OpenAI's GPT-6.1-sol outperforms Jev on paused tasks, but lags in real-time

Paired tests show that the larger GPT‑6.1‑sol model beats the faster Jev model when the games are paused, but Jev overtakes GPT‑6.1‑sol when the games keep running. To address this gap, the authors propose Two‑Timescale Offline Recursive Self‑Improvement (TORSI). A slow “teacher” analyses failures in paused games, writes reusable instructions, and a fast “student” applies them without further teacher calls during live play. After two TORSI rounds, Jev’s average test score rises from 13.51 to 24.19 on a 0‑100 scale, improving across all six game types, and the same instructions lift three other models without additional tuning.

These results suggest a path toward real‑time intelligence, where deliberation and action are coordinated across different timescales to keep agents effective in continuously changing environments.