// ⚖ Alignment & Preference Optimization
SimPO (Simple Preference Optimization)
Uses length-normalized log-likelihood as an implicit reward, reference-free.
SimPO 2024 Intermediate core
IN PLAIN WORDS
DPO needs a reference model and favors length; SimPO uses average log-probability plus a margin, cheaper and less verbose.
Used in: Open-model preference tuning
01Original paper
SimPO: Simple Preference Optimization with a Reference-Free Reward
Yu Meng, Mengzhou Xia, Danqi Chen · 2024
arXiv ↗ ✓ Verified · arXiv API
02Where it sits
Learn the left column first; the right column builds on this one.
Learn first
Then you can learn
- Nothing in the atlas builds on it yet
Related:ORPO (Odds Ratio Preference Optimization)
Left: prerequisites · right: builds on this · dashed: related. Full map →
03Learning resources
Official docs, explainers, courses and reference code. All links checked.
Official docs
04Where the course teaches it
On learning paths:Post-training & Reasoning Models
For AI / data
- Verification: verified, checked 2026-10-21.
- JSON: /api/learn/algorithms.json (id:
simpo)