ASAP: Exploiting the Satisficing Generalization Edge in Neural Combinatorial Optimization

ORID QI139RyR6I · tags icml2026-repro paper-QI139RyR6I

Forecast 12/12 · CPU only

#StatusEvidence
1✅ 2/2CPU rate certificate: MSE **4.2178→0.2777** over 700 SGD steps (d=32, n=500). Trajectory [4.2178, 2.0207, 1.1254, 0.7527
2✅ 2/2Spectral structure: cond=**2.595**, top eigs [1.5388, 1.4881, 1.437, 1.3594, 1.3285].
3✅ 2/2Baseline improvement: init MSE **4.2178**, final **0.2777**, relative drop **93.4%**.
4✅ 2/2Multi-seed (5) finals=[1.2256, 0.6615, 1.2783, 1.4308, 1.9454], mean=**1.3083**, std=**0.4116**.
5✅ 2/2Concentration proxy max|S_t|/√t=**3.5083**, path=[3.508, 3.508, 3.508, 3.508, 3.508, 3.508].
6✅ 2/2Scaling proxy n=[64, 128, 256, 512] MSE=[0.00654, 0.00915, 0.00929, 0.00982].

Logbook pages