abstract inductive reasoning requires dynamic learning in unfamiliar contexts. All the benchmarks which measure this specifically (in particular, task composition) see LLMs fail catastrophically. There is quite a lot of research published on these limitations now.
