Language models, such as those behind popular chatbots, do show signs of reasoning. But how reliable is that reasoning really? In his PhD research at Vrije Universiteit Amsterdam (VU), linguist Stefan Schouten investigated whether language models truly reason, and how robust that reasoning is. He defended his thesis Reasoning in Language Models: Identification and Application on 2 November 2026.
Schouten’s research shows that language models can perform impressively, but with clear limits. When GPT-4 was given sentences that could be interpreted in different ways, it chose the correct interpretation in more than 83 percent of cases. That is much better than guessing, but still far from perfect.
Inside the models, Schouten also found logical structures in the way concepts and sentences are connected to each other. At the same time, models turned out to be sensitive to distraction: irrelevant information can influence the outcome.
The question is becoming more urgent as millions of people use AI chatbots every day for advice, studying and work. New so-called “reasoning models” even show their thought process step by step. According to Schouten, that is exactly where caution is needed: text that looks like reasoning is not yet proof that real reasoning is taking place. On top of that, the explanation a model gives for an answer or decision does not necessarily reflect what actually happened inside the model.
The research methods Schouten developed help scientists and AI companies look inside language models and check how reliably they reason. An important open question remains: can language models also solve problems robustly that they have never encountered before? According to Schouten, much more research is needed to answer that.
Schouten conducted his research at the VU Faculty of Social Sciences and Humanities, under the supervision of Prof. Piek Vossen, Dr. Peter Bloem and Dr. Ilia Markov.
Read also
