PKU & Samsung: slow model coaching lifts fast agent's live game score from 13.51 to 24.19
rohanpaul_ai · x · 2026-10-17
A Peking University and Samsung paper shows game benchmarks that pause the world during inference massively overstate slow reasoning models: on RealtimeBench (180 tasks, 6 continuously running games), GPT-6.1-sol beat the faster Jev in 5 of 6 paused games but scored 1.51 vs 13.51 live.
Their fix, TORSI (Two-Timescale Offline Recursive Self-Improvement), has the slow teacher study the fast student's failures offline and write reusable instructions the student follows in real time. Two rounds raised Jev's average score from 13.51 to 24.19 across all six game types, and the instructions also improve three other models without retraining.
Takeaway for real-time agents: keep a fast model in the loop and use a slow model offline to write and validate its instructions.
More from coding & agent
- 一行 PowerShell 实现 11 秒截图转表格，作者零手写代码 — dfinke · 2026-10-17
- HeyGen demos 3-step AI video pipeline: ChatGPT script, MCP avatar, auto-editing — HeyGen · 2026-10-17
- DeepSeek launches standalone Search & Fetch APIs at ~$0.70 per 1,000 successful requests, failures free — teortaxesTex · 2026-10-17
- OpenAI's new Python and Node SDKs make browser and computer use much easier — OpenAIDevs · 2026-10-17
- .NET 10 Lambda ARM64 base image ships with 170MB of leftover Yum cache — DanielLockyer · 2026-10-17
- A one-shot cron job prompt got Claude to crack ciphers and advance a 1932 number theory bound — IgorCarron · 2026-10-17