SpreadsheetBench · reports/results_tables.md source

Version A. Sorted by mean test gain (champion − same-window seed)

# System Train (80) Val (40) Test (280): seed → champion (gain) Gain SD across seeds Test gain per seed (s0 / s1 / s2)
1 GEPA-full n/a 0.706 → 0.872 (+0.167) 0.638 → 0.767 (+0.129) 0.034 +0.145 / +0.090 / +0.152
2 SkillOpt n/a 0.697 → 0.886 (+0.189) 0.640 → 0.759 (+0.119) 0.018 +0.098 / +0.123 / +0.134
3 GEPA n/a 0.716 → 0.867 (+0.150) 0.614 → 0.732 (+0.118) 0.054 +0.179 / +0.100 / +0.075
4 WikiSkill 0.667 → 0.690 (+0.023), 2 of 3 seeds 0.692 → 0.831 (+0.139) 0.623 → 0.719 (+0.097) 0.016 +0.089 / +0.086 / +0.115
5 MIPROv2 n/a 0.719 → 0.817 (+0.097) 0.637 → 0.734 (+0.096) 0.036 +0.093 / +0.135 / +0.062
6 CapEvolve v2 0.727 → 0.813 (+0.086) 0.730 → 0.860 (+0.130) 0.631 → 0.726 (+0.095) 0.041 +0.075 / +0.068 / +0.143
7 CapEvolve v1 0.670 → 0.763 (+0.092) 0.686 → 0.814 (+0.128) 0.641 → 0.730 (+0.090) 0.045 +0.142 / +0.066 / +0.061
8 EvoSkill 0.688 → 0.688 (+0.000), 1 of 3 seeds 0.712 → 0.826 (+0.114) 0.643 → 0.732 (+0.087) 0.081 -0.006* / +0.143 / +0.123
9 Trace2Skill 0.600 → 0.600 (+0.000), 1 of 3 seeds 0.700 → 0.719 (+0.019) 0.642 → 0.699 (+0.054) 0.048 +0.000* / +0.069 / +0.093
10 Meta-Harness 0.683 → 0.717 (+0.033) 0.731 → 0.806 (+0.075) 0.640 → 0.692 (+0.052) 0.010 +0.041 / +0.060 / +0.055
11 ACE n/a n/a 0.616 → 0.656 (+0.040) 0.057 +0.015 / +0.105 / +0.000*

Version B. Sorted by mean absolute test score of the champion

# System Train (80) Val (40) Test (280): seed → champion (gain) Gain SD across seeds Test gain per seed (s0 / s1 / s2)
1 GEPA-full n/a 0.706 → 0.872 (+0.167) 0.638 → 0.767 (+0.129) 0.034 +0.145 / +0.090 / +0.152
2 SkillOpt n/a 0.697 → 0.886 (+0.189) 0.640 → 0.759 (+0.119) 0.018 +0.098 / +0.123 / +0.134
3 MIPROv2 n/a 0.719 → 0.817 (+0.097) 0.637 → 0.734 (+0.096) 0.036 +0.093 / +0.135 / +0.062
4 GEPA n/a 0.716 → 0.867 (+0.150) 0.614 → 0.732 (+0.118) 0.054 +0.179 / +0.100 / +0.075
5 EvoSkill 0.688 → 0.688 (+0.000), 1 of 3 seeds 0.712 → 0.826 (+0.114) 0.643 → 0.732 (+0.087) 0.081 -0.006* / +0.143 / +0.123
6 CapEvolve v1 0.670 → 0.763 (+0.092) 0.686 → 0.814 (+0.128) 0.641 → 0.730 (+0.090) 0.045 +0.142 / +0.066 / +0.061
7 CapEvolve v2 0.727 → 0.813 (+0.086) 0.730 → 0.860 (+0.130) 0.631 → 0.726 (+0.095) 0.041 +0.075 / +0.068 / +0.143
8 WikiSkill 0.667 → 0.690 (+0.023), 2 of 3 seeds 0.692 → 0.831 (+0.139) 0.623 → 0.719 (+0.097) 0.016 +0.089 / +0.086 / +0.115
9 Trace2Skill 0.600 → 0.600 (+0.000), 1 of 3 seeds 0.700 → 0.719 (+0.019) 0.642 → 0.699 (+0.054) 0.048 +0.000* / +0.069 / +0.093
10 Meta-Harness 0.683 → 0.717 (+0.033) 0.731 → 0.806 (+0.075) 0.640 → 0.692 (+0.052) 0.010 +0.041 / +0.060 / +0.055
11 ACE n/a n/a 0.616 → 0.656 (+0.040) 0.057 +0.015 / +0.105 / +0.000*