Newest papers

Uncovering Audio-Text Alignment in Large Audio-Language Models through Cross-Modal Concept Neurons

Tzu-Quan Lin, Chih-Kai Yang, Yi-Cheng Lin, Hao Tang, Hung-yi Lee · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28189 · cited by 0 works (OpenAlex) · open-access PDF

Portuguese

Enquanto a espinha dorsal do modelo de idioma é pré-treinado em texto sozinho, grandes modelos de áudio-linguagem (LALMs) podem, em grande medida, reconhecer atributos acústicos, como emoção, gênero do falante, e linguagem falada. No entanto, não está claro como a representação de áudio é integrada, se em tudo, com a outra representação de texto antes que o modelo produza texto. Hipotetizamos que há um subespaço compartilhado onde as representações de áudio e texto estão alinhadas. Para testar esta hipótese, identificamos neurônios cross-modal que são responsáveis por representar tanto áudio e texto. Encontramos alinhamento entre neurônios de áudio e texto a vários graus em três LALMs e quatro tipos de atributos ac

Machine translation (Meta M2M100); the English original follows.

Abstract

While the language model backbone is pretrained on text alone, large audio-language models (LALMs) can, to a great extent, recognize acoustic attributes, such as emotion, speaker gender, and spoken language. However, it is unclear how the audio representation is integrated, if at all, with the other text representation before the model produces text. We hypothesize that there is a shared subspace where the audio and text representations are aligned. To test this hypothesis, we identify cross-modal neurons that are responsible for representing both audio and text. We find alignment between audio and text neurons to various degrees across three LALMs and four acoustic attribute types. Causal interventions, steering and deactivating the neurons, confirm that the cross-modal neurons are not just necessary for prediction but can also be manipulated.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Representation (politics), Speech recognition, Subspace topology, Natural language processing, Artificial intelligence

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange