I am less interested in questioning training data corruption than I am in questioning claims like this:

  test reasoning abilities such as pattern recognition, lateral thinking, abstraction, contextual reasoning (accounting for British cultural references), and multi-step inference.... its emphasis on clever reasoning rather than knowledge recall, Only Connect provides an ideal challenge for benchmarking LLMs' reasoning capabilities.

In general a lot of LLM benchmarks don't adequately consider that LLMs can solve certain things better than humans without using reasoning or knowledge. The most stupid example is how common multiple choice benchmarks are, despite us all learning as children that multiple-choice questions can be partially gamed with shallow statistical-linguistic tricks even if you have no clue how to answer the question honestly[1]; it stands to reason that a superhuman statistical-linguistic computer could accumulate superhuman statistical-linguistic tricks without ever properly learning the subject matter. AI folks have always been quick to say "if it quacks like a duck it reasons like a duck" but these days computers are quite good at playing duck recordings.

[1] "When in doubt, C your way out," sniffing out suspicious answers, shallow pattern-matching to answer reading comprehension, etc etc. One thing humans and LLMs actually do have in common is that multiple-choice tests are terrible ways to assess their knowledge or intelligence.