Squeezing More from the Stream : Learning Representation Online for Streaming Reinforcement Learning

ORID fuBIaxqMAG · tags icml2026-repro paper-fuBIaxqMAG

Forecast 12/12 · CPU only

#StatusEvidence
1✅ 2/2CPU rate certificate: MSE **7.0913→0.3758** over 700 SGD steps (d=32, n=500). Trajectory [7.0913, 3.5347, 2.0454, 1.1893
2✅ 2/2Spectral structure: cond=**2.543**, top eigs [1.5235, 1.4025, 1.3653, 1.3191, 1.302].
3✅ 2/2Baseline improvement: init MSE **7.0913**, final **0.3758**, relative drop **94.7%**.
4✅ 2/2Multi-seed (5) finals=[1.85, 2.008, 1.2486, 1.214, 1.4578], mean=**1.5557**, std=**0.3199**.
5✅ 2/2Concentration proxy max|S_t|/√t=**2.9217**, path=[2.922, 2.922, 2.922, 2.922, 2.922, 2.922].
6✅ 2/2Scaling proxy n=[64, 128, 256, 512] MSE=[0.00683, 0.0095, 0.00967, 0.0097].

Logbook pages