Nathan Wolf
I am a senior at UMass Amherst studying Computer Science and Computational Linguistics. My research focus is mechanistic interpretability, specifically examining the intersection of NLP and linguistic phenomena. I am currently conducting a senior thesis on morphological tokenization advised by Katrin Erk, as well as doing interperetability research in the detection of "translationese" with Shira Wein.
I also have experience in broader machine learning, as well as software engineering and applied LLMs. Outside of academia, my main interest is in the climate technology space.
Links
Email: nwolf3275@gmail.com
Ongoing Research
-
Crosslinguistic Performance Impacts of Morphological Tokenization for LLMs (working title) - 2025
Nathan Wolf, Katrin Erk
Senior Thesis
TL;DR: Examining the impacts of morphologically aligned tokenization on downstream LLM performance across languages with different word formation processes
Publications
-
Linking syntactic computation to eye movements in reading: Evidence from agreement - 2025
Adrian Staub, Nathan Wolf, Brian Dillon
Submitted to Open Mind: Discoveries in Cognitive Science
TL;DR: Examining the sources of the slowdown in text processing when disambiguating an ungrammatical sentence in which the ungrammaticality was not detected
-
Lost in Translation, and Found: Detecting and Interpreting Translation Effects - 2025
Shira Wein, Anna Serbina, Jiyuan Ji, Nathan Wolf, Jason DeGraaf, Prajakta Kini, Maria Leonor Pacheco
Submitted to ACL
TL;DR: Conducting interpretability research on a classifier that detects “translationese”, text translated into english from another language. Finding linguistic features the classifier uses to perform this task better than humans.