Merve Noyan Trains Qwen Model for Low Poly Worlds

Merve Noyan posted that she is training Qwen3.5-9B to generate isometric low poly worlds, using GLM 5.3 Flash as judge and testing GEPA with varying rewards. She reported early attempts produced no renders while later outputs lacked semantic quality, then moved to GRPO. In a follow-up post she noted issues with the judge's structured responses and overlit low-resolution images passed to the judge, which she plans to fix in the next run. She shared project links on Hugging Face based on TRL and OpenEnv.