RSIGame self-improvement loops lift Godot game quality from 50.26 to 64.53
jiqizhixin · x · 2026-11-01
Researchers from UC San Diego, ByteDance and CMU propose RSIGame, making coding agents iterate usefully by verifying behavior instead of blindly rewriting code. On GameCraft-Bench Godot tasks, it improves games initialized with Codex and GPT-5.5 from an overall quality score of 50.26 to 64.53.
The method uses two cooperating loops:
- Local loop: explore the game, diagnose issues, edit code, verify fixes via interaction, with a shared checklist tracking evidence and regressions
- Global monitor: compares progress, preserves best checkpoints, and detects when a development direction stops paying off
They also train Qwen3.8-27B on verified development experience; combined with RSIGame it scores 61.38 on Godot, above the 50.26 one-shot GPT-5.5 baseline.
More from coding & agent
- Don't default to summarize: simple /bookmark /apply-to skills for agent article handling — gethackteam · 2026-11-01
- Dev Open-Sources Sentinel to Block Indirect Prompt Injection in MCP Tool Outputs — Infamous_You7977 · 2026-11-01
- Developer implements the RSA algorithm from scratch — cneuralnetwork · 2026-11-01
- rakyll Teases AX Studio: Test Agents on a Local Cluster, Ship to Production Unchanged — rakyll · 2026-11-01
- Hackathon study finds agent swarm language shifts measurably under deadline pressure — Miles_Brundage · 2026-11-01
- Karpathy shares tip: prompt LLMs with ASD-STE100, the aerospace controlled language spec — LeviTurk · 2026-11-01