Ningxin Pan and colleagues test agents in games that keep moving while a model chooses its next action. A good decision can arrive after the collision. Their RealtimeBench uses 180 tasks across six game types, with separate tasks for learning, selecting updates and final testing.
- With games paused, GPT-6.1-sol beats the faster Jev in five of six game types. Under the live clock, their average progress scores are 1.51 and 13.51 out of 100.
- Their TORSI method moves slow reasoning between attempts. A teacher examines failures, practices alternatives in resettable games and writes instructions that the fast student follows alone during play.
- Validation selects which instructions survive. After two rounds, Jev’s average test score rises from 13.51 to 24.19; the second-round changes for Mario and Temple are rejected.
- Rewriting instructions from failures without practice scores 13.08. Combining teacher practice with live student checks scores 19.35 after one round, although practice budgets differ.
The experiment measures game progress, not the percentage of tasks completed. Its strongest engineering pattern is to keep expensive reflection away from time-sensitive execution, then test new guidance before replacing the old version.
Caveat: these are author-reported results with structured text observations and resettable games. Transfer tests reuse the same environments, and do not establish performance in unfamiliar worlds or from raw images.