Finding the Translation Switch: Discovering and Exploiting the Task-Initiation Features in LLMs

2026-01-16

AI Summary

While large language models (LLMs) demonstrate strong performance in translation tasks, the internal mechanisms that trigger such capabilities remain poorly understood. This work employs sparse autoencoders to identify and validate critical “activation features” within LLMs that initiate translation behavior. The functional role of these features is confirmed through causal interventions—specifically, amplification and ablation—and further supported by principal component analysis (PCA) consistency checks. Leveraging this mechanistic insight, the study proposes a data selection strategy targeting challenging translation samples, which substantially improves fine-tuning efficiency and mitigates hallucination. Moreover, the identified activation mechanism exhibits transferability, as evidenced by its successful validation on larger-scale models within the same family.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

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. The codes are available at https://github.com/flamewei123/AAAI26-translation-Initiation-Features.

Problem

• translation initiation
• Large Language Models
• mechanistic interpretability
• task-specific features
• innate translation competency

Innovation

• translation initiation features
• Sparse Autoencoders
• mechanistic interpretability
• data-efficient fine-tuning
• causal intervention