“Large language models have so far performed poorly on the ARC benchmark.”