Reasearch Intern

Investigated how LLMs represent source code through Sajaniemi's variable roles. Developed a static-analysis pipeline to extract 11 semantic variable roles from large-scale, multilingual codebases. Explored their internal representations through probing, sparse autoencoders (SAEs), controlled code edits, and generation interventions. Contributed to SharedSAE, a framework for learning a shared dictionary of interpretable sparse features across multiple model families.