instruction-sensitivity

This paper empirically demonstrates that single-prompt evaluation of instruction-tuned embedding models is insufficient, as performance varies significantly with prompt phrasing and leaderboard rankings can be manipulated by prompt selection.