The World Won’t Wait: Evaluating and Teaching Agents Under a Live Clock

Summary

Large language model agents can solve interactive tasks, but a correct action may arrive too late when the environment keeps changing. The paper introduces RealtimeBench, a benchmark of 180 tasks spanning six game types in which games continue running during inference. In paired evaluations, GPT-6.1-sol beats the faster Jev when games are paused, while Jev performs better when the games remain live. The authors propose Two-Timescale Offline Recursive Self-Improvement (TORSI), in which a slow teacher examines failures and tests corrections in paused games, then produces reusable instructions for a fast student. The student acts without teacher calls during gameplay, and validation scores determine which updates are retained. After two rounds, Jev’s average test score rises from 13.51 to 24.19 on a 0–100 progress scale, with gains across all six game types. The resulting instructions also improve the average scores of three other models without retuning. The authors present the approach as a way to coordinate deliberation, learning, and action across different timescales for agents operating in continuously changing environments.