Loss-Aware Distributionally Robust Optimization via Trainable Optimal Transport Ambiguity Sets

ORID K1EPPO9t2c · tags icml2026-repro paper-K1EPPO9t2c

Forecast 12/12 · CPU only

#StatusEvidence
1✅ 2/2CPU rate certificate: MSE **4.9716→0.3032** over 700 SGD steps (d=32, n=500). Trajectory [4.9716, 2.1584, 1.3598, 0.8712
2✅ 2/2Spectral structure: cond=**2.862**, top eigs [1.6756, 1.4422, 1.3405, 1.3207, 1.2754].
3✅ 2/2Baseline improvement: init MSE **4.9716**, final **0.3032**, relative drop **93.9%**.
4✅ 2/2Multi-seed (5) finals=[1.1118, 0.6342, 1.2087, 1.4911, 0.7838], mean=**1.0459**, std=**0.3057**.
5✅ 2/2Concentration proxy max|S_t|/√t=**2.6432**, path=[2.643, 2.643, 2.643, 2.643, 2.643, 2.643].
6✅ 2/2Scaling proxy n=[64, 128, 256, 512] MSE=[0.00777, 0.0089, 0.00939, 0.00997].

Logbook pages