We further demonstrate an important moderating role of embedding extraction method and confirm results suggesting that final-layer embeddings underrepresent psycholinguistic information.

We show that accessibility cascades from lexical to semantic information similarly across all transformer language models, but is still realized differently in encoder and decoder models.

Submit your abstract by 4 February 2026 (AOE): machinebehavior.science/behavioral-c...

The workshop focuses on machines trained to replicate human behavior, as scientific models of human cognition and action, as tools for large-scale behavioral simulation, and as objects of study in their own right. We invite submissions for talks, lightning talks, and posters.

I am sorry to hear. If you are interested, there is research showing the usefulness of LLMs for psychological research, which in part derives from the models' ingestion of academic literature during training. I see no reason why LLM capabilities wouldn't generalize to this context to some extent.

I think I grasped your point. But all good, no problem if we disagree.

Logically correct---not quite what I said. I genuinely think that LLMs can provide useful signals on such questions. But we need to do our best to assess error and bias, like we should do with every other approach used to address such questions.

The ratings should be validated, of course. But there are good reasons LLMs might "know" this up to some reasonable error and bias.

With @tikhomirova.bsky.social, Valentin Kriegmair, Fritz Günther, Aliona Petrenco, Louis Schiekiera, @ercbravenewword.bsky.social, and @marcelbinz.bsky.social

Join one of Europe’s largest and best psychology departments in one of the world’s most liveable cities. We offer competitive conditions. Further information: jobs.univie.ac.at/job/Tenure-T... (deadline for applications: December 12, 2025)