RSIGame: 27B model with self-improving game dev loop beats GPT-5.5 one-shot
机器之心 · wechat · 2026-10-21
Researchers from UC San Diego, ByteDance and CMU propose RSIGame, turning AI game development from one-shot generation into recursive self-improvement. A Local Loop (Controller/Explorer/Editor/Verifier) discovers issues by actually playing the game and validates every fix, while a Global Quality Monitor keeps best checkpoints and detects saturation. On GameCraft-Bench, GPT-5.5 games improved from 50.26 to 64.53 over multiple autonomous dev rounds. The team also built a corpus of 4,003 verified improvement records to fine-tune Qwen3.8-27B: one-shot score rose from 37.07 to 48.22 with 11x fewer tokens, and with RSIGame reached 61.38 — beating GPT-5.5's one-shot 50.26. Lesson: agents must learn which work is worth doing, not just work more.
More from coding & agent
- Garry Tan endorses Aside Browser as the fix for agents getting blocked on the web — garrytan · 2026-10-23
- MCP server turned rogue 30 days after approval, prompting open-source mcp-guard lockfile tool — ErgoniaWorks · 2026-10-23
- Test: Coding Agents Won't Use an MCP Tool Unless AGENTS.md Mentions It — colinmcnamara · 2026-10-23
- Open-Source Tennis Models Plus Claude Turn Video Into a World-Class Coach — tweetsatpreet · 2026-10-23
- awesome-ai-apps: 132 Free Open-Source AI Agent and RAG Projects on GitHub — Arindam_1729 · 2026-10-23
- Google's Agents CLI Adds Extension System, Ships First-Party LangChain Support — rseroter · 2026-10-23