Resolves YES if there are no more improvements in WinoGrande and BigBench-Hard beyond the state of the art as of market creation. Gemini Ultra 1.0 results will be considered included in this and do not count, even if leaderboards haven't been updated.
To the best of my knowledge that means 83.6% on BBH https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf
And 87.5% for WinoGrande https://arxiv.org/abs/2303.08774v4 though Gemini Ultra is not reported yet. Please let me know if I am missing something in comments.
Updated Gemini pro 1.5 is 89.2 on BBH: https://storage.googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf (page 5)
@CDBiddulph no, if there is improvement on either, it resolves NO. Is that unclear from the description?
@Thomas42 Oh lol, I knew that. Or rather, I meant "someone has to improve on both of those benchmarks for this to resolve NO." But sounds like an improvement in either would count as LLM progress not stalling, got it
@vluzko yea no improvement at all. I think that means 83.6% on BBH https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf
And 87.5% for WinoGrande https://arxiv.org/abs/2303.08774v4 though Gemini Ultra is not reported yet.
@ProjectVictory I'm sure whoever publishes it will call it a breakthrough ;) but yea poor choice of words, I'll replace with "improvement".