Safeyah Alshemali

KU & Columbia Alumina | Graduate student at Imperial College London | Photo by Neuron Visuals

[Safeyah Alshemali @SafAlshemali]

🎉 Happy to share that our paper on thematic fit and autoregressive LLMs has been accepted to CoNLL 2026!

We study whether LLMs learned which entities plausibly fill event roles (e.g., people eat pizza, but pizza does not eat people).

[Safeyah Alshemali @SafAlshemali]

📄 Paper (arXiv): https://t.co/qrpvmInhyK
📝 More details: https://t.co/6gFpQjIaia

[Safeyah Alshemali @SafAlshemali]

New Paper Out 📣📄

🔍We study whether autoregressive LLMs learned to generalize how well entities (event participants) fit to their assigned roles in events, a classic psycholinguistic problem known as thematic fit.

[Safeyah Alshemali @SafAlshemali]

🎉 Happy to share that our paper on thematic fit and autoregressive LLMs has been accepted to CoNLL 2026!

We study whether LLMs learned which entities plausibly fill event roles (e.g., people eat pizza, but pizza does not eat people).

[Safeyah Alshemali @SafAlshemali]

📌 Takeaway: thematic fit is still not "solved"

Beyond whether models possess event knowledge, understanding how different prompting and input formats elicit that knowledge remains an open question.

🌷I am grateful for the guidance of Daniel Bauer and Yuval Marton on this work

[Safeyah Alshemali @SafAlshemali]

📌Takeaway: Thematic fit isn’t "solved".

LLMs store rich event knowledge, but reliably eliciting it remains an open problem. Prompting choices matters, no single setup works everywhere, and no clear trend of improvement with newer models.

arXiv link: https://t.co/R5QjrMBDqv

[Safeyah Alshemali @SafAlshemali]

New Paper Out 📣📄

🔍We study whether autoregressive LLMs learned to generalize how well entities (event participants) fit to their assigned roles in events, a classic psycholinguistic problem known as thematic fit.

[Safeyah Alshemali @SafAlshemali]

🔸Not surprising: closed models (still?) yielded higher results than open models.
🔹Surprising: older models are better than newer ones! (4-turbo vs )