Interesting work.

I'm surprised that hypothesis 2 (that CSV serialization format mangles table columns) was falsified. Back in the gpt-3.5-turbo and gpt-4o era, I did needle-haystack tests and found that table format mattered a lot (csv, tsv, markdown). Most models "could not read vertically" for csv (they were horrible), but they could for markdown. I concluded that serialization format or tokenization played a major role.

Nowadays, LLM performance on csvs is much improved (I'm guessing after being explicitly trained on CSV question-answering.) But I still carry the impression that LLMs read columns only by "memorizing" column positions in a format-dependent manner. Maybe this impression is out of date.