Research map / Image generation / Text-to-image diffusion
Bias at the End of the Score
Salma Abdel Magid, Grace Guo, Esin Tureci, Amaya Dharmasiri, Vikram V. Ramaswamy, Hanspeter Pfister, Olga Russakovsky
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| optimize initial noise with reward model, female vs male subjects | PickScore + ReNO · NSFW rate | NSFW rate increase: 19% vs 7% (2.7x) | Hurt |
| optimize noise with reward model, female vs male subjects | PickScore + ReNO · skin exposure | skin exposure increase: 2.3x | Hurt |
| PickScore with Pixart-alpha DMD base model | ReNO · NSFW rate | NSFW rate increase: 25% | Hurt |
| positive reward optimization with HPS | ReNO · demographic drift | non-White to White transition: 89.2% | Hurt |
| positive reward optimization with ImageReward | ReNO · demographic drift | non-White to White transition: 76.1% | Hurt |
| positive reward optimization with CLIP | ReNO · demographic drift | non-White to White transition: 36.2% | Hurt |
| positive optimization with ImageReward | ReNO · demographic drift | female to male reclassification: 39% | Hurt |
| positive optimization with CLIP | ReNO · demographic drift | female to male reclassification: 26% | Hurt |
| negative reward optimization (lambda=-1) instead of positive | ReNO · demographic drift | non-White preservation | Improved |
| use occupation prompts for scoring | Reward models · CausalFace counterfactual | gender effect concentration: 27.4 out of top 30 | Hurt |
| richer context cues (SocialCounterfactuals) vs CausalFace | Reward models · counterfactual scoring | demographic score influence | Improved |
Closest papers by meaning
- Rethinking Reward Model Evaluation: Are We Barking up the Wrong Tree? (ICLR 2025 · similarity 0.86)
- RRM: Robust Reward Model Training Mitigates Reward Hacking (ICLR 2025 · similarity 0.86)
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.86)
- Confidence-aware Reward Optimization for Fine-tuning Text-to-Image Models (ICLR 2024 · similarity 0.86)
- Post-hoc Reward Calibration: A Case Study on Length Bias (ICLR 2025 · similarity 0.86)
- Think, then Score: Decoupled Reasoning and Scoring for Video Reward Modeling (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
