RSIGame: recursive self-improvement pushes Qwen-27B past GPT-5.5 one-shot at 11x fewer tokens
RSIGame · hf · 2026-10-01
RSIGame is an autonomous agentic game development framework with recursive self-improvement, addressing the overfitting that plagues naive iterative refinement.
A local explore-diagnose-improve loop explores the executable game and performs evidence-grounded fixes backed by an evolving checklist; a global loop tracks quality, keeps best checkpoints, and detects saturation or regression. Successful development experience is also internalized into the generator via training. Across 140 GameCraft-Bench tasks, two engines, and five generators, experience internalization lets Qwen3.8-27B reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.5 one-shot scores while cutting generation tokens 11x.
More from coding & agent
- codemode + general classification models demoed in pi draws developer praise — ricklamers · 2026-10-01
- Ex-Cursor engineer runs 6 Grok bots: from prompting to hiring a bot team — lasas · 2026-10-01
- Agent-built custom Lego sets: dev lets AI design and order real sets — noahsolomon · 2026-10-01
- A Month Delegating Real Paid Work to an AI Agent: Verification Beats Intelligence — alexksteadman · 2026-10-01
- HF researcher admits a well-maintained monorepo is the superior way to run an AI lab — soldni · 2026-10-01
- uv replaces five Python tools at 10-100x pip speed, written in Rust — blaizedsouza · 2026-10-01