Cross-Lingual Generalization and Compression: From Language-Specific to Shared Neurons
This provides insights into cross-lingual generalization mechanisms in AI, which is incremental for improving multilingual NLP systems.
The study analyzed how multilingual language models evolve during pre-training, finding that they initially form language-specific representations that compress into cross-lingual abstractions over time, with neurons gradually aligning to encode the same semantic concepts across languages.
Multilingual language models (MLLMs) have demonstrated remarkable abilities to transfer knowledge across languages, despite being trained without explicit cross-lingual supervision. We analyze the parameter spaces of three MLLMs to study how their representations evolve during pre-training, observing patterns consistent with compression: models initially form language-specific representations, which gradually converge into cross-lingual abstractions as training progresses. Through probing experiments, we observe a clear transition from uniform language identification capabilities across layers to more specialized layer functions. For deeper analysis, we focus on neurons that encode distinct semantic concepts. By tracing their development during pre-training, we show how they gradually align across languages. Notably, we identify specific neurons that emerge as increasingly reliable predictors for the same concepts across languages.