Skip to content
TheChatArchiveThe Chat Archive
GeminiGemini (model unspecified)· GoogleHugging Face ShareLM (browser-extension contributions)

The last sentence (as part of a larger paper abstract) refers to comparing two models to…

  1. User

    The last sentence (as part of a larger paper abstract) refers to comparing two models to the best/worst model means most comparisons end up with the best (worst) model winning over both and not giving new information. With as little words as possible convey this idea instead of the last sentence here: a poor anchor can reduce the correlation with human rankings by up to x\textbf{x}. We identify that the most common anchor choices (best performing and worse performing model) make poor anchors. They produce skewed win-rate distributions that lack informativeness, effectively wasting the evaluation budget on non-discriminative samples.

  2. Gemini Assistant

    To streamline your last sentence while keeping the technical weight, you can replace it with:"Extreme anchors (best/worst models) produce skewed, uninformative win-rates that fail to discriminate between the remaining candidates."Why this works:"Extreme anchors" covers both best and worst in two words."Uninformative win-rates" captures the idea that the result is a foregone conclusion."Fail to discriminate" explains why the evaluation budget is wasted (it doesn't help you tell the other models apart).Would you like me to suggest a few more variations based on a specific word count limit?

  3. User

    Rewrite both sentences to be clearer with less jargon (informative win-rates, skewed, and discriminative)

  4. Gemini Assistant

    Here is a more direct way to phrase those two sentences:"Commonly used anchors, such as the best or worst performing models, are ineffective. Because these extreme benchmarks beat or lose to everything else, they fail to reveal the meaningful differences between the models being compared."Key ImprovementsReplaces "skewed/uninformative": Uses "beat or lose to everything else" to show why the data is useless.Replaces "non-discriminative": Uses "meaningful differences" to explain what you are actually looking for.Improved Flow: Connects the choice of anchor directly to the loss of useful data.Would you like me to try an even shorter version for a strict word count?