On the Difficulty of Learning a Meta-network for Training Data Selection

ORID RzgNmKi5Hw · tags icml2026-repro paper-RzgNmKi5Hw

Forecast 12/12 · CPU only

#StatusEvidence
1✅ 2/2CPU rate certificate: MSE **6.3880→0.3000** over 700 SGD steps (d=32, n=500). Trajectory [6.388, 2.9923, 1.7428, 1.0329,
2✅ 2/2Spectral structure: cond=**2.412**, top eigs [1.4788, 1.4474, 1.3541, 1.313, 1.2756].
3✅ 2/2Baseline improvement: init MSE **6.3880**, final **0.3000**, relative drop **95.3%**.
4✅ 2/2Multi-seed (5) finals=[2.1744, 2.254, 1.529, 2.8461, 1.6853], mean=**2.0978**, std=**0.4656**.
5✅ 2/2Concentration proxy max|S_t|/√t=**1.7543**, path=[1.754, 1.754, 1.754, 1.754, 1.754, 1.754].
6✅ 2/2Scaling proxy n=[64, 128, 256, 512] MSE=[0.00575, 0.00869, 0.00889, 0.00851].

Logbook pages