The conclusions of the authors that LLMs can reason inductively very well runs counter to what I've read elsewhere. A big part of doing induction is the ability to generalize a shared pattern from multiple disparate examples, recognizing the essential elements that are necessary and sufficient to satisfy that pattern's operators' constraints. To date, I've seen consensus that LLMs can match verbs or basic relational operators across examples, thereby associating the mechanisms in similar events that lead to similar outcomes. But extending that facility further, to employing predicate logic operators, or even the simpler propositional ones appears to fall largely outside LLM capabilities. To suggest then that LLMs can then perform higher-order reasoning skills yet, like the modeling of contrapositives, this seems quite a stretch.