stability are strictly obtained. Furthermore, practical PT synchronization of MCNs is also investigated, i.e., its synchronization error can enter to a prescribed-precision region after prescribed time surely and exponentially converge to zero. At last, numerical simulations are provided to evaluate the effectiveness of theoretical results.