← Back to concept explainers

Concept Fable

rejection sampling

Conclusion first:Rejection Sampling in AI training often refers to generating multiple candidate outputs and keeping only the qualified samples according to quality standards, used to construct higher-quality fine-tuning or preference data. It can improve data quality, but depends on the reliability of the filter, reward model, or human annotation. Enterprises can use it to accumulate high-quality customer service replies, tool-call traces, RAG answers, and task execution examples.

Rejection Sampling · Fine-tuning

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

Fire a hundred cups first, then keep twelve

The porcelain kiln fired cups, and after coming out of the kiln they were always uneven. The master could not send every one as tribute, so he picked them one by one under the light: cracked, crooked-bottomed, and muddy-glazed ones were set aside, and only the qualified ones were stamped.

The apprentice asked: if they will be discarded anyway, why not fire only the good cup from the start? The master said the heat fluctuates and the clay has its own temper; firing several and then choosing is often more reliable than betting on one perfect piece.

The first round of selection was too coarse—many cups looked round but only seeped water when filled with hot tea. In the second round, the master added a water test, a tap test, and a level check. More were rejected, and those that remained were more reliable.

But the kiln owners also looked at the ledger: firing too much wastes firewood, and selecting too strictly leaves no cups to deliver. So they set different thresholds for tribute cups, daily-use cups, and practice cups.

The kiln owner once tried firing only one most carefully made cup. The master watched over it all night, but when it came out of the kiln it still cracked from a gust of cold wind. He then tried lowering the standard; the cups could all be sold, but half a cart was returned when they were sent as tribute.

Later, an extra batch was fired in the kiln according to use, and after coming out of the kiln they were selected in three rounds: under the lamp, on the table, and in water. Cracked ones were discarded, slightly crooked ones were sold to tea stalls, and those with proper glaze color were sent to the government office. Discarding bad cups may seem wasteful, but it made the cups finally delivered more reliable. Fire and clay involve uncertainty; generating multiple candidates first and then rejecting the unqualified ones by standard is often more reliable than gambling on one perfect attempt.

The kiln master also learned to control quantity by cost. Too few candidates, and you can't pick good cups; firing too many wastes firewood. True craftsmanship is knowing how many to produce, what standard to filter by, and whether rejected cups can be repurposed. The kiln master also learned to control quantity by cost. Too few candidates, and you can't pick good cups; firing too many wastes firewood. True craftsmanship is knowing how many to produce, what standard to filter by, and whether rejected cups can be repurposed. The kiln master also learned to control quantity by cost. Too few candidates, and you can't pick good cups; firing too many wastes firewood. True craftsmanship is knowing how many to produce, what standard to filter by, and whether rejected cups can be repurposed.

This method relies on first generating multiple candidates, then rejecting the unqualified ones according to criteria. It cannot replace good kiln craft, but amid uncertain kiln outcomes it can raise the quality of what is finally delivered.

Reveal

This story is about: rejection sampling

Rejection Sampling in AI training often refers to generating multiple candidate outputs and keeping only the qualified samples according to quality standards, used to construct higher-quality fine-tuning or preference data. It can improve data quality, but depends on the reliability of the filter, reward model, or human annotation. Enterprises can use it to accumulate high-quality customer service replies, tool-call traces, RAG answers, and task execution examples.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • Burn more cups: generate multiple candidate outputs
  • Explicit standard filtering: rules, reward model, eval or manual filtering
  • Leave sample cups to teach new apprentices: continue training with filtered samples
  • Review preference error: screening criteria shape model behavior
  • More generation is not the goal: quality thresholds determine data value

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

best-of-Nself-improvementSFTRLHF
京ICP备2026014272号