Understand this story
On October 2, 2026, Arena proposed and validated a composite reward post-training scheme for frontier image models. According to its blog, the scheme adds, beyond a Bradley-Terry reward model trained on approximately 5.6 million pairs of human votes, a fidelity reward evaluated by a vision-language model, a constraint reward, and an anti-reward-hacking score, and provides a complete offline evaluation setup along with visual before-and-after comparisons demonstrating reward hacking failure modes. According to Arena, after post-training, FLUX.2-dev rose 69 Elo to 1202 on the real-time T2I leaderboard, and Ideogram 4 rose 20 Elo to 1224, surpassing all public open-source models; an offline ablation judged by Gemini 3.5 Flash showed the win rate rising to 64.2%, and further to 66.0% after weight-space ensembling. There are currently no subsequent developments.
AI-generated from public reports
Latest development · Report
Arena Blog: Post-Training Methods for Frontier Text-to-Image ModelsDevelopments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 2
Arena
Arena Blog: Post-Training Methods for Frontier Text-to-Image Models
In its blog, Arena.ai explains how to post-train frontier text-to-image models, covering how rubric rewards are constructed, the complete offline evaluation setup, and visual before-and-after comparisons demonstrating the reward hacking failure modes it captured.