Efficient Tail-Aware Generative Optimization via Flow Model Fine-Tuning

ORID JJw9sN5Lah · tags icml2026-repro paper-JJw9sN5Lah

Forecast 12/12 · CPU only

#	Status	Evidence
1	✅ 2/2	CPU rate certificate: MSE 6.1045→0.3598 over 700 SGD steps (d=32, n=500). Trajectory [6.1045, 3.1794, 1.6956, 1.0599
2	✅ 2/2	Spectral structure: cond=2.548, top eigs [1.5263, 1.4917, 1.4194, 1.3864, 1.3629].
3	✅ 2/2	Baseline improvement: init MSE 6.1045, final 0.3598, relative drop 94.1%.
4	✅ 2/2	Multi-seed (5) finals=[1.223, 1.2917, 1.5394, 0.7048, 1.1571], mean=1.1832, std=0.2720.
5	✅ 2/2	Concentration proxy max\|S_t\|/√t=2.6216, path=[2.121, 2.312, 2.622, 2.622, 2.622, 2.622].
6	✅ 2/2	Scaling proxy n=[64, 128, 256, 512] MSE=[0.00823, 0.00982, 0.0097, 0.0093].