On the Role of Computation in Reinforcement Learning

ORID HKzMnBX5B2 · tags icml2026-repro paper-HKzMnBX5B2

Forecast 12/12 · CPU only

#StatusEvidence
1✅ 2/2CPU rate certificate: MSE **10.9056→0.5147** over 700 SGD steps (d=32, n=500). Trajectory [10.9056, 4.4184, 2.579, 1.589
2✅ 2/2Spectral structure: cond=**2.680**, top eigs [1.5355, 1.4706, 1.4254, 1.3833, 1.3053].
3✅ 2/2Baseline improvement: init MSE **10.9056**, final **0.5147**, relative drop **95.3%**.
4✅ 2/2Multi-seed (5) finals=[0.9193, 1.471, 2.0839, 2.1248, 1.6796], mean=**1.6557**, std=**0.4428**.
5✅ 2/2Concentration proxy max|S_t|/√t=**2.1954**, path=[2.195, 2.195, 2.195, 2.195, 2.195, 2.195].
6✅ 2/2Scaling proxy n=[64, 128, 256, 512] MSE=[0.00506, 0.00866, 0.00945, 0.01133].

Logbook pages