tyrone white

Research

Publications & talks

Mostly on the interpretability of language models from a linguistics-oriented perspective. Newest first.

  1. 2026 Accepted

    FreqBLiMP: Frequency-Controlled Minimal Pairs Reveal Robustness and Fragility of LLMs Under Lexical Rarity

    EMNLP 2026 (Main Conference)

    Introduces FreqBLiMP, a frequency-controlled extension of BLiMP that tests whether LLMs keep their grammatical preferences when minimal pairs use rare rather than common words. Models stay robust on morphosyntactic phenomena but degrade on phenomena that depend on specific lexical knowledge.

    arXiv
  2. Feb 2026

    Large Language Models Are Robust to Low-Frequency Words in Grammatical Evaluation

    言語処理学会 (NLP) 2026, poster

    A smaller, earlier version of the grammatical-generalization work, presented as a poster.

    Paper (PDF)

Research interests

mechanistic interpretabilitylarge language modelslinguistic representations in neural networksgrammatical generalizationsyntaxlexical frequencymultilingualitysparse autoencodersrepresentation analysisNLP evaluation