A model improved on AlpacaEval but its Arena ELO did no...

When a model shows improvement on AlpacaEval but not on Arena ELO, you should first investigate whether the gains are specific to prompt types that are rare in Arena’s live traffic. You should also inspect whether the model has become more verbose, as verbosity often boosts AlpacaEval scores but may not impress human voters in side-by-side comparisons. Finally, check if the model is new to the Arena and lacks a sufficient number of comparisons to produce a stable, reliable ELO estimate.