Research

Legal-language models hallucinate, and in law a fabricated citation is worse than silence. The thesis went after faithful citation: getting a model to quote its sources instead of inventing them.

Together with my supervisor, started on copy mechanisms, in the spirit of "Copy Is All You Need," then moved to the decoding side. That shift produced CoCoLex: confidence-guided, copy-based decoding. When the model is unsure, it copies from the source instead of improvising. Accepted at ACL 2025, main conference. Further taking the CoCoLex codebase to the community as an open repository, so anyone can reproduce and build on confidence-guided, copy-based decoding for grounded legal text. Actively cooperating with researchers who want to extend it.