Tribina Laboratorije za eksperimentalnu psihologiju
Understanding Space by Large Language Models
Dražen Domijan
Faculty of Humanities and Social Sciences, University of Rijeka, Rijeka, Croatia
Četvrtak, 26. mart 2026. u 15 časova
Laboratorija za eksperimentalnu psihologiju
(Suteren Filozofskog fakulteta u Beogradu)
______________________________
*Tribina će biti fotografisana
Understanding Space by Large Language Models
Dražen Domijan
1
Enio Ibrišagić
1
1 Faculty of Humanities and Social Sciences,
University of Rijeka, Rijeka, Croatia
drazen.domijan@uniri.hr
Large language models (LLMs) have recently attracted significant interest and
controversy due to their potential to perform various tasks at a level comparable
to humans, leading some authors to claim that they are achieving artificial
general intelligence (Sharkar, 2023). However, several recent studies have found
that LLMs tend to produce non-human errors in language comprehension tests
(Dentella et al., 2024; Murphy et al., 2025). In this study, we tested the ability of
LLMs to reason about space and spatial relations using a maze-solving task as a
benchmark. The task required finding a path from point X to point Y without
crossing the walls of the maze and drawing it on a supplied image. We tested
Grok, Gemini, ChatGPT-4o, and ChatGPT-5.2 on 20 problems (10 with a path from
X to Y and 10 without one). Each problem was presented five times in random
order, resulting in a total of 100 trials. We found that only ChatGPT-5.2 was able to
solve the task and draw the path, if it existed. The others made errors such as
crossing the walls, providing a verbal rather than a visual response, or inventing a
new maze (a form of visual hallucination). We also asked the LLMs to provide
metacognitive judgements about their performance. Interestingly, ChatGPT-5.2
typically gave lower confidence estimates (around 95%) than the others, which
were 100% confident they had solved the task. We further tested ChatGPT-5.2’s
spatial abilities by devising new tests involving contour tracing and inside-out
relations. Again, ChatGPT-5.2 was able to find a solution without making errors.
We interpret this finding as suggesting that ChatGPT-5.2 is not a pure LLM but
rather integrates an LLM with other techniques from classical AI, such as
breadth-first search and the A* algorithm, which have been developed to find a
path on abstract representations such as graphs.
