PhD student, University of Washington
2 papers at NeurIPS 2025
Language models are surprisingly robust to non-canonical tokenizations of the input, which can even lead to improved performance
We design a scheme to embed upto 24576 fingerprints into an LLM for better security
