Humans apply many cognitive processes to coordinate smoothly in complex traffic scenarios. While many engineering tasks from perception to the automation of driving have been successfully studied, the core cognitive task, however, remains to be tackled: How can a formally grounded and cognitively inspired representation of the way humans mentally simulate possibilities in specific traffic situations be performed? Based on insights from cognitive science and formal knowledge representation and reasoning, we outline some constraints and formal foundations for such a framework. Limitations are discussed.

Large Language Models (LLMs) are increasingly evaluated in terms of their ability to demonstrate human-like behaviour on tasks inspired by the Turing Test. We examine this evaluation from the perspective of an Imitation Game comparing human and LLM performance on spatial reasoning tasks. We use the Region Connection Calculus (RCC-8), a domain in which complex spatial relations can be communicated with few premises yet still requires genuine spatial understanding. While several state-of-the-art LLMs achieve substantially higher logical accuracy than humans, accuracy alone does not fully characterize their suitability as assistant systems. By analysing response distributions, we identify systematic differences between the relations preferred by humans and those preferred by LLMs, even when LLMs are prompted to respond in a human-like manner. These differences show that cognitive adequacy is not met in these tests of spatial cognition in LLMs.

The sentence: If he had been the thief then he would have fled, is interpretable as a counterfactual conditional: he is neither the thief nor did he flee. Theorists often presuppose that counterfactuals are subjunctive conditionals of this sort. We present a new theory of counterfactuals based on mental models. Four experiments corroborated it, and its computational implementation simulates its dual systems of intuition and deliberation. As it assumes, no linguistic cue in English signals that a sentence is counterfactual. Depending on its context, any assertion can describe a factual or a counterfactual domain, whether it is indicative or subjunctive (Experiment 1). Assertions can have mixed moods with one indicative clause and one subjunctive clause, e.g.: She paid her mortgage or she would have moved. People reason that this disjunction implies that she did pay her mortgage and did not move (Experiments 2 and 3). Any sort of compound assertion based, say, on and, if, and or, can have a counterfactual interpretation. It yields inferences from models in both factual and counterfactual domains, and from a model in one domain to its denial in another domain. Their difficulty depends on the number of intuitive models of their premise (Experiment 4). We discuss how counterfactuals can elicit certain correct inferences more readily than factual assertions do, how verifications of counterfactuals can depend on mental simulations using kinematic models, and whether alternative theories of our results are feasible. We conclude that factuals and counterfactuals can have the same language, models, and inferences.

Over the past decades, human reasoning research has identified a variety of effects and processes, of which several have been compiled into comprehensive theories. Based on such theories, cognitive models were developed that made the theoretical findings applicable and testable. However, the models often consist of a variety on sub-processes and effects internally, but are not built in a modular way, hindering the transfer of findings between different models and their comparability. We approach this problem by proposing a different perspective: By treating the generation of cognitive process models as a search problem, process models can be derived from cognitive operations automatically in an objective way. Our method is illustrated on the domain of syllogistic reasoning, where we show that it generates a process model that outperforms state-of-the-art models while preserving their explanatory meaning. Finally, we discuss our approach as a framework for streamlining and facilitating cognitive modeling endeavors.

Syllogistic reasoning tasks typically contain a response option stating that no valid conclusion can be drawn from the given information. In fact, this response is the logically correct response for the majority of all syllogisms. However, the same response may also be chosen by participants who are simply unable to find the correct solution. In the current study, we suggest using an additional ‘don't know’ response option so that unsuccessfully aborted reasoning processes are not systematically classified as successful. We demonstrate that, when given the opportunity, participants make systematic use of a ‘don't know’ response option that. Further, the inclusion and use of the response option significantly impacted participants’ response distribution when compared to a group of participants that did not have such a response option. Our findings suggest that without an opt-out option, participants engage in biased guessing behavior when they ‘give-up’ on a problem, leading to incorrect ability estimates in testing situations. We conclude that future tests employing syllogistic reasoning tasks should include an additional ‘don’t know’ response option to obtain unbiased measures of reasoning abilities.