AutoRubric-T2I: Robust Rule-Based Reward Model for Text-to-Image Alignment
The paper introduces AutoRubric-T2I, a novel framework for improving text-to-image (T2I) alignment through automatic rubric generation. This approach synthesizes reasoning traces from preference pairs to create explicit scoring rubrics, enhancing the evaluation of generated images. The results indicate that AutoRubric-T2I significantly reduces the need for extensive training data while outperforming existing reward models in quality and interpretability.
- ▪AutoRubric-T2I synthesizes and selects explicit rubrics for guiding Vision-Language Model judges.
- ▪The framework uses less than 0.01% of annotated preference data to produce high-quality reward signals.
- ▪Extensive evaluations show that AutoRubric-T2I outperforms strong reward model baselines on image reward benchmarks.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.17602 (cs) [Submitted on 17 May 2026] Title:AutoRubric-T2I: Robust Rule-Based Reward Model for Text-to-Image Alignment Authors:Kuei-Chun Kao, Daixuan Huo, Yuanhao Ban, Cho-Jui Hsieh View a PDF of the paper titled AutoRubric-T2I: Robust Rule-Based Reward Model for Text-to-Image Alignment, by Kuei-Chun Kao and 3 other authors View PDF HTML (experimental) Abstract:Aligning Text-to-Image (T2I) generation models with human preferences increasingly relies on image reward models that score or rank generated images according to prompt alignment and perceptual quality.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
