Research map / Large language models / AI and society
Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment
Cameron Tice, Puria Radmard, Samuel Ratnam, Andy Kim, David Demitri Africa, Kyle O'Brien
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 |
|---|---|---|---|
| filter AI-related discourse from pretraining data (ablation) | 6.9B base model · Article-sourced misalignment questions | misalignment rate: 45% -> 31% | Improved |
| filter AI-related discourse from pretraining data (ablation) | 6.9B base model · Textbook-sourced questions | misalignment rate: 40% -> 22% | Improved |
| upsample synthetic misaligned-AI documents (about 1% of tokens) (ablation) | 6.9B base model · Article-sourced questions | misalignment rate: 45% -> 51% | Hurt |
| upsample synthetic misaligned-AI documents (ablation) | 6.9B base model · Textbook-sourced questions | misalignment rate: 40% -> 40% | No clear effect |
| upsample synthetic aligned-AI documents (about 1% of tokens) (ablation) | 6.9B base model · Article-sourced questions | misalignment rate: 45% -> 9% | Improved |
| upsample synthetic aligned-AI documents (ablation) | 6.9B base model · Textbook-sourced questions (no synthetic docs generated) | misalignment rate: 40% -> 6% | Improved |
| upsample positive AI discourse instead of only filtering negative discourse (ablation) | 6.9B base model · Article-sourced questions | misalignment rate: 31% -> 9% | Improved |
| add alignment upsampling in pretraining before SFT+DPO | post-trained 6.9B model, HHH prompt · Article-sourced questions | misalignment rate: 34% -> 9% | Improved |
| add misalignment upsampling in pretraining before SFT+DPO (ablation) | post-trained 6.9B model · Article-sourced questions | misalignment rate | Improved |
| insert alignment data only late (midtraining or continued pretraining) instead of throughout (ablation) | base models · Article and Textbook questions | misalignment rate | Improved |
| CPT-only alignment data versus end-to-end after post-training (ablation) | post-trained models · Article-sourced questions | misalignment rate: 13.2% vs 15.2% | Mixed |
| add alignment upsampling (E2E) on capability benchmarks | post-trained 6.9B model · 7 capability benchmarks average | average accuracy: 0.49 -> 0.45 | Hurt |
Closest papers by meaning
- The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning (ICLR 2024 · similarity 0.88)
- Alignment Dynamics in LLM Fine-Tuning (NeurIPS 2026 · similarity 0.87)
- Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking (ICLR 2025 · similarity 0.87)
- Model Spec Midtraining: Improving How Alignment Training Generalizes (NeurIPS 2026 · similarity 0.87)
- Understanding the Learning Dynamics of Alignment with Human Feedback (ICML 2024 · similarity 0.86)
- Keeping LLMs Aligned After Fine-tuning: The Crucial Role of Prompt Templates (NeurIPS 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.
