Language representations & the brain
Language Intelligence Lab, Georgia Tech · With Dr. Anna Ivanova
How do language models represent meaning, and how do those representations relate to the human language network? At Georgia Tech’s Language Intelligence Lab, I investigate the geometry of hidden activations and their alignment with neural responses.
The work combines linear probes, prompt recovery, and representational similarity analysis. These methods help test which representational subspaces predict neural responses, what information is retained in hidden activations, and how semantic alignment changes across model layers.
My contribution
- Probed hidden activations to identify subspaces that predict neural responses.
- Explored prompt recovery as a diagnostic for semantic compression.
- Compared LLM hidden states with voxel-wise fMRI responses.