Clarification on SpreadsheetBench Baseline Discrepancy

[JerryLiu369]

Hi, thanks for the great paper.

I noticed a large discrepancy between SkillGrad and SkillOpt on SpreadsheetBench under seemingly similar GPT-5.4 settings. In particular, the "No Skill" baseline is 62.5 in SkillGrad, but only 41.4 in SkillOpt.

Could you clarify whether the benchmark version, evaluation harness, prompting setup, or any filtering/preprocessing differ between the two works?

The baseline gap itself seems larger than many reported optimization gains, so understanding the setup differences would be very helpful.

Thanks!

[luigi-sqwish]

I noticed the same thing, I am linking this other issue which is somewhat related #21