On February 5, 2026 we will host a guest lecture as part of the research seminars in language technology by our visiting researcher from the University of Amsterdam, Seth Aycock.
Place and Time:
Thursday, February 5, 2026 at 14:15 – 15:45
Live streaming: Zoom link
Title: Reasoning about Translation in LLMs
Abstract: What form of reasoning does translation between languages involve? Both humans and large language models (LLMs) perform this task competently; in this talk, we present recent work showing divergences between LLM and human strategies for reasoning about translation. First, human translators tend to decompose difficult translations with a research step before refinement; we show that such a chain-of-thought decomposition has no positive impact on the translation quality of LLMs. Next, humans tend to learn a second language (L2) effectively from worked examples, such as those in grammar books. We show empirically that LLMs fail to exploit grammatical rules for translation, relying instead on parallel examples, a form of discovery learning. In ongoing work, we continue this line of enquiry by asking whether metalinguistic awareness and reasoning, which can accelerate human L2 acquisition, are useful inductive biases for models learning to translate unseen or low-resource languages. We conclude with a discussion of whether we should expect model and human translation processes to align a priori.