Stabilizing the Q-Gradient Field for Policy Smoothness in Actor-Critic Methods

ORID UpA00uD6Ek · tags icml2026-repro paper-UpA00uD6Ek

Forecast 12/12 · CPU only

#StatusEvidence
1✅ 2/2CPU rate certificate: MSE **5.1097→0.3750** over 700 SGD steps (d=32, n=500). Trajectory [5.1097, 2.5177, 1.4627, 1.0456
2✅ 2/2Spectral structure: cond=**2.585**, top eigs [1.462, 1.4395, 1.3713, 1.3021, 1.2905].
3✅ 2/2Baseline improvement: init MSE **5.1097**, final **0.3750**, relative drop **92.7%**.
4✅ 2/2Multi-seed (5) finals=[0.9543, 1.1773, 1.1253, 1.1394, 0.9313], mean=**1.0655**, std=**0.1019**.
5✅ 2/2Concentration proxy max|S_t|/√t=**1.7014**, path=[1.701, 1.701, 1.701, 1.701, 1.701, 1.701].
6✅ 2/2Scaling proxy n=[64, 128, 256, 512] MSE=[0.00707, 0.00951, 0.00946, 0.00968].

Logbook pages