Research map / Probabilistic and scientific modeling / Molecules and proteins
Curated Synthetic Data Doesn’t Have to Collapse: A Theoretical Study of Generative Retraining with Pluralistic Preferences
Ali Falahati, Mohammad Mohammadi Amiri, Kate Larson, Lukasz Golab
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 |
|---|---|---|---|
| increase number of balanced preferences from two to five (ablation) | pluralistic curation · CIFAR-10 retraining | FID: 77.6 -> 35.9 | Improved |
| increase number of balanced preferences from two to five (ablation) | pluralistic curation · CIFAR-10 retraining | class entropy: 1.098 -> 1.687 | Improved |
| increase number of balanced preferences from two to five (ablation) | pluralistic curation · CIFAR-10 retraining | KL divergence: 1.204 -> 0.616 | Improved |
| use two preferences with balanced q=0.5 instead of single preference q=0.0 (ablation) | pluralistic curation · CIFAR-10 retraining | FID: 100.0 -> 77.6 | Improved |
| use balanced q=0.5 instead of single preference q=0.0 (ablation) | pluralistic curation · CIFAR-10 retraining | class entropy: 0.032 -> 1.098 | Improved |
| skew q toward 0.1-0.3 (ablation) | two-preference curation · CIFAR-10 retraining | class entropy: 0.010 at q=0.1, 0.036 at q=0.3 | No clear effect |
| reduce distance between reward optima below about 2 (ablation) | pluralistic curation · synthetic 2D Gaussian mixture | reward variance | Hurt |
| use pluralistic instead of single-reward curation | GMM retraining · synthetic 2D Gaussian mixture | reward variance | Improved |
| increase distance between conflicting length targets (ablation) | GPT-2 pluralistic retraining · length preference text | length entropy | Improved |
Closest papers by meaning
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.92)
- Self-Consuming Generative Models with Co-Evolving Human Preferences (NeurIPS 2026 · similarity 0.90)
- A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops (ICLR 2025 · similarity 0.88)
- Stability and Diversity of Networked Self-Consuming Generative Ecosystems (NeurIPS 2026 · similarity 0.88)
- Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World (ICML 2025 · similarity 0.86)
- Towards Theoretical Understandings of Self-Consuming Generative Models (ICML 2024 · similarity 0.86)
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.
