AI Trend Research

← AI glossary

AI glossary · Basic · Language models · Reinforcement learning

Post-training

Post-training is the stage that refines a model that has finished pretraining so people can use it well. It additionally teaches the model things like following instructions, holding conversations, answering safely and reasoning.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-08
In plain words It is like an onboarding program that uses work manuals and feedback to teach a new hire, who has read a lot and knows plenty but does not yet know how to do the job, "how to do the job well."

How it works

It usually goes through several stages. First, SFT (supervised fine-tuning) shows the model data pairing questions with model answers and has it imitate them. Next, methods such as RLHF or DPO adjust the model's answering tendencies based on answers that people chose as better. Recently, reinforcement learning such as RLVR, which gives rewards on tasks where correct answers can be checked automatically, like math problems or code, is also widely used.

The typical post-training flowBase model → SFT; SFT → Preference/reward learning; Preference/reward learning → Assistant modelPost-trainingBase modelResult of pretrainingSFTImitating modelanswersPreference/rewardlearningRLHF · DPO · RLVRAssistant modelSafety rules also at this stageAlignment · reducing sycophancy
The typical post-training flow A flow in which a base model built through pretraining is refined in turn with SFT and preference/reward learning to create an assistant model. The actual order and setup differ from company to company.

Why it matters

A model that has only finished pretraining may know a lot but can continue a question in odd ways or answer dangerous requests as is. After post-training, it becomes a usable product such as a conversational assistant, coding agent or reasoning model. Because it costs less than pretraining, models with different characters can be made from the same base model just by varying post-training. That is why post-training is taking up a growing share of today's model competition.

Things to know

Post-training is closer to drawing out abilities the model already has and changing how it behaves than to adding a lot of new knowledge. If rewards are poorly designed, problems can arise such as reward hacking, where the model learns tricks just to collect rewards, or sycophancy, where it flatters the user. Alignment work that makes the model follow safety rules is also mostly done at this stage.

Example

Even from the same base model, depending on how post-training was done, you get a general chat model, a reasoning model that thinks at length before answering, or a model for coding agents that fixes code. This is also why AI companies often announce that they "improved post-training so the model follows instructions better."

Related terms

Pre-training SFT RLHF DPO RLVR Alignment