Arena unveils post-training method combining preference and rubric rewards

Arena detailed a new post-training method for text-to-image models that combines a reward model trained on 5 million human preference votes with VLM-generated rubric rewards, addressing reward hacking. Applied to FLUX.2-dev, it gained 69 Elo.

2026-10-02 ~ 2026-10-02 · 2 related posts

• Arena's reward recipe lifts post-trained FLUX.2-dev by 69 Elo on live T2I leaderboard — arena · 2026-10-02
• Arena details post-training T2I models with 5M preference votes and VLM-built rubric rewards — arena · 2026-10-02