Research map / Large language models / AI and society
What happens when generative AI models train recursively on each others' outputs?
Hung Anh Vu, Galen Reeves, Emily Wenger
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 |
|---|---|---|---|
| mix generated data from the other model with real data (alpha=0.5 vs alpha=0) (ablation) | OPT-350m, K=2 · other model's task loss, beta=0.5 | eval loss: 3.1 -> 1.8 (alpha=0.5) vs 3.1 -> 3.5 (alpha=0) | Improved |
| mix generated data from the other model with real data (alpha=0.5 vs alpha=0) (ablation) | Llama 3.2 1B, K=2 · other model's task loss, beta=0.5 | eval loss: 3.7 -> 3.0 (alpha=0.5) vs 3.7 -> 4.2 (alpha=0) | Improved |
| use only generated data (alpha=1.0) instead of mixing in real data (ablation) | OPT-350m, K=2 · own task loss, Model 2 on Task 2, beta=0.5 | eval loss: 1.7 -> 2.2 | Hurt |
| use only generated data (alpha=1.0) instead of alpha=0.5 (ablation) | Llama 3.2 1B, K=2 · own task loss, Model 2 on Task 2, beta=0.5 | eval loss: 1.3 -> 1.9 | Hurt |
| add synthetic data at all (alpha=0.5 vs alpha=0) on own task (ablation) | OPT-350m and Llama 3.2 1B, K=2 · own task loss, beta=0.5 | eval loss: 3.3 -> 3.3 (OPT Model 1) | No clear effect |
| add public real data D* (beta=0.5 vs beta=0) at alpha=0 (ablation) | Llama 3.2 1B, K=2 · own task loss, Model 1 on Task 1 | eval loss: 2.8 -> 4.5 (beta=0) vs 2.8 -> 2.9 (beta=0.5) | Improved |
| introduce synthetic data exchange (alpha>0) (ablation) | Llama 3.2 1B, K=2 · embedding cosine similarity of outputs on D1 | cosine similarity: 0.73 -> 0.88 (alpha=0.5) vs 0.73 -> 0.70 (alpha=0) | Mixed |
| set alpha=0.5, beta=0.5 versus other settings (theory) (ablation) | K=4 linear regression system · asymptotic relative efficiency | relative efficiency | Improved |
Closest papers by meaning
- Stability and Diversity of Networked Self-Consuming Generative Ecosystems (NeurIPS 2026 · similarity 0.89)
- Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy and Research (NeurIPS 2025 · similarity 0.87)
- Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World (ICML 2025 · similarity 0.86)
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.86)
- Self-Consuming Generative Models with Co-Evolving Human Preferences (NeurIPS 2026 · similarity 0.86)
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · 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.
