Benchmark Leaderboards
Loading leaderboardsβ¦
Loading leaderboardsβ¦
Track which models hold the current SOTA on the benchmarks the field actually uses. Click a column header to sort; the leading row is highlighted.
Sample-efficient deep RL on the 26-game Atari 100K benchmark β agents train on only 100K environment steps (~2 hours of gameplay). Headline metric is the human-normalized median score across the 26 games (Human = 1.0).
Metric: Human-Normalized Median β higher is better Β· Source β
| # β² | Model / Paper | Human-Normalized Median | Ξ vs SOTA | Links |
|---|---|---|---|---|
| 1 | EfficientZero V2SOTA | 1.29 | β | arxiv.org β |
| 2 | EfficientZero | 1.09 | -0.20 | arxiv.org β |
| 3 | BBF (Bigger, Better, Faster) | 0.917 | -0.37 | arxiv.org β |
| 4 | DreamerV3 | 0.49 | -0.80 | arxiv.org β |
| 5 | SPR (Self-Predictive Representations) | 0.415 | -0.87 | arxiv.org β |
| 6 | DrQ | 0.268 | -1.02 | arxiv.org β |
| 7 | MuZero (Atari 100K) | 0.227 | -1.06 | arxiv.org β |
| 8 | CURL | 0.175 | -1.11 | arxiv.org β |
| 9 | SimPLe | 0.144 | -1.14 | arxiv.org β |