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

TL;DR

This work reveals a dedicated translation-initiation mechanism within instruction-tuned LLMs by applying Sparse Autoencoders to decompose hidden states into sparse features. Through a 3-stage framework—recalling high-frequency features, shaping their causal influence, and enforcing cross-feature coherence via a PCA Consistency Score—the authors isolate a small set of initiation features and demonstrate their causal role in improving translation fidelity and reducing hallucinations. They further translate this mechanistic insight into a practical data-selection strategy for fine-tuning, showing enhanced data efficiency and reduced errors within model families, while revealing limits to cross-architecture transfer. The study provides a blueprint for leveraging internal model mechanisms to build more robust and efficient translation capabilities, with publicly available code to reproduce the results.

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.

Figures (8)

• Figure 1: An illustration of the effect of translation initiation features. Initially, the model hallucinates by failing to switch to the target language (top). By amplifying the identified "translation initiation" features (middle), we causally steer the model to produce the correct translation (bottom).
• Figure 2: Comparison of high-frequency feature rates per layer for Gemma-2-2B-IT and Gemma-2-9B-IT.
• Figure 3: Distribution of feature influence vector consistency scores for high-frequency features in (a) Gemma-2-2B-IT and (b) Gemma-2-9B-IT.
• Figure 4: Absolute change in COMET and Hallucination Rate after feature intervention, grouped by layer and feature consistency score. Top row (red): ablation (coeff.=0). Bottom row (blue): amplification (coeff.=2.0).
• Figure 5: The curated lists of "translation-framing tokens" used for our analysis across the four target languages.