← AI Terminology
Rejection Sampling (Alignment)
Rejection sampling in alignment generates multiple candidates and keeps only those that pass a reward threshold or rank filter, often to build higher-quality SFT/preference sets.
Also a classic statistics method; in LLMs it usually means filter-by-reward sampling.
Also a classic statistics method; in LLMs it usually means filter-by-reward sampling.
Why It Matters in AI
Before or beside RL, labs sample many outputs and keep winners — simple, parallel, effective. It improves datasets and can approximate heavier RL at inference (best-of-N).
Key Points
| Aspect | Description |
|---|---|
| Use | Curate SFT data; inference best-of-N |
| Cost | Linear in N samples |
| Loop | Sample N → score → keep top / above threshold |
| Risk | Reward hacking if scorer flawed |
| Related | Best-of-N, RLAIF, RAFT-style methods |
| Stats sense | Sample until proposal accepted under criterion |
Simple Analogy
Photographing twenty takes and publishing only the shots that pass the art director’s bar.
Common Usage Examples
- Generate 16 answers; keep highest RM score for SFT
- Best-of-N at serving time
- Filter code by unit tests
- Watch diversity collapse if filter too strict
Summary
In short: Rejection sampling keeps only high-scoring generations — a simple filter used to clean training data and boost inference quality.