// ⚖ Alignment & Preference Optimization
Reward Modeling (Preference Learning)
Learns a scoring function from pairwise human comparisons.
奖励模型 2017 Intermediate core
IN PLAIN WORDS
A Bradley–Terry loss pushes the preferred answer's score above the other. It is central to RLHF and vulnerable to reward hacking.
Used in: RLHF pipelines, best-of-N reranking
TIMELINE · prerequisites → this → what builds on it
01Original paper
Deep reinforcement learning from human preferences
Paul Christiano, Jan Leike, Tom B. Brown, et al. · 2017
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
Related:Reinforcement Learning from Human Feedback (RLHF)
Left: prerequisites · right: builds on this · dashed: related. Full map →
03Learning resources
Official docs, explainers, courses and reference code. All links checked.
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:
reward-model)