| \caption{\textbf{Cross-task gains depend on the task and evaluation delay.} | |
| Mixed-latency training on Demon Attack yields higher delayed returns on | |
| Atlantis and AirRaid, while Asterix has lower return at four frames than with | |
| zero-latency training. Gray and purple denote zero-latency and mixed-latency | |
| training, respectively. From left to right, panels show Demon Attack, | |
| Asterix, Atlantis, and AirRaid, evaluated at delays of 0, 2, and 4 raw frames. | |
| Each point is the mean episode return over 20 evaluation episodes.} | |