Post-Training Frontier Text-to-Image Models by Composing Preference and Rubric Rewards ▲ 23 💬 1 2026-10-10 arXiv HF 原文 摘要 （无摘要）