Alignment & Preference Learning · 2024
Deliberative Alignment: Reasoning Enables Safer Language Models
Teaches a model the text of its safety spec and trains it to reason over that spec at inference, the alignment method behind OpenAI's o-series that improves jailbreak robustness without raising overrefusal.
Editorial record
Plain-language summary
Teaches a model the text of its safety spec and trains it to reason over that spec at inference, the alignment method behind OpenAI's o-series that improves jailbreak robustness without raising overrefusal.
Knowledge graph
Relationships
Descendants
Conceptual ancestorEvidence: Strongly supported
Constitutional AI: Harmlessness from AI Feedback (RLAIF)
Builds on the alignment lineage in the archive
freshness sweep 2026
Source record
Provenance
- Record ID
- P-638
- Record created
- 2026-07-13
- Last reviewed
- 2026-07-14
- Record version
- 2
- https://arxiv.org/abs/2412.16339
- arXiv:2412.16339
Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.