Information Abundance Paradox: Long-Context Training Undermines Parametric Knowledge
In submission · Long-context training · Parametric knowledge
Arda Uzunoğlu
I'm a fourth year undergraduate studying computer science and applied mathematics at Johns Hopkins University, where I am fortunate to work with Daniel Khashabi, Benjamin Van Durme, and Eric Nalisnick. I am currently applying to PhD programs for Fall 2027 entry.
I am interested in developing a scientific understanding of large language models, as well as in broader questions about data, including how it governs model behavior, contributes to the emergence of intelligence, and shapes our understanding of models.
Recently
Received the Pistritto Research Fellowship ($6,000).
Trust Functions was accepted to ICML 2026.
Received a Summer PURA Research Grant ($6,000).
Feel free to contact me at auzunog1@jh.edu.
Selected Works
Trust Functions: Near-Lossless Weak-to-Strong Generalization by Learning When to Trust the Weak Teacher
ICML 2026 · Weak-to-strong generalization