PKU and Samsung paper: slow model coaches fast one offline, lifting live game scores from 13.51 to 24.19
rohanpaul_ai · x · 2026-10-17
A Peking University and Samsung paper introduces RealtimeBench — 180 tasks across six game types where the world keeps running during inference, so a correct action arriving too late is worthless.
Key findings:
- When games are paused, the stronger reasoning model (GPT-6.1-sol) beats the faster Jev; when they keep running, it falls behind.
- The proposed TORSI method (Two-Timescale Offline Recursive Self-Improvement) has a slow teacher study failures in paused games between attempts and write reusable instructions for a fast student, who acts without teacher calls during gameplay; validation scores gate updates.
- Two rounds raise Jev's average score from 13.51 to 24.19 (0–100 scale), improving all six game types, and the instructions also lift three other models without retuning.
The work points toward "real-time intelligence": organizing deliberation, learning, and action across timescales.
More from coding & agent
- Dev proposes shared cloud-local filesystem to unify agent bot/code/work workflows — sytelus · 2026-10-17
- AI didn't replace developers — it replaced our understanding of our own code — loveolya · 2026-10-17
- Reddit user pairs Mac mini M5 Pro with old M4 as sub-agent for local coding agents — Calm-Set6265 · 2026-10-17
- Go team adds interactive refactoring to gopls, proposes new LSP standard for multi-step dialogs — rseroter · 2026-10-17
- research-rag: local CPU-only PDF/ebook search with browser, CLI and MCP access — kishkisan · 2026-10-17
- Single-video 3D reconstruction via agent→MCP→Blender simulates industrial hazards — RexDouglass · 2026-10-17