Analogical mappings of facts and counterfactuals in the human mind and Peirce’s abduction: limitations in LLMs

In this work, it is proposed that the human mind engages in an analogical mapping between facts found in “expert knowledge” and the abductive reasoning process described by Charles Sanders Peirce (1839–1914). This mapping connects the human mind with the causal world and enables the generation of hypotheses—whether scientific, artistic, or related to everyday life. Artificial Neural Networks (ANNs), including Large Language Models (LLMs) (Vaswani et al., 2017) and models incorporating Generative Adversarial Networks (GANs) (Goodfellow et al., 2014), face two key limitations: (1) They cannot work with counterfactuals, relying only on correlational datasets. (2) They are unable to perform true abductive reasoning. These systems may appear to “create” mappings with varying degrees of amplitude, but this impression arises from hyperparameters—such as Temperature (T) (Agarwal et al., 2024, Peeperkorn et al., 2024) and Top–K (Noarov et al., 2025)—configured by system supervisors or users via prompts. These parameters control the model’s output variability: Temperature influences the distribution of logits, while Top–K limits the prediction to the top K probable tokens, thus managing how deterministic or aleatoric the output becomes.