Papers by Andrea Gregor de Varda
Different types of syntactic agreement recruit the same units within large language models (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models can reliably distinguish grammatical from ungrammatically sentences, but how gramatical knowledge is represented within the models remains an open question. |
| Approach: | They use a functional localization approach inspired by cognitive neuroscience to identify the LLM units most responsive to 67 English syntactic phenomena in seven open-weight models. |
| Outcome: | The proposed model is most responsive to 67 English syntactic phenomena and consistently supports model performance. |
The Emergence of Semantic Units in Massively Multilingual Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Massively multilingual models can process text in several languages relying on a shared set of parameters, but little is known about the encoding of multilingual information in single network units. |
| Approach: | They propose to use a shared set of parameters to encode multilingual information in single network units. |
| Outcome: | The proposed model achieves higher scores in semantic encoding in languages with more cross-lingual alignment than those with more shared cross-linguistic substrate. |