JumpStart Your Policy Learning with Lessons from 160,000 Training Runs
The study finds offline policy-learning rankings are unstable once scale, tuning, and benchmark mix are taken seriously.
The authors trained more than 160,000 policies across 114 datasets and report that no single algorithm wins overall. Strong methods often land close in aggregate, while leaders change across environments. Hyperparameter tuning can reorder rankings, and different benchmark compositions can support conflicting conclusions. The release includes JumpStart, with trained policies, scores, hyperparameters, baselines, code, and a website for analysis. HF Daily Papers' note
The authors trained more than 160,000 policies across 114 datasets and report that no single algorithm wins overall. Strong methods often land close in aggregate, while leaders change across environments. Hyperparameter tuning can reorder rankings, and different benchmark compositions can support conflicting conclusions. The release includes JumpStart, with trained policies, scores, hyperparameters, baselines, code, and a website for analysis. HF Daily Papers' note
score 6