Unveiling the Influence of Amplifying Language-Specific Neurons

TL;DR

This work investigates language‑specific neurons in multilingual LLMs and examines whether amplifying them can steer outputs toward a target language. It introduces the Language Steering Shift ($LSS$) metric and distinguishes language‑activated (Baseline) from language‑specific (LAPE) neurons, evaluating patched factors ($p_{max}$, $p_{median}$) and test‑time variants. Across 18 languages and three models, patched amplification achieves high language steering, with self‑language interventions improving or preserving performance on reasoning, knowledge, translation, and perplexity tasks, while cross‑language interventions generally degrade cross‑lingual transfer. The findings suggest that neuron‑level amplification can bolster language capabilities for low‑resource languages but offers limited cross‑lingual benefits, emphasizing both potential and boundary conditions for multilingual interventions. Overall, the study provides a framework for neuron‑level language control in LLMs and highlights the nuanced effects of amplifying language‑specific representations on downstream tasks.

Abstract

Language-specific neurons in LLMs that strongly correlate with individual languages have been shown to influence model behavior by deactivating them. However, their role in amplification remains underexplored. This work investigates the effect of amplifying language-specific neurons through interventions across 18 languages, including low-resource ones, using three models primarily trained in different languages. We compare amplification factors by their effectiveness in steering to the target language using a proposed Language Steering Shift (LSS) evaluation score, then evaluate it on downstream tasks: commonsense reasoning (XCOPA, XWinograd), knowledge (Include), and translation (FLORES). The optimal amplification factors effectively steer output toward nearly all tested languages. Intervention using this factor on downstream tasks improves self-language performance in some cases but generally degrades cross-language results. These findings highlight the effect of language-specific neurons in multilingual behavior, where amplification can be beneficial especially for low-resource languages, but provides limited advantage for cross-lingual transfer.

Figures (87)

• Figure 1: Illustration of our methodology. There are three sequential stages involving: 1). Identification of language-specific neurons, 2). Finding the optimal amplifying steering factor, and 3). Evaluation of the optimal amplifying steering factor on downstream tasks to understand its impact on the models' behavior.
• Figure 2: Comparison of averaged LSS scores of various steering factors across languages using LAPE and Baseline neurons.
• Figure 3: Delta Include-lite accuracy under LAPE neurons on Gemma2 9B. Blue highlights reductions, whereas red highlights increases.
• Figure 4: Delta XWinograd accuracy on Qwen2.5 7B (left) and Qwen2.5 0.5B (right) after amplification of LAPE neurons. Blue highlights reductions, whereas red highlights increases.
• Figure 5: Delta XCOPA accuracy after intevention of LAPE neurons for SeaLLMs3 7B (right). Row i depicts the initial language and column j is the intervention language.