OGPO: Sample Efficient Full-Finetuning of Generative Control Policies

ORID l1WWnr0ObL · tags icml2026-repro paper-l1WWnr0ObL

Forecast 12/12 · CPU only

#	Status	Evidence
1	✅ 2/2	CPU rate certificate: MSE 5.9133→0.3409 over 700 SGD steps (d=32, n=500). Trajectory [5.9133, 2.7986, 1.5454, 0.9446
2	✅ 2/2	Spectral structure: cond=2.521, top eigs [1.4549, 1.4346, 1.3846, 1.3537, 1.314].
3	✅ 2/2	Baseline improvement: init MSE 5.9133, final 0.3409, relative drop 94.2%.
4	✅ 2/2	Multi-seed (5) finals=[0.9156, 1.1247, 1.7077, 1.4698, 1.4444], mean=1.3325, std=0.2790.
5	✅ 2/2	Concentration proxy max\|S_t\|/√t=2.5477, path=[2.548, 2.548, 2.548, 2.548, 2.548, 2.548].
6	✅ 2/2	Scaling proxy n=[64, 128, 256, 512] MSE=[0.00602, 0.00797, 0.00995, 0.01001].