Indic-TunedLens：印度语言中的多语言模型解释

Multilingual large language models (LLMs) are increasingly utilized in linguistically diverse regions such as India, yet most interpretability tools remain focused on English. Prior studies indicate that LLMs often function within English-centric representation spaces, highlighting the urgent need for cross-lingual interpretability. Indic-TunedLens is introduced as a novel interpretability framework specifically designed for Indian languages, learning shared affine transformations. Unlike the standard Logit Lens, which directly decodes intermediate activations, Indic-TunedLens adjusts hidden states for each target language, aligning them with target output distributions to facilitate more faithful decoding of model representations. The framework has been evaluated across 10 Indian languages, demonstrating its effectiveness.