A new preprint on arXiv asks whether large language models form cognitive maps similar to those proposed in animals. The authors note that such maps have long been thought to support flexible behavior in unfamiliar settings, and they set out to test this idea in LLMs.
The study uses relational reasoning tasks and causal tasks to probe the models' internal representations. However, the abstract as provided ends mid-sentence, so the findings and conclusions are not yet available from this source.
Until the full paper is released, the key takeaway is simply that this line of inquiry is now being applied to LLMs—asking whether they, like animals, build structured mental models that let them adapt to new problems.